CN117131160A - Intelligent query response interaction information mining method and system based on artificial intelligence - Google Patents
Intelligent query response interaction information mining method and system based on artificial intelligence Download PDFInfo
- Publication number
- CN117131160A CN117131160A CN202311208533.4A CN202311208533A CN117131160A CN 117131160 A CN117131160 A CN 117131160A CN 202311208533 A CN202311208533 A CN 202311208533A CN 117131160 A CN117131160 A CN 117131160A
- Authority
- CN
- China
- Prior art keywords
- text
- query response
- semantic
- initial
- interaction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000003993 interaction Effects 0.000 title claims abstract description 650
- 230000004044 response Effects 0.000 title claims abstract description 578
- 238000005065 mining Methods 0.000 title claims abstract description 72
- 238000000034 method Methods 0.000 title claims abstract description 45
- 238000013473 artificial intelligence Methods 0.000 title claims abstract description 19
- 230000006698 induction Effects 0.000 claims abstract description 247
- 230000002452 interceptive effect Effects 0.000 claims abstract description 225
- 239000013598 vector Substances 0.000 claims description 215
- 238000012545 processing Methods 0.000 claims description 156
- 238000012937 correction Methods 0.000 claims description 30
- 238000013139 quantization Methods 0.000 claims description 27
- 238000004590 computer program Methods 0.000 claims description 11
- 230000009191 jumping Effects 0.000 claims description 9
- 238000004891 communication Methods 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 abstract description 21
- 238000012549 training Methods 0.000 abstract description 20
- 230000002829 reductive effect Effects 0.000 abstract description 17
- 230000008569 process Effects 0.000 abstract description 9
- 238000010801 machine learning Methods 0.000 abstract description 8
- 238000013528 artificial neural network Methods 0.000 abstract description 7
- 238000003062 neural network model Methods 0.000 abstract description 6
- 238000007418 data mining Methods 0.000 abstract description 2
- 238000013461 design Methods 0.000 description 37
- 230000001965 increasing effect Effects 0.000 description 13
- 238000009826 distribution Methods 0.000 description 11
- 230000001939 inductive effect Effects 0.000 description 6
- 238000010586 diagram Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 230000009193 crawling Effects 0.000 description 2
- 230000003247 decreasing effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000036961 partial effect Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3344—Query execution using natural language analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3346—Query execution using probabilistic model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2216/00—Indexing scheme relating to additional aspects of information retrieval not explicitly covered by G06F16/00 and subgroups
- G06F2216/03—Data mining
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Probability & Statistics with Applications (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The application relates to the technical field of machine learning and big data mining, and provides an intelligent query response interaction information mining method and system based on artificial intelligence, which can provide high-quality training samples for training of a neural network. In detail, when the induction update instruction is updated based on the condition that the offset index of the selected interactive text is reduced, the semantic difference variable corresponding to the initial induction query response interactive text can be gradually reduced in the updating process, so that the similarity of the final induction query response interactive text and the query response interactive text of the selected online dialogue theme in the text detail level is improved, and the generation quality of the induction text example is improved. Therefore, the debugging quality of the neural network model can be improved based on the induced text examples, so that the operation quality of the neural network model is guaranteed to improve the accuracy of dialogue topic summary analysis.
Description
Technical Field
The application relates to the technical field of machine learning and big data mining, in particular to an intelligent query response interaction information mining method and system based on artificial intelligence.
Background
Machine Learning (Machine Learning) is a common research hotspot in the fields of artificial intelligence and pattern recognition, and its theory and method have been widely used to solve complex problems in engineering applications and scientific fields. With the continuous maturity of AI question-answering technology, machine learning is increasingly widely applied in the field of intelligent question-answering interaction. In the actual implementation process, in order to improve the quality of intelligent question-answer interaction, dialogue topic summarization is generally required, and dialogue topic summarization analysis by using a neural network model is one of main means. How to ensure the operation quality of the neural network model to improve the accuracy of the dialogue topic summary analysis is a technical problem which needs to be improved at present.
Disclosure of Invention
The application provides at least one intelligent query response interaction information mining method and system based on artificial intelligence.
The technical scheme of the application is realized by at least partial embodiments as follows.
An intelligent query response interaction information mining method based on artificial intelligence is applied to an intelligent query response processing system, and the method comprises the following steps:
acquiring a selected query response interaction text of a selected online dialogue topic and an initial induced query response interaction text corresponding to an induced online dialogue topic;
Performing question-answer interaction semantic mining on the selected query response interaction text to obtain a selected question-answer interaction semantic vector, and combining the selected question-answer interaction semantic vector to obtain an auxiliary question-answer interaction semantic vector;
carrying out question-answer interaction semantic mining on the initial induced query response interaction text to obtain an induced question-answer interaction semantic vector;
determining semantic difference variables between the induction question-answer interaction semantic vectors and the auxiliary question-answer interaction semantic vectors, and combining the semantic difference variables to obtain a selected interaction text offset index;
the semantic difference variable has a first quantization relation with the selected interactive text offset index;
and obtaining an induction updating instruction by combining the selected interactive text offset index, and updating a text unit description value of the initial induction inquiry response interactive text by combining the induction updating instruction to obtain a final induction inquiry response interactive text corresponding to the selected inquiry response interactive text.
In some exemplary embodiments, the selecting query response interaction text is a plurality of, the performing the query response interaction text on the selected query response interaction text to obtain a selected query response interaction semantic vector, and combining the selected query response interaction semantic vector to obtain an auxiliary query response interaction semantic vector includes:
Performing question-answer interaction semantic mining on each selected query response interaction text to obtain selected question-answer interaction semantic vectors corresponding to each selected query response interaction text;
acquiring an initial semantic thermodynamic relationship network, and determining the initial existence probability of each selected question-answer interaction semantic vector in the initial semantic thermodynamic relationship network;
summarizing the initial existence probabilities corresponding to the selected question-answer interaction semantic vectors to obtain initial probability summarizing results, updating the relation network variables corresponding to the initial semantic thermodynamic relation network to adjust the initial probability summarizing results to obtain a target semantic thermodynamic relation network;
and obtaining a target linear output field corresponding to the target semantic thermodynamic relationship network to obtain the auxiliary question-answering interaction semantic vector.
In some exemplary embodiments, the initial semantic thermodynamic relationship network comprises a first initial semantic thermodynamic relationship network and a second initial semantic thermodynamic relationship network, and the initial existence probabilities comprise a first initial existence probability corresponding to the first initial semantic thermodynamic relationship network and a second initial existence probability corresponding to the second initial semantic thermodynamic relationship network;
summarizing the initial existence probabilities corresponding to the selected question-answer interaction semantic vectors to obtain initial probability summarizing results, updating the relation network variables corresponding to the initial semantic thermodynamic relation network to adjust the initial probability summarizing results, and obtaining a target semantic thermodynamic relation network comprises the following steps:
Acquiring a first initial thermodynamic index corresponding to the first initial semantic thermodynamic relationship net and a second initial thermodynamic index corresponding to the second initial semantic thermodynamic relationship net;
combining the first initial heat index and the first initial existence probability, and calculating the second initial heat index and the second initial existence probability to obtain initial existence probability corresponding to the selected question-answer interaction semantic vector;
summarizing initial existence probabilities corresponding to the selected question-answer interaction semantic vectors to obtain initial probability summarizing results, updating relation network variables corresponding to a first initial semantic thermodynamic relation network and a second initial semantic thermodynamic relation network, the first initial thermodynamic index and the second initial thermodynamic index, and adjusting the initial probability summarizing results to obtain a first target semantic thermodynamic relation network, a second target semantic thermodynamic relation network, a first target thermodynamic index and a second target thermodynamic index.
In some exemplary embodiments, the obtaining the auxiliary question-answer interaction semantic vector includes:
determining a first target existence probability of the induction question-answer interaction semantic vector in the first target semantic thermodynamic relationship network, and combining the first target thermodynamic index and the first target existence probability to obtain a first correction probability;
Determining a second target existence probability of the induction question-answer interaction semantic vector in the second target semantic thermodynamic relationship network, and combining the second target thermodynamic index and the second target existence probability to obtain a second correction probability;
selecting a semantic thermodynamic relationship network with the maximum probability value from the first target semantic thermodynamic relationship network and the second target semantic thermodynamic relationship network by combining the first correction probability and the second correction probability as a significance semantic thermodynamic relationship network;
and obtaining a target linear output field corresponding to the significance semantic thermodynamic relationship network as the auxiliary question-answer interaction semantic vector.
In some exemplary embodiments, the deriving the selected interaction text offset index in combination with the semantic difference variable comprises:
combining the semantic difference variable to obtain a first query response interaction text cost coefficient, wherein the semantic difference variable has a first quantization relation with the first query response interaction text cost coefficient;
determining a first credibility coefficient of an initial induced query response interaction text on the selected online dialogue topic by combining the induced question response interaction semantic vector;
combining the first trusted coefficient to obtain a second query response interactive text cost coefficient;
The second query response interaction text cost coefficient is in second quantization relation with the first trusted coefficient;
and combining the first query response interaction text cost coefficient with the second query response interaction text cost coefficient to obtain a selected interaction text offset index.
In some exemplary embodiments, the obtaining the second query response interaction text cost coefficient in combination with the first confidence coefficient includes:
determining a second credibility coefficient of the initial induction inquiry response interaction text in the induction online dialogue topic by combining the induction inquiry response interaction semantic vector;
combining a first trusted comparison result between the first trusted coefficient and the second trusted coefficient to obtain a second query response interaction text cost coefficient; and a second quantization relation exists between the second query response interaction text cost coefficient and the first trusted comparison result.
In some exemplary embodiments, the deriving the selected interaction text offset index in combination with the semantic difference variable comprises:
combining the semantic difference variable to obtain a first query response interaction text cost coefficient, wherein the semantic difference variable has a first quantization relation with the first query response interaction text cost coefficient;
Determining the priori credibility coefficients of the initial induced query response interaction text on each priori online dialogue topic by combining the induced question response interaction semantic vectors;
selecting the largest trusted coefficient from the prior trusted coefficients as a third trusted coefficient;
combining a second trusted comparison result between the first trusted coefficient and the third trusted coefficient to obtain a third query response interaction text cost coefficient; the third query response interaction text cost coefficient has a second quantization relation with the second trusted comparison result;
and combining the first query response interaction text cost coefficient with the third query response interaction text cost coefficient to obtain a selected interaction text offset index.
In some exemplary embodiments, the updating the text unit description value of the initial induced query response interaction text in combination with the induced update instruction, and obtaining the final induced query response interaction text corresponding to the selected query response interaction text includes:
when the second trusted comparison result is smaller than the trusted comparison set value, the text unit description value of the initial induced query response interactive text is updated in combination with the induced update instruction, and the updated initial induced query response interactive text is obtained;
And jumping to the step of carrying out question-answer interaction semantic mining on the initial induction query response interaction text to obtain induction question-answer interaction semantic vectors until the second trusted comparison result reaches a trusted comparison set value, and taking the initial induction query response interaction text as a final induction query response interaction text corresponding to the selected query response interaction text.
In some exemplary embodiments, the updating the text unit description value of the initial induced query response interaction text in combination with the induced update instruction, and obtaining the final induced query response interaction text corresponding to the selected query response interaction text includes:
updating a text unit description value of the initial induction query response interactive text by combining the induction updating instruction to obtain an optimized initial induction query response interactive text;
determining an initial query response interaction text difference variable between the optimized initial induction query response interaction text and a basic induction query response interaction text corresponding to the initial induction query response interaction text, and jumping to a step of carrying out question-answer interaction semantic mining on the initial induction query response interaction text when the initial query response interaction text difference variable is smaller than a text difference variable set value to obtain an induction question-answer interaction semantic vector until the initial query response interaction text difference variable reaches the text difference variable set value, and taking the initial induction query response interaction text as a final induction query response interaction text corresponding to the selected query response interaction text.
In some exemplary embodiments, the induction update instruction includes description update values corresponding to respective text units, and the updating the text unit description values of the initial induction query response interaction text in combination with the induction update instruction includes:
processing the selected interactive text offset index by combining text unit description values of all text units of the initial induction query response interactive text to obtain description update values corresponding to all the text units in the initial induction query response interactive text;
and updating the text unit description value of the text unit in the initial induced query response interaction text by combining the description update values respectively corresponding to the text units to obtain the final induced query response interaction text corresponding to the selected query response interaction text.
In some exemplary embodiments, the method further comprises:
loading the final induction query response interaction text into an AI interaction text processing network to be debugged corresponding to the selected online dialogue topic to obtain an induction credibility coefficient of the final induction query response interaction text corresponding to the selected online dialogue topic;
Determining a target network cost coefficient by combining the induction credible coefficient; the target network cost coefficient is in first quantization connection with the induction credible coefficient;
and updating network variables of the AI interactive text processing network by combining the target network cost coefficient to obtain the AI interactive text processing network for completing debugging.
An intelligent query response processing system, comprising: a processor, a memory, and a network interface; the processor is connected with the memory and the network interface; the network interface is for providing data communication functions, the memory is for storing program code, and the processor is for invoking the program code to perform the above-described method.
A computer readable storage medium having stored thereon a computer program which, when run, performs an artificial intelligence based intelligent query response mutual information mining method.
A computer program product comprising a computer program or computer executable instructions that when executed by a processor implement an artificial intelligence based intelligent query response interaction information mining method.
According to one embodiment of the application, a selected query response interaction text of a selected online dialogue topic and an initial induction query response interaction text corresponding to the induction online dialogue topic are obtained, question-answer interaction semantic mining is carried out on the selected query response interaction text to obtain a selected question-answer interaction semantic vector, auxiliary question-answer interaction semantic vectors are obtained by combining the selected question-answer interaction semantic vectors, question-answer interaction semantic mining is carried out on the initial induction query response interaction text to obtain an induction question-answer interaction semantic vector, semantic distinction variables between the induction question-answer interaction semantic vector and the auxiliary question-answer interaction semantic vector are determined, a selected interaction text offset index is obtained by combining the semantic distinction variables, a first quantized relation exists between the semantic distinction variables and the selected interaction text offset index, induction update indication is obtained by combining the selected interaction text offset index, and a text unit description value of the initial induction query response interaction text is updated by combining the induction update indication to obtain a final induction query response interaction text corresponding to the selected query response interaction text.
In view of the fact that the auxiliary question-answer interaction semantic vector is obtained according to the selected question-answer interaction semantic vector, the auxiliary question-answer interaction semantic vector can represent text details of a query response interaction text of a selected online dialog theme, so that a semantic distinction variable can represent distinction between the induction question-answer interaction semantic vector and the text details of the query response interaction text of the selected online dialog theme, a first quantized relation exists between a selected interaction text offset index and the semantic distinction variable, and the purposes that when an induction update instruction is updated based on a condition that the selected interaction text offset index is reduced, the semantic distinction variable corresponding to an initial induction query response interaction text is gradually reduced in an update process can be achieved, and the similarity of the final induction query response interaction text and the query response interaction text of the selected online dialog theme in a text detail layer is further improved, and the generation quality of induction text examples is improved. Therefore, the debugging quality of the neural network model can be improved based on the induced text examples, so that the operation quality of the neural network model is guaranteed to improve the accuracy of dialogue topic summary analysis.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the application as claimed.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are necessary for the embodiments to be used are briefly described below, the drawings being incorporated in and forming a part of the description, these drawings showing embodiments according to the present application and together with the description serve to illustrate the technical solutions of the present application. It is to be understood that the following drawings illustrate only certain embodiments of the application and are therefore not to be considered limiting of its scope, for the person of ordinary skill in the art may admit to other equally relevant drawings without inventive effort.
Fig. 1 shows a schematic diagram of an intelligent query response processing system according to an embodiment of the present application.
Fig. 2 shows a flowchart of an intelligent query response interaction information mining method based on artificial intelligence according to an embodiment of the present application.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present application more apparent, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application, and it is apparent that the described embodiments are only some embodiments of the present application, not all embodiments. The components of the embodiments of the application generally described and illustrated herein may be arranged and designed in a wide variety of different configurations.
Thus, the following detailed description of the embodiments of the application is not intended to limit the scope of the application, as claimed, but is merely representative of selected embodiments of the application. All other embodiments, which can be made by a person skilled in the art without making any inventive effort, are intended to be within the scope of the present application.
Fig. 1 is a schematic diagram of an intelligent query response processing system according to an embodiment of the present application, where the intelligent query response processing system 100 includes a processor 110, a memory 120, and a network interface 130. The processor 110 is connected to the memory 120 and the network interface 130. Further, the network interface 130 is configured to provide data communication functions, the memory 120 is configured to store program codes, and the processor 110 is configured to invoke the program codes to perform an artificial intelligence based intelligent query response interaction information mining method.
Fig. 2 is a flow diagram illustrating an artificial intelligence based intelligent query response mutual information mining method that may implement embodiments of the present application, which may be implemented by the intelligent query response processing system 100 shown in fig. 1, and which illustratively includes steps 202-210.
Step 202, obtaining a selected query response interactive text of a selected online dialogue topic and an initial induced query response interactive text corresponding to an induced online dialogue topic.
In the embodiment of the invention, the online dialogue theme can be understood as the query response type corresponding to the query response interaction text, can be obtained according to the induction of the query response interaction text information, for example, can be obtained according to the induction of key words and sentences included in the query response interaction text, for example, when the query response interaction text comprises a "meta space digital space", the query response interaction text can be determined to belong to the meta space concept type, and when the query response interaction text comprises a "digital finance", the query response interaction text can be determined to belong to the digital finance concept type. The selected online dialog topic may be any online dialog topic, such as a meta-universe concept category. The induced online conversation topic is an online conversation topic that is different from the selected online conversation topic, such as may be a digital financial concept category.
The query response type of the query response interactive text can be determined based on the pre-debugged AI interactive text processing network to be induced and debugged, and when the pre-debugged AI interactive text processing network is a neural network for analyzing the induced interactive text in the query response interactive text, the query response type of the query response interactive text can be determined according to the dialogue event included in the query response interactive text. For example, when the pre-debugged AI interactive text processing network is a neural network for analyzing the "meta-universe concept type", the online dialogue topic of the "meta-universe concept type" can be used as a selected online dialogue topic, the query response interactive text of the "meta-universe concept type" can be used as a selected query response interactive text, and when the analysis performance of an induced text example obtained by the pre-debugged AI interactive text processing network on the query response interactive text of the "digital financial concept type" is to be improved, the induced online dialogue topic can be set as the "digital financial concept type", the query response interactive text of the "digital financial concept type" can be used as an initial induced query response interactive text, or the query response interactive text of the "digital financial concept type" can be updated to obtain an initial induced query response interactive text.
The selected query response interaction text may be understood as query response interaction text belonging to the selected online dialog topic, i.e. the correct online dialog topic of the selected query response interaction text is the selected online dialog topic. The initial induction query response interactive text can be a basic induction query response interactive text which is not updated, or can be a query response interactive text obtained by carrying out one or more times of noise update on the basic induction query response interactive text by using the method provided by the embodiment of the invention, wherein the basic induction query response interactive text is not updated and belongs to an induction online dialogue theme. The selected query response interaction text and the underlying induced query response interaction text may be the actual query response interaction text that the text crawling thread crawls directly.
The intelligent query response processing system may obtain a basic induced query response interaction text of the induction online dialogue theme, and may use the basic induced query response interaction text as an initial induced query response interaction text, or perform confusion processing on the basic induced query response interaction text, for example, modify a text unit description value of the basic induced query response interaction text or introduce an influence feature on the basic induced query response interaction text, and use the confused query response interaction text as the initial induced query response interaction text.
Under some exemplary design ideas, the online dialogue platform server may send an induced debugging application (which may also be understood as a disturbance debugging application or an countermeasure debugging application) to the intelligent query response processing system, where the induced debugging application may carry a neural network to be subjected to induced debugging and an induced online dialogue topic, the intelligent query response processing system may obtain a query response interaction text of the induced online dialogue topic, obtain an initial induced query response interaction text according to the query response interaction text of the induced online dialogue topic, and the intelligent query response processing system may determine an online dialogue topic for analysis by the neural network to be subjected to induced debugging as the selected online dialogue topic.
And 204, carrying out question-answer interaction semantic mining on the selected query response interaction text to obtain a selected question-answer interaction semantic vector, and combining the selected question-answer interaction semantic vector to obtain an auxiliary question-answer interaction semantic vector.
In the embodiment of the invention, the selected question-answer interaction semantic vector is text details obtained by conducting question-answer interaction semantic mining on the selected query response interaction text. The auxiliary question-answer interaction semantic vector can be text details obtained by summarizing and determining the selected question-answer interaction semantic vector corresponding to each selected query-answer interaction text when a plurality of selected query-answer interaction texts are selected. For example, the summary determination may include an average operation or a covariance operation.
By way of example, the intelligent query response processing system may utilize a related AI algorithm to perform question-answer interaction semantic mining on the selected query response interaction text to obtain a selected question-answer interaction semantic vector, for example, may perform question-answer interaction semantic mining on the selected query response interaction text through a pre-debugged AI interaction text processing network to obtain the selected question-answer interaction semantic vector. The question-answer interaction semantic mining can be understood as extracting features of related texts, so that corresponding question-answer interaction semantic vectors, namely question-answer interaction content features, are obtained.
Under some exemplary design ideas, the intelligent query response processing system can utilize a pre-debugged AI interactive text processing network to conduct question-answer interaction semantic mining on the initial induced query response interactive text to obtain an induced question-answer interaction semantic vector, and determine auxiliary question-answer interaction semantic vectors by combining the induced question-answer interaction semantic vectors, wherein the auxiliary question-answer interaction semantic vectors corresponding to different induced question-answer interaction semantic vectors can be the same or different. For example, the intelligent query response processing system can determine auxiliary question-answer interaction semantic vectors corresponding to the induction question-answer interaction semantic vectors according to the difference between the induction question-answer interaction semantic vectors and each selected question-answer interaction semantic vector, for example, the selected question-answer interaction semantic vector with the smallest difference between the induction question-answer interaction semantic vectors can be used as the auxiliary question-answer interaction semantic vector.
Under some exemplary design ideas, the intelligent query response processing system can acquire selected question-answer interaction semantic vectors corresponding to a plurality of selected query response interaction texts, summarize each selected question-answer interaction semantic vector, determine a target semantic thermodynamic relationship net matched with each selected question-answer interaction semantic vector, and determine auxiliary question-answer interaction semantic vectors corresponding to the induced question-answer interaction semantic vectors by combining the target semantic thermodynamic relationship net. For example, the intelligent query response processing system may use the text details with the discrimination corresponding to the target semantic thermal relationship network as auxiliary question-answer interaction semantic vectors, for example, the text details with the highest probability of existence in the target semantic thermal relationship network may be used as auxiliary question-answer interaction semantic vectors, for example, when the target semantic thermal relationship network is normally distributed, the average result of the target semantic thermal relationship network may be used as the auxiliary question-answer interaction semantic vectors.
And 206, carrying out question-answer interaction semantic mining on the initial induced query response interaction text to obtain an induced question-answer interaction semantic vector.
The induction question-answer interaction semantic vector is text details obtained by conducting question-answer interaction semantic mining on an initial induction query-answer interaction text. The concept of performing the question-answer interaction semantic mining on the initial induced query response interaction text may be the same as the concept of performing the question-answer interaction semantic mining on the selected query response interaction text, for example, the induced question-answer interaction semantic vector and the selected question-answer interaction semantic vector may be information generated by the same question-answer interaction semantic mining layer of the pre-debugged AI interaction text processing network. The question-answer interaction semantic mining layer is used for question-answer interaction semantic vectors and can comprise at least one of linear mining nodes or nonlinear mining nodes. The nonlinear mining node may be, for example, a RELU node.
Under some exemplary design ideas, the pre-debugged AI interactive text processing network comprises a plurality of question-answer interaction semantic mining layers, the intelligent query response processing system can load selected query response interaction texts into the pre-debugged AI interactive text processing network to obtain selected question-answer interaction semantic vectors respectively generated by at least two question-answer interaction semantic mining layers, load initial induced query response interaction texts into the pre-debugged AI interactive text processing network to obtain induced question-answer interaction semantic vectors respectively generated by each question-answer interaction semantic mining layer, and determine auxiliary question-answer interaction semantic vectors corresponding to the induced question-answer interaction semantic vectors generated by the question-answer interaction semantic mining layers according to the selected question-answer interaction semantic vectors and the induced question-answer interaction semantic vectors generated by the same question-answer interaction semantic mining layer.
Under some exemplary design ideas, the intelligent query response processing system can load a selected query response interaction text into a pre-debugged AI interaction text processing network, acquire text details generated by a nonlinear mining node of the pre-debugged AI interaction text processing network, obtain a selected question-answer interaction semantic vector corresponding to the selected query response interaction text, and load an initial induced query response interaction text into the pre-debugged AI interaction text processing network, so as to obtain text details generated by the nonlinear mining node, and obtain an induced question-answer interaction semantic vector. Wherein the nonlinear mining nodes may have one or more. A plurality is understood to be at least two.
Step 208, determining semantic difference variables between the induction question-answer interaction semantic vectors and the auxiliary question-answer interaction semantic vectors, and combining the semantic difference variables to obtain a selected interaction text offset index; the semantic difference variable has a first quantitative association with the selected interactive text offset index.
Illustratively, the semantic difference variable may be understood as a difference between the induced question-answer interaction semantic vector and the auxiliary question-answer interaction semantic vector. The intelligent query response processing system can determine the vector difference between the induction question-answer interaction semantic vector and the auxiliary question-answer interaction semantic vector as a semantic difference variable, for example, can determine the cosine distance between the induction question-answer interaction semantic vector and the auxiliary question-answer interaction semantic vector, and obtains the semantic difference variable. The intelligent query response processing system can also determine the similarity between the interaction semantic vector of the induced question and answer and the interaction semantic vector of the auxiliary question and answer, so as to obtain semantic similarity, determine a semantic difference variable according to the semantic similarity, and make a second quantitative connection with the semantic similarity, for example, the reciprocal of the semantic similarity can be used as the semantic difference variable. Wherein the first quantized relationship is a positive correlation relationship and the second quantized relationship is a negative correlation relationship. Further, the selected interactive text offset index may be understood as a selected interactive text loss value or a selected interactive text loss variable.
Under some exemplary design ideas, the intelligent query response processing system can acquire a selected question-answer interaction semantic vector generated by a preset question-answer interaction semantic mining layer of a pre-debugged AI interaction text processing network, an induced question-answer interaction semantic vector and an auxiliary question-answer interaction semantic vector determined according to the selected question-answer interaction semantic vector, determine a semantic difference variable (characteristic difference value) between the induced question-answer interaction semantic vector and the auxiliary question-answer interaction semantic vector, transpose the semantic difference variable to obtain a target difference variable, combine the semantic difference variable and the target difference variable to obtain a hierarchical interaction text cost coefficient, and have a first quantized relation with the target difference variable. For example, when the auxiliary question-answer interaction semantic vector is an average result of normal distribution corresponding to the target semantic thermodynamic relationship network, the intelligent query response processing system may perform multiplication processing on the target differential variable and the covariance list of normal distribution to obtain a first quantization processing result, perform multiplication processing on the first quantization processing result and the semantic differential variable to obtain a hierarchical interaction text cost coefficient corresponding to the preset question-answer interaction semantic mining layer, obtain a selected interaction text offset index by combining the query response interaction text cost coefficient, where the selected interaction text offset index and the query response interaction text cost coefficient have a first quantization relationship, for example, the preset question-answer interaction semantic mining layer may have multiple, the intelligent query response processing system may obtain hierarchical interaction text cost coefficients corresponding to each set mining layer, and perform weight-based operation processing on each hierarchical interaction text cost coefficient to obtain the selected interaction text offset index.
Under some exemplary design ideas, the intelligent query response processing system can acquire an induction analysis result of an initial induction query response interaction text generated by the pre-debugged AI interaction text processing network, the induction analysis result can comprise a trusted coefficient of the initial induction query response interaction text belonging to the selected online dialogue topic, and a second query response interaction text cost coefficient is obtained by combining the trusted coefficient of the initial induction query response interaction text belonging to the selected online dialogue topic, and has a second quantitative relation with the trusted coefficient of the initial induction query response interaction text belonging to the selected online dialogue topic. The intelligent query response processing system may obtain the selected interactive text deviation index based on the first query response interactive text cost coefficient and the second query response interactive text cost coefficient. The confidence coefficient is used for representing the probability that the query response interaction text belongs to each online dialogue topic, and the larger the confidence coefficient is, the larger the probability is, and the numerical interval of the confidence coefficient can be 0-1.
And 210, obtaining an induction updating instruction by combining the offset index of the selected interactive text, and updating a text unit description value of the initial induction query response interactive text by combining the induction updating instruction to obtain a final induction query response interactive text corresponding to the selected query response interactive text.
The induction update instruction is used for updating a text unit description value of a text unit in the initial induction query response interactive text, and may include induction update instructions corresponding to each text unit in the initial induction query response interactive text, for the text unit in the initial induction query response interactive text. The AI-interactive text processing network corresponding to the selected online dialog topic may be understood as a neural network for parsing the query response interactive text of the selected online dialog topic, and is a debugged pre-debugged AI-interactive text processing network. In addition, the AI interactive text processing network corresponding to the selected online conversation topic can also analyze and obtain the query response interactive text of the online conversation topic outside the selected online conversation topic.
The final induced query response interaction text is a query response interaction text obtained by updating a text unit description value of the initial induced query response interaction text. The final induced query response interactive text may be used as an example of the induced text. The induced text examples are training examples that are misguided by training. Training misguidance can be understood as the introduction of adaptive aliasing processing in conventional training examples, such that neural networks incorporating AI techniques output erroneous information. Conventional training examples may be understood as original samples that have not been misled by training, such as query response interactive text crawled by a text crawling thread.
The induction debugging (in other words, the induction debugging can be understood as anti-debugging or noise debugging) is one of core means for strengthening the anti-interference performance of the machine learning model, in the process of the induction debugging, adaptive confusion processing is introduced into the training sample, and the induction debugging is expected to enable the machine learning model to receive the change, so that the machine learning model has the anti-interference performance on the induction text sample, and the original sample and the induction text sample are accurately distinguished.
The intelligent query response processing system determines a derivative of the offset index of the selected interactive text relative to the interactive text of the initial induced query response, and obtains induction update instructions corresponding to each text unit in the interactive text of the initial induced query response. The text unit stores the difference and the corresponding induced update indication (perturbation update value) may be different. For one text unit in the initial induction query response interaction text, the intelligent query response processing system can determine a result obtained by summing/multiplying a text unit description value (which can be understood as a characteristic value) of the text unit and a corresponding induction update instruction to obtain an optimized text unit description value corresponding to the text unit, obtain a final induction query response interaction text according to the initial induction query response interaction text after updating the text unit description value, take the initial induction query response interaction text after updating the text unit description value as the final induction query response interaction text, and can also continuously update the text unit description value of the initial induction query response interaction text after updating the text unit description value until the text unit description value is up to meet the update termination requirement of the text unit description value, and take the initial induction query response interaction text meeting the update termination requirement of the text unit description value as the final induction query response interaction text. The text unit description value updating termination requirement comprises, but is not limited to, that the fluctuation value of the selected interactive text deviation index is smaller than the preset deviation cost fluctuation, and the difference between the optimized initial induced query response interactive text and the basic induced query response interactive text corresponding to the initial induced query response interactive text reaches the text difference variable set value.
In the embodiment of the invention, the text unit can be a part of the related query response interactive text, for example, the text unit can be a word, a phrase, a sentence or a segment. The text distinction variable setting value may be understood as a set text distinction value.
Under some exemplary design ideas, after the final induced query response interaction text is obtained, induced debugging can be performed on the AI interaction text processing network corresponding to the selected online dialogue topic by utilizing the final induced query response interaction text, so that query response interaction text analysis can be performed by utilizing the AI interaction text processing network after completion of debugging. For example, the intelligent query response processing system can take the final induced query response interactive text as a negative training sample, and conduct induced debugging on the AI interactive text processing network corresponding to the selected online dialogue theme, namely conduct induced debugging on the pre-debugged AI interactive text processing network, so that the anti-interference performance of the AI interactive text processing network on the induced text example is improved. By way of example, the intelligent query response processing system may debug the AI-interactive text processing network that has not undergone AI-interactive text processing network by using a conventional training sample, so as to obtain a pre-debugged AI-interactive text processing network, and then debug the pre-debugged AI-interactive text processing network by combining the conventional training sample and the induced text sample, so as to obtain the AI-interactive text processing network that completes the debugging.
It can be seen that, the method is applied to step 202-step 210, obtain the selected query response interaction text of the selected online dialogue topic and the initial induced query response interaction text corresponding to the induced online dialogue topic, question-answer interaction semantic mining is performed on the selected query response interaction text, obtain the selected question-answer interaction semantic vector, obtain the auxiliary question-answer interaction semantic vector in combination with the selected question-answer interaction semantic vector, question-answer interaction semantic mining is performed on the initial induced query response interaction text, obtain the induced question-answer interaction semantic vector, determine the semantic difference variable between the induced question-answer interaction semantic vector and the auxiliary question-answer interaction semantic vector, obtain the selected interaction text offset index in combination with the semantic difference variable, have a first quantized relation with the selected interaction text offset index, obtain the induced update indication in combination with the selected interaction text offset index, update the text unit description value of the initial induced query response interaction text in combination with the induced update indication, obtain the final induced query response interaction text corresponding to the selected query response interaction text, and use the final induced query response interaction text to perform induced debugging on the AI interaction processing network corresponding to the selected online dialogue topic, so as to use the completed AI interaction text processing network to perform query interaction text analysis and debugging. In view of the fact that the auxiliary question-answer interaction semantic vector is obtained according to the selected question-answer interaction semantic vector, the auxiliary question-answer interaction semantic vector can represent text details of a query response interaction text of a selected online dialog theme, so that a semantic distinction variable can represent distinction between the induction question-answer interaction semantic vector and the text details of the query response interaction text of the selected online dialog theme, a first quantized relation exists between a selected interaction text offset index and the semantic distinction variable, and the purposes that when an induction update instruction is updated based on a condition that the selected interaction text offset index is reduced, the semantic distinction variable corresponding to an initial induction query response interaction text is gradually reduced in an update process can be achieved, and the similarity of the final induction query response interaction text and the query response interaction text of the selected online dialog theme in a text detail layer is further improved, and generation quality is improved. When the network is debugged by utilizing the final induced query response interactive text, the anti-interference performance and stability of the network can be improved.
The intelligent query response interaction information mining method based on artificial intelligence can limit text details generated by a question-answer interaction semantic mining layer of an AI interaction text processing network.
Under some exemplary design ideas, the online dialogue platform server can send a query response interactive text analysis request to the intelligent query response processing system, the query response interactive text analysis request can carry query response interactive text to be analyzed, the intelligent query response processing system can analyze the query response interactive text to be analyzed by utilizing the AI interactive text processing network which completes debugging to obtain a query response interactive text analysis result corresponding to the query response interactive text to be analyzed, the query response interactive text analysis result is returned to the online dialogue platform server, and the online dialogue platform server can output the query response interactive text analysis result.
For example, the online dialogue platform server may trigger sending a query response interaction text analysis request for the selected online dialogue topic to the intelligent query response processing system, where the intelligent query response processing system obtains a query response interaction text to be analyzed carried in the query response interaction text analysis request, loads the query response interaction text to be analyzed into an AI interaction text processing network for completing debugging corresponding to the selected online dialogue topic, and obtains probabilities that the query response interaction text to be analyzed belongs to various online dialogue topics, for example, the probability of belonging to a "digital financial concept category" is 0.95, the probability of belonging to a "meta universe concept category" is 0.35, and when the probability of a "digital financial concept category" is greater than a preset probability of 0.7, the intelligent query response processing system may also output a query response interaction text analysis result that includes a digital financial concept category "in the query response interaction text.
Under some exemplary design ideas, the method comprises the steps of performing question-answer interaction semantic mining on a plurality of selected query response interaction texts to obtain selected question-answer interaction semantic vectors, and combining the selected question-answer interaction semantic vectors to obtain auxiliary question-answer interaction semantic vectors, wherein the steps comprise 402-408.
And step 402, carrying out question-answer interaction semantic mining on each selected query response interaction text to obtain selected question-answer interaction semantic vectors respectively corresponding to each selected query response interaction text.
Step 404, acquiring an initial semantic thermodynamic relationship network, and determining the initial existence probability of each selected question-answer interaction semantic vector in the initial semantic thermodynamic relationship network.
Step 406, summarizing the initial existence probabilities corresponding to the selected question-answer interaction semantic vectors to obtain initial probability summarizing results corresponding to the selected question-answer interaction semantic vectors, and updating the relation network variables corresponding to the initial semantic thermodynamic relation network to adjust the initial probability summarizing results to obtain the target semantic thermodynamic relation network.
Step 408, obtaining a target linear output field corresponding to the target semantic thermodynamic relationship network as an auxiliary question-answering interaction semantic vector.
The initial semantic thermodynamic relationship network may be any semantic thermodynamic feature distribution, such as a normal distribution. The relationship network variables corresponding to the initial semantic thermodynamic relationship network are used to determine the initial semantic thermodynamic relationship network, for example, when the initial semantic thermodynamic relationship network is normally distributed, the relationship network variables may include an average value list and a covariance list. The initial semantic thermodynamic relationship network includes probabilities of semantic content existence, through which the probabilities of semantic content existence can be determined. The initial existence probability can be understood as the probability that the selected question-answer interaction semantic vector exists in the initial semantic thermodynamic relationship network.
The initial probability summarizing result is obtained by summarizing initial existence probabilities corresponding to the selected question-answer interaction semantic vectors respectively, and the initial probability summarizing result and the selected question-answer interaction semantic vectors are in first quantitative connection, for example, the selected question-answer interaction semantic vectors can be summed to obtain the initial probability summarizing result, or multiplication processing is carried out on the selected question-answer interaction semantic vectors to obtain the initial probability summarizing result, inverse exponentiation operation can also be carried out on the selected question-answer interaction semantic vectors to obtain target question-answer interaction semantic vectors corresponding to the selected question-answer interaction semantic vectors, and summarizing and determining are carried out on the target question-answer interaction semantic vectors to obtain the initial probability summarizing result, for example, the summed result of the target question-answer interaction semantic vectors can be used as the initial probability summarizing result.
The target linear output field may be understood as text details with a degree of identification in the target semantic thermodynamic relationship network, for example, may be a target key vector of the target semantic thermodynamic relationship network, and the target key vector may be understood as text details with a maximum probability of existence in the target semantic thermodynamic relationship network. The target linear output field may also be text details in the target semantic thermodynamic relationship network that have a vector difference from the target key vector that is less than a feature vector difference threshold.
The intelligent query response processing system may update the relationship network variables corresponding to the initial semantic thermodynamic relationship network according to the rule of increasing the initial probability summary result, obtain the optimized relationship network variables, determine whether the update termination requirement is met, if not, skip to the step of updating the relationship network variables corresponding to the initial semantic thermodynamic relationship network according to the rule of increasing the initial probability summary result until the update termination requirement is met, and take the initial semantic thermodynamic relationship network corresponding to the optimized relationship network variables as the target semantic thermodynamic relationship network. The update termination requirements may include a distinguishing variable between two consecutive relationship network variables being less than a relationship network variable difference limit and a distinguishing variable between two consecutive initial probabilistic summary results being less than a probabilistic summary difference limit. The relationship network variable variance limits and the probability summary variance limits may be adjusted based on design considerations.
Under some exemplary design considerations, the initial semantic thermodynamic relationship network may include a plurality of initial local relationship networks, a plurality being understood as at least two. For each selected question-answer interaction semantic vector, the intelligent query response processing system can determine initial local existence probabilities of the selected question-answer interaction semantic vector in each initial local relation network respectively, and summarize and determine each initial local existence probability, for example, based on weight operation processing, so as to obtain initial existence probabilities corresponding to the selected question-answer interaction semantic vector, wherein the initial local existence probabilities can be understood as probabilities of the selected question-answer interaction semantic vector in the local relation network.
Under some exemplary design ideas, the local relation network corresponds to sub-relation network variables, the relation network variables can include sub-relation network variables corresponding to each initial local relation network, the intelligent query response processing system can continuously update the sub-relation network variables corresponding to each initial local relation network according to a rule of increasing initial probability summarizing results until the update termination requirement is met, the initial local relation network corresponding to the optimized sub-relation network variables is used as a target local relation network, and a target semantic thermodynamic relation network is obtained according to each target local relation network, for example, each target local relation network is integrated to obtain the target semantic thermodynamic relation network.
Under some exemplary design ideas, the intelligent query response processing system can acquire linear output fields (which can be understood as characterization vectors or representative features) corresponding to the target local relationship networks respectively to obtain local linear output fields, determine auxiliary question-answer interaction semantic vectors according to the target local relationship networks, for example, the auxiliary question-answer interaction semantic vectors can be obtained by summarizing and determining the local linear output fields corresponding to the target local relationship networks respectively, for example, the auxiliary question-answer interaction semantic vectors can be obtained by weight-based operation processing, auxiliary question-answer interaction semantic vectors can be selected from the local linear output fields, for example, the similarity between the induction question-answer interaction semantic vectors and the local linear output fields can be determined, and the local linear output field with the largest similarity is used as the auxiliary question-answer interaction semantic vector. The local linear output field can be flexibly set, for example, the text detail with the highest possibility of existence in the target local relation network can be adopted.
In the embodiment of the invention, the relation network variables corresponding to the initial semantic thermodynamic relation network are updated according to the rule of increasing the initial probability summarizing result, so that the initial semantic thermodynamic relation network gradually tends to the actual distribution matched with the text details of the query response interaction text of the selected online dialogue theme, the target semantic thermodynamic relation network can accurately reflect the actual distribution matched with the text details of the query response interaction text of the selected online dialogue theme, the target linear output field corresponding to the target semantic thermodynamic relation network is obtained and is used as an auxiliary question-answer interaction semantic vector, the text details with identity in the target semantic thermodynamic relation network are used as auxiliary question-answer interaction semantic vectors, and the accuracy of the auxiliary question-answer interaction semantic vectors is improved.
Under some exemplary design ideas, the initial semantic thermodynamic relationship net comprises a first initial semantic thermodynamic relationship net and a second initial semantic thermodynamic relationship net, and the initial existence probability comprises a first initial existence probability corresponding to the first initial semantic thermodynamic relationship net and a second initial existence probability corresponding to the second initial semantic thermodynamic relationship net; summarizing initial existence probabilities corresponding to the selected question-answer interaction semantic vectors to obtain initial probability summarizing results, updating relation network variables corresponding to the initial semantic thermodynamic relation network to adjust the initial probability summarizing results, and obtaining a target semantic thermodynamic relation network comprises the following steps: acquiring a first initial heat index corresponding to a first initial semantic heat relation network and a second initial heat index corresponding to a second initial semantic heat relation network; combining the first initial heat index and the first initial existence probability, and calculating the second initial heat index and the second initial existence probability to obtain the initial existence probability corresponding to the selected question-answer interaction semantic vector; summarizing initial existence probabilities corresponding to the selected question-answer interaction semantic vectors to obtain initial probability summarizing results, updating relation network variables, first initial heat indexes and second initial heat indexes corresponding to the first initial semantic heat relation network and the second initial semantic heat relation network, and adjusting the initial probability summarizing results to obtain a first target semantic heat relation network, a second target semantic heat relation network, a first target heat index and a second target heat index.
The first initial semantic thermodynamic relationship network and the second initial semantic thermodynamic relationship network may be the same class of likelihood maps, for example, may be normal distributions. The first initial existence probability can be understood as the existence probability of the selected question-answer interaction semantic vector in the first initial semantic thermodynamic relationship network, and the second initial existence probability can be understood as the existence probability of the selected question-answer interaction semantic vector in the second initial semantic thermodynamic relationship network.
The first initial heat index represents the probability percentage that the selected question-answer interaction semantic vector belongs to the first initial semantic heat relation network, and the second initial heat index represents the probability percentage that the selected question-answer interaction semantic vector belongs to the second initial semantic heat relation network. The sum of weights corresponding to the members included in the initial semantic thermodynamic relationship net is equal to 1, for example, when the initial semantic thermodynamic relationship net includes only the first initial semantic thermodynamic relationship net and the second initial semantic thermodynamic relationship net, the result of the summation of the first initial thermodynamic index and the second initial thermodynamic index is 1. The default values of the first initial heat index and the second initial heat index may be set in advance, and the default values may be, for example, 50% and 50%, respectively.
The target semantic thermodynamic relationship network comprises a first target semantic thermodynamic relationship network and a second target semantic thermodynamic relationship network. The first target semantic thermodynamic relationship net, the second target semantic thermodynamic relationship net, the first target thermodynamic index and the second target thermodynamic index are relationship net variables, the first initial thermodynamic index and the second initial thermodynamic index corresponding to the first initial semantic thermodynamic relationship net and the second initial semantic thermodynamic relationship net when the update termination requirement is met respectively.
The intelligent query response processing system may determine a multiplication result of the first initial heat index and the first initial existence probability to obtain a first weighted probability, determine a multiplication result of the second initial heat index and the second initial existence probability to obtain a second weighted probability, and sum the first weighted probability and the second weighted probability to determine an initial existence probability corresponding to the selected question-answer interaction semantic vector.
Under some exemplary design ideas, the intelligent query response processing system can update the relation network variables, the first initial heat index and the second initial heat index corresponding to the first initial semantic heat relation network and the second initial heat relation network according to a rule of increasing initial probability summary results until the update termination requirement is met, wherein the optimized first initial semantic heat relation network is used as a first target semantic heat relation network, the optimized second initial semantic heat relation network is used as a second target semantic heat relation network, the optimized first initial heat index is used as a first target heat index, and the optimized second initial heat index is used as a second target heat index. The update termination requirements may also include a difference between two consecutive initial heat indices being less than the heat index comparison limit.
In the embodiment of the invention, the relation network variables, the first initial heat index and the second initial heat index corresponding to the first initial semantic heat relation network and the second initial heat relation network are updated according to the rule of increasing the initial probability summarizing result to obtain the first target semantic heat relation network, the second target semantic heat relation network, the first target heat index and the second target heat index, so that the target semantic heat relation network comprises the first target semantic heat relation network and the second target semantic heat relation network, the distribution situation of text detail matching of the query response interaction text type can be selected through better reaction of the two heat relation networks, and the accuracy of the target semantic heat relation network is improved.
Under some exemplary design ideas, obtaining a target linear output field corresponding to a target semantic thermodynamic relationship network, and obtaining an auxiliary question-answer interaction semantic vector includes steps 502-508.
Step 502, determining a first target existence probability of the induction question-answer interaction semantic vector in the first target semantic thermodynamic relationship network, and combining the first target thermodynamic index and the first target existence probability to obtain a first correction probability.
Step 504, determining a second target existence probability of the induction question-answer interaction semantic vector in the second target semantic thermodynamic relationship network, and combining the second target thermodynamic index and the second target existence probability to obtain a second correction probability.
Step 506, selecting the semantic thermodynamic relationship net with the largest probability value from the first target semantic thermodynamic relationship net and the second target semantic thermodynamic relationship net as the significant semantic thermodynamic relationship net by combining the first correction probability and the second correction probability.
And step 508, obtaining a target linear output field corresponding to the significance semantic thermodynamic relationship network as an auxiliary question-answering interaction semantic vector.
The first target existence probability can be understood as the possibility of inducing the existence of the question-answer interaction semantic vector in the first target semantic thermodynamic relationship network, and the second target existence probability can be understood as the possibility of inducing the existence of the question-answer interaction semantic vector in the second target semantic thermodynamic relationship network. The saliency semantic thermodynamic relationship net may be any one of a first target semantic thermodynamic relationship net or a second target semantic thermodynamic relationship net. The target linear output field can be understood as text details with identification corresponding to the significance semantic thermodynamic relationship network, for example, text details with maximum existence probability in the significance semantic thermodynamic relationship network can be considered.
The intelligent query response processing system may determine a multiplication result of the first target heat index and the first target existence probability to obtain a first correction probability, determine a multiplication result of the second target heat index and the second target existence probability to obtain a second correction probability, compare the first correction probability with the second correction probability, determine a larger correction probability of the first correction probability and the second correction probability, and use the semantic thermodynamic relationship network corresponding to the target correction probability as a target correction probability, and use the semantic thermodynamic relationship network corresponding to the target correction probability as a saliency semantic thermodynamic relationship network, for example, when the first correction probability is greater than the second correction probability, the first target semantic thermodynamic relationship network corresponding to the first correction probability may be used as a target correction probability.
Under some exemplary design ideas, the intelligent query response processing system can combine the semantic distinction variable to obtain a first query response interaction text cost coefficient, obtain a selected interaction text offset index according to the first query response interaction text cost coefficient, wherein the selected interaction text offset index has a first quantized relation with the first query response interaction text cost coefficient, and the first query response interaction text cost coefficient has a first quantized relation with the semantic distinction variable.
Under some exemplary design ideas, the target semantic thermodynamic relationship net may be normal distribution, the first target semantic thermodynamic relationship net and the second target semantic thermodynamic relationship net are local relationship nets in the normal relationship net respectively, the intelligent query response processing system may determine a local relationship net from the normal distributed local relationship net as a significant semantic thermodynamic relationship net, for example, the intelligent query response processing system may determine the possibility of inducing the question-answer interaction semantic vector to exist in each local relationship net, so as to obtain the possibility distribution, and the local relationship net corresponding to the largest probability in the possibility relationship net is used as the significant semantic thermodynamic relationship net.
Under some exemplary design ideas, when the significance semantic thermal relationship network is a normal distributed local relationship network, the intelligent query response processing system can determine the existence probability of the induction question-answer interaction semantic vector in the local relationship network corresponding to the significance semantic thermal relationship network, and update the text unit description value of the initial induction query response interaction text according to the rule of increasing the existence probability to obtain the target induction query response interaction text corresponding to the selected query response interaction text.
The intelligent query response processing system can update the text unit description value of the initial induced query response interaction text according to the rule of reducing the cost coefficient of the first query response interaction text, namely update the text unit description value of the initial induced query response interaction text according to the rule of reducing the difference between the initial query response interaction semantic vector and the auxiliary query response interaction semantic vector, and can reduce the vector difference between the initial induced query response interaction text and the selected query response interaction text of the selected online dialogue theme in a high-order vector coordinate system.
In the embodiment of the invention, the semantic thermodynamic relationship net with the maximum probability value is selected from the first target semantic thermodynamic relationship net and the second target semantic thermodynamic relationship net and is used as the saliency semantic thermodynamic relationship net, so that the possibility of inducing the existence of the question-answer interaction semantic vector in the saliency semantic thermodynamic relationship net is higher than that of inducing the existence of the question-answer interaction semantic vector in the non-saliency semantic thermodynamic relationship net, thereby obtaining the target linear output field corresponding to the saliency semantic thermodynamic relationship net and being used as the auxiliary question-answer interaction semantic vector, improving the stability speed of the induction updating instruction, being more beneficial to obtaining the final induction query response interaction text under the condition of smaller induction updating instruction, and improving the efficiency of generating the final induction query response interaction text. In addition, in view of the fact that the semantic difference variable has a first quantitative relation with the selected interactive text offset index, namely, when the selected interactive text offset index is reduced, the semantic difference variable is reduced, the difference between the induction question-answer interaction semantic vector and the auxiliary question-answer interaction semantic vector is reduced, and therefore the difference between text details of the induction text example and the query response interactive text characteristics of the selected online dialogue topic can be reduced, and the fact that the text details of the induction text example are buried in a relation network matched with the text details of the query response interactive text of the selected online dialogue topic is achieved.
Under some exemplary design considerations, deriving the selected interaction text deviation index in combination with the semantic difference variable includes: combining the semantic difference variable to obtain a first query response interaction text cost coefficient, wherein the semantic difference variable has a first quantization relation with the first query response interaction text cost coefficient; determining a first credibility coefficient of the initial induction inquiry response interaction text on the selected online dialogue topic by combining the induction inquiry response interaction semantic vector; combining the first trusted coefficient to obtain a second query response interactive text cost coefficient; the second query response interaction text cost coefficient is in second quantization relation with the first trusted coefficient; and combining the first query response interaction text cost coefficient with the second query response interaction text cost coefficient to obtain a selected interaction text offset index.
Wherein the first confidence coefficient may be understood as a likelihood that the initially induced query response interaction text belongs to the selected online dialog topic. The second quantization relation exists between the second query response interaction text cost coefficient and the first trusted coefficient, and the reciprocal of the first trusted coefficient can be used as the second query response interaction text cost coefficient. The selected interactive text offset index is in a first quantized relationship with the first query response interactive text cost coefficient and the second query response interactive text cost coefficient. And both the cost factor and the offset index can understand the loss value or the loss variable.
The intelligent query response processing system may load the initial induced query response interaction text into the pre-debugged AI interaction text processing network to obtain a query response interaction text analysis result generated by the pre-debugged AI interaction text processing network, where the query response interaction text analysis result may include that the initial induced query response interaction text belongs to a selected online dialog topic, may further include that the initial induced query response interaction text belongs to a priori trusted coefficient of a priori online dialog topic, and the priori online dialog topic may be any query response category outside the selected online dialog topic, and may include at least one of an induced online dialog topic or an induced online dialog topic outside the online dialog topic. The priori trusted coefficient can be understood as a trusted coefficient for initially inducing the query response interaction text to belong to the priori online dialogue topic, and the intelligent query response processing system can combine the first trusted coefficient and the priori trusted coefficient to obtain a second query response interaction text cost coefficient, for example, can combine a distinguishing variable between the first trusted coefficient and the priori trusted coefficient to obtain the second query response interaction text cost coefficient.
Under some exemplary design ideas, the intelligent query response processing system may perform weight-based operation processing on the first query response interactive text cost coefficient and the second query response interactive text cost coefficient, and use a result of the weight-based operation processing as a selected interactive text offset index, or may use a result of summing the first query response interactive text cost coefficient and the second query response interactive text cost coefficient as the selected interactive text offset index.
In the embodiment of the invention, the second quantitative relation exists between the cost coefficient of the second query response interactive text and the first trusted coefficient, and when the cost coefficient of the second query response interactive text is gradually reduced and the first trusted coefficient is gradually increased, the possibility that the initial induction query response interactive text is analyzed into the selected online dialogue theme is increased, and the possibility that the initial induction query response interactive text is an induction text example of the selected online dialogue theme is improved.
Under some exemplary design considerations, combining the first confidence coefficient to obtain a second query response interaction text cost coefficient includes: determining a second credibility coefficient of the initial induction inquiry response interaction text in the induction online dialogue topic by combining the induction inquiry response interaction semantic vector; combining a first trusted comparison result between the first trusted coefficient and the second trusted coefficient to obtain a second query response interactive text cost coefficient; the second query response interaction text cost coefficient has a second quantized relationship with the first trusted comparison result.
Wherein, the second confidence coefficient can be understood as the possibility that the initial induced query response interactive text belongs to the induced online dialogue topic. The first confidence comparison result may have a first quantization relationship with a result of subtracting the second confidence coefficient from the first confidence coefficient, e.g., the result of subtracting the second confidence coefficient from the first confidence coefficient. The second query response interaction text cost coefficient has a second quantized relationship with the first trusted coefficient.
For example, the intelligent query response processing system may use the reciprocal of the first trusted comparison result as the cost coefficient of the second query response interaction text, or adjust the first trusted comparison result, and use the reciprocal corresponding to the adjusted result as the cost coefficient of the second query response interaction text.
In the embodiment of the invention, the second quantitative relation exists between the cost coefficient of the second query response interaction text and the first trusted comparison result, so that the distinction between the first trusted coefficient and the second trusted coefficient is reduced, and the possibility that the initial induced query response interaction text is analyzed as the first online dialogue theme is improved.
Under some exemplary design considerations, deriving the selected interaction text deviation index in combination with the semantic difference variable includes: combining the semantic difference variable to obtain a first query response interaction text cost coefficient, wherein the semantic difference variable has a first quantization relation with the first query response interaction text cost coefficient; determining priori trusted coefficients of the initial induced query response interactive text on each priori online dialogue topic by combining the induced question response interaction semantic vectors; selecting the largest trusted coefficient from all the priori trusted coefficients as a third trusted coefficient; combining a second trusted comparison result between the first trusted coefficient and the third trusted coefficient to obtain a third query response interactive text cost coefficient; and combining the first query response interaction text cost coefficient with the third query response interaction text cost coefficient to obtain a selected interaction text offset index.
Wherein the a priori online dialog topic may understand the reference dialog topic and the a priori trusted coefficient may be understood as the reference trusted coefficient. The third confidence coefficient is the largest confidence coefficient among the a priori confidence coefficients. The second confidence comparison result may be understood as a difference between the first confidence coefficient and the third confidence coefficient, and may have a first quantization relationship with a result obtained by subtracting the third confidence coefficient from the first confidence coefficient, for example, may be a result obtained by subtracting the third confidence coefficient from the first confidence coefficient. And the third query response interaction text cost coefficient has a second quantization relation with the second trusted comparison result. The third query response interaction text cost coefficient has a second quantized relationship with the first confidence coefficient.
The intelligent query response processing system may determine, for example, by performing weight-based operation processing, a selected interactive text offset index, where the selected interactive text offset index has a first quantized relationship with the first query response interactive text cost coefficient and the third query response interactive text cost coefficient, and the intelligent query response processing system may determine, for example, by performing weight-based operation processing, a summary of the first query response interactive text cost coefficient, the second query response interactive text cost coefficient, and the third query response interactive text cost coefficient, and obtain the selected interactive text offset index.
In the embodiment of the invention, the largest trusted coefficient is selected from all prior trusted coefficients and used as the third trusted coefficient, and the third query response interactive text cost coefficient is obtained by combining the second trusted comparison result between the first trusted coefficient and the third trusted coefficient, when the third query response interactive text cost coefficient is gradually reduced, the second trusted comparison result can be gradually increased, namely, when the first trusted coefficient is gradually increased, the third trusted coefficient is gradually decreased, so that the first trusted coefficient is larger than the third trusted coefficient, namely, the first trusted coefficient can be the largest trusted coefficient in all the trusted coefficients included in the query response interactive text analysis result, and thus, the initial induced query response interactive text is analyzed into the first online dialogue topic.
Under some exemplary design ideas, updating the text unit description value of the initial induced query response interaction text in combination with the induced update instruction, the obtaining the final induced query response interaction text corresponding to the selected query response interaction text includes: when the second trusted comparison result is smaller than the trusted comparison set value, the text unit description value of the initial induced query response interactive text is updated in combination with the induced update instruction, and the updated initial induced query response interactive text is obtained; and jumping to the step of carrying out question-answer interaction semantic mining on the initial induction query response interaction text to obtain induction question-answer interaction semantic vectors until the second trusted comparison result reaches the trusted comparison set value, and taking the initial induction query response interaction text as a final induction query response interaction text corresponding to the selected query response interaction text.
The trusted comparison setting value may be adjusted based on design conditions, or may be set in advance, for example. The second trusted comparison result may be a result obtained by subtracting the third trusted coefficient from the first trusted coefficient, and the text unit description value updating termination requirement may further include that the second trusted comparison result reaches a trusted comparison set value, when the second trusted comparison result reaches a set value, the intelligent query response processing system may determine that the text unit description value updating termination requirement is met, stop the step of updating the text unit description value of the initial induced query response interaction text in combination with the induction update instruction, and use the initial induced query response interaction text when the text unit description value updating termination requirement is met as the final induced query response interaction text.
In the embodiment of the invention, when the second trusted comparison result is smaller than the trusted comparison set value, the text unit description value of the initial induced query response interaction text is updated in combination with the induced update instruction to obtain the updated initial induced query response interaction text, the step of jumping to question-answer interaction semantic mining on the initial induced query response interaction text to obtain the induced question-answer interaction semantic vector is performed until the second trusted comparison result reaches the trusted comparison set value, the initial induced query response interaction text is used as the final induced query response interaction text corresponding to the selected query response interaction text, and the degree that the first trusted coefficient is higher than the third trusted coefficient can be adjusted, so that the first trusted coefficient is not excessively larger than the third trusted coefficient, and the network overfitting is avoided.
Under some exemplary design ideas, updating the text unit description value of the initial induced query response interaction text in combination with the induced update instruction, the obtaining the final induced query response interaction text corresponding to the selected query response interaction text includes: updating a text unit description value of the initial induction query response interactive text by combining the induction update instruction to obtain an optimized initial induction query response interactive text; determining an initial query response interaction text difference variable between the optimized initial induction query response interaction text and a basic induction query response interaction text corresponding to the initial induction query response interaction text, and jumping to the step of carrying out question-answer interaction semantic mining on the initial induction query response interaction text when the initial query response interaction text difference variable is smaller than a text difference variable set value to obtain an induction question-answer interaction semantic vector until the initial query response interaction text difference variable reaches a text difference variable set value, wherein the initial induction query response interaction text is used as a final induction query response interaction text corresponding to the selected query response interaction text.
Wherein the basic induced query response interactive text is not misled through training. The initial induced query response interaction text may be obtained by a confusion process that is not apparent in the underlying induced query response interaction text, such as modifying the text element description value. The optimized initial induced query response interactive text and basic induced query response interactive text are distinguished, for example, the difference of text unit description values of each text unit between the initial induced query response interactive text and the basic induced query response interactive text can be obtained, the description value distinguishing variable (the difference value of the description value) corresponding to each text unit is obtained, the description value distinguishing variables are summarized, and the initial query response interactive text distinguishing variable is obtained, for example, the result of weight-based operation processing of each description value distinguishing variable is used as the initial query response interactive text distinguishing variable. The initial query response interaction text distinction variable may also be a distinction between text details of the optimized initial induced query response interaction text and text details of the underlying induced query response interaction text.
The intelligent query response processing system may integrate the induction update instruction as the optimized text unit description value based on the text unit description value of the initial induction query response interaction text, so as to obtain the optimized initial induction query response interaction text, where at least one induction update instruction may be included, and a plurality of induction update instructions may be understood as at least two. When the induction update instruction is only one, the same update can be carried out on each text unit description value in the initial induction query response interactive text through the induction update instruction, for example, when the induction update instruction is 0.05, the text unit description value obtained in the initial induction query response interactive text can be increased by 0.05 on the basis of each text unit description value, and the optimized text unit description value is obtained.
Under some exemplary design ideas, the text unit description value updating termination requirement can further comprise the step of obtaining an induced question-answer interaction semantic vector by jumping to question-answer interaction semantic mining on the initial induced query response interaction text when the initial query response interaction text difference variable is smaller than the text difference variable set value, wherein the intelligent query response processing system can determine a difference variable between the optimized initial induced query response interaction text and a basic induced query response interaction text corresponding to the initial induced query response interaction text, and the step of obtaining the induced question-answer interaction semantic vector is carried out until the initial query response interaction text difference variable reaches the text difference variable set value, and takes the initial induced query response interaction text as a final induced query response interaction text corresponding to the selected query response interaction text.
Under some exemplary design ideas, there may be respective induction update indications corresponding to a plurality of text units, where each induction update indication may be the same or different, for example, an induction update indication corresponding to one text unit is-0.05, and an induction update indication corresponding to one text unit is 0.05. The intelligent inquiry response processing system can update the corresponding text units in the initial induction inquiry response interactive text by combining the induction update instructions respectively corresponding to the text units to obtain the optimized initial induction inquiry response interactive text.
Under some exemplary design ideas, the initial query response interaction text distinguishing variable can be calculated by introducing norms. The text distinguishing variable set values are different, and the obtained final induced query response interactive text is also different.
In the embodiment of the invention, the initial induction query response interactive text difference variable between the initial induction query response interactive text before and after updating and the basic induction query response interactive text corresponding to the initial induction query response interactive text after optimizing is determined, when the initial induction query response interactive text difference variable is smaller than the text difference variable set value, the step of carrying out question-answer interaction semantic mining on the initial induction query response interactive text is skipped to obtain the induction question-answer interaction semantic vector until the initial induction query response interactive text difference variable is larger than the text difference variable set value, and the initial induction query response interactive text is used as the final induction query response interactive text corresponding to the selected query response interactive text, so that the difference between the initial induction query response interactive text before and after updating and the basic induction query response interactive text can be adjusted, the initial induction query response interactive text before and after updating is more difficult to be distinguished from the basic induction query response interactive text, and the precision and quality of the final induction query response interactive text are improved.
Under some exemplary design ideas, the induction update instruction includes description update values corresponding to each text unit, and the updating of the text unit description values of the initial induction query response interaction text in combination with the induction update instruction includes: processing the selected interactive text offset index by combining the text unit description value of each text unit of the initial induction query response interactive text to obtain description update values respectively corresponding to each text unit in the initial induction query response interactive text; and updating the text unit description values of the text units in the initial induced query response interactive text by combining the description update values respectively corresponding to the text units to obtain the final induced query response interactive text corresponding to the selected query response interactive text.
The intelligent query response processing system may process the text unit description value of the text unit in the initial induced query response interaction text by using the selected interaction text offset index to obtain description update values corresponding to the text units respectively, update the text unit description value of the corresponding text unit by using the description update values corresponding to the text units to obtain the optimized initial induced query response interaction text, and obtain the final induced query response interaction text according to the optimized initial induced query response interaction text.
Under some exemplary design ideas, the intelligent query response processing system can process text unit description values of text units in the initial induced query response interactive text by using the selected interactive text offset index to obtain derivative results corresponding to the text units respectively, determine text unit update factors according to the derivative results, obtain text unit description value update step sizes, and determine multiplication operation results of the text unit update factors and the text unit update step sizes to obtain description update values corresponding to the text units. The text unit update step size is the maximum text unit description value which can be increased or decreased each time the text unit description value is updated, and can be preset, for example, can be 2, for example, the text unit description value of the text unit is 20, and the range of the optimized text unit description value is 18 to 22. The intelligent query response processing system can compare the derivative result with a set derivative value, and determine a text unit update factor according to the comparison result, wherein when the derivative result is larger than the set derivative value, the text unit update factor corresponding to the derivative result is determined to be a first variable value, when the derivative result is smaller than the set derivative value, the text unit update factor corresponding to the derivative result is determined to be a second variable value, the set derivative value can be adjusted based on design conditions, for example, the set derivative value can be 0, the first variable value is larger than the second variable value, for example, the first variable value is 1, and the second variable value is-1.
In the embodiment of the invention, the text unit description values of the text units in the initial induced query response interactive text are updated by combining the description update values respectively corresponding to the text units to obtain the final induced query response interactive text corresponding to the selected query response interactive text, so that the text unit description values of the text units are updated respectively, and the flexibility and the precision of updating the text unit description values are improved.
Under some exemplary design considerations, the method further comprises: loading the final induction inquiry response interaction text into an AI interaction text processing network to be debugged corresponding to the selected online dialogue topic to obtain an induction credibility coefficient corresponding to the final induction inquiry response interaction text in the selected online dialogue topic; determining a target network cost coefficient by combining the induction credibility coefficient; the target network cost coefficient and the induction credible coefficient have a first quantization relation; and updating network variables of the AI interactive text processing network by combining the target network cost coefficient to obtain the AI interactive text processing network after debugging.
The AI interactive text processing network to be debugged corresponding to the selected online dialogue theme can be understood as a pre-debugged AI interactive text processing network. The induced credibility coefficient is the credibility coefficient of the final induced query response interactive text generated by the AI interactive text processing network belonging to the selected online dialogue topic. Network variables can be understood as structural parameters within the network, and also as algorithm weights for AI algorithms.
The intelligent query response processing system can take the final induced query response interaction text as a negative training sample of the AI interaction text processing network to be debugged, debug the AI interaction text processing network to be debugged, obtain the real online dialogue subject as the selected query response interaction text of the selected online dialogue subject, take the selected query response interaction text as a positive training sample of the AI interaction text processing network to be debugged, and debug the AI interaction text processing network to be debugged by using the positive training sample and the negative training sample to obtain the AI interaction text processing network to be debugged. The number of positive training samples and the number of negative training samples in each debugging process can be determined based on requirements.
Under some exemplary design ideas, the intelligent query response processing system can debug the AI interactive text processing network by adopting a circularly debugged ideas, and the intelligent query response processing system can update network variables of the AI interactive text processing network according to a rule for reducing cost coefficients of a target network for circularly debugging until the network stability requirement is met, and the AI interactive text processing network after updating the network variables is used as the AI interactive text processing network for completing debugging. The network stability requirement may include that the fluctuation of the target network cost coefficient is less than a preset cost coefficient fluctuation value.
In the embodiment of the invention, the target network cost coefficient is determined by combining the induction credible coefficient, and in view of the first quantitative relation between the target network cost coefficient and the induction credible coefficient, when the target network cost coefficient is reduced, the induction credible coefficient is also reduced, namely the probability that the final induction query response interactive text is analyzed to be a selected online dialogue theme is reduced, so that the probability that the AI interactive text processing network which completes debugging analyzes the final induction query response interactive text to be the selected online dialogue theme can be reduced, the anti-interference performance of the AI interactive text processing network on induction text examples is improved, and the stability of the AI interactive text processing network is improved.
Under some exemplary design ideas, a query response interactive text parsing method is provided, comprising the following steps.
Step 602, obtaining a selected query response interaction text of a selected online dialogue topic and an initial induced query response interaction text corresponding to the induced online dialogue topic.
Step 604, performing question-answer interaction semantic mining on each selected query response interaction text to obtain selected question-answer interaction semantic vectors corresponding to each selected query response interaction text, obtaining an initial semantic thermodynamic relationship network, determining initial existence probabilities of each selected question-answer interaction semantic vector in the initial semantic thermodynamic relationship network, summarizing the initial existence probabilities corresponding to each selected question-answer interaction semantic vector to obtain an initial probability summarizing result, updating relationship network variables corresponding to the initial semantic thermodynamic relationship network to adjust the initial probability summarizing result, and obtaining a target semantic thermodynamic relationship network.
The target semantic thermodynamic relationship network may include a first target semantic thermodynamic relationship network and a second target semantic thermodynamic relationship network, where the first target semantic thermodynamic relationship network corresponds to a first target thermodynamic index, the second target semantic thermodynamic relationship network corresponds to a second target thermodynamic index, and the first target semantic thermodynamic relationship network and the second target semantic thermodynamic relationship network are combined to perform weight-based operation processing on the first target semantic thermodynamic relationship network and the second target semantic thermodynamic relationship network, so that the target semantic thermodynamic relationship network may be obtained, that is, the target semantic thermodynamic relationship network is a relationship network obtained by integrating the first target semantic thermodynamic relationship network and the second target thermodynamic relationship network.
And step 606, carrying out question-answer interaction semantic mining on the initial induced query response interaction text to obtain an induced question-answer interaction semantic vector.
Step 608, determining a first target existence probability of the induction question-answer interaction semantic vector in the first target semantic thermodynamic relationship network, obtaining a first correction probability by combining the first target thermodynamic index and the first target existence probability, determining a second target existence probability of the induction question-answer interaction semantic vector in the second target semantic thermodynamic relationship network, and obtaining a second correction probability by combining the second target thermodynamic index and the second target existence probability.
Step 610, selecting a semantic thermodynamic relationship net with the largest probability value from the first target semantic thermodynamic relationship net and the second target semantic thermodynamic relationship net by combining the first correction probability and the second correction probability, and using the semantic thermodynamic relationship net as a saliency semantic thermodynamic relationship net to acquire a target linear output field corresponding to the saliency semantic thermodynamic relationship net, and using the target linear output field as an auxiliary question-answer interaction semantic vector.
Step 612, determining semantic difference variables between the induction question-answer interaction semantic vector and the auxiliary question-answer interaction semantic vector, and combining the semantic difference variables to obtain a selected interaction text offset index.
Wherein the semantic difference variable has a first quantitative association with the selected interactive text offset index.
Step 614, determining a first trusted coefficient of the initial induced query response interaction text on the selected online dialogue topic by combining the induced question-answer interaction semantic vector, determining a second trusted coefficient of the initial induced query response interaction text on the induced online dialogue topic by combining the induced question-answer interaction semantic vector, and obtaining a second query response interaction text cost coefficient by combining a first trusted comparison result between the first trusted coefficient and the second trusted coefficient.
And the second query response interaction text cost coefficient is in second quantization relation with the first trusted comparison result.
Step 616, combining the first query response interactive text cost coefficient and the second query response interactive text cost coefficient to obtain a selected interactive text offset index, and combining the selected interactive text offset index to obtain an induced update indication.
Step 618, determining whether the second trusted comparison result is smaller than the trusted comparison setting value, and whether the initial query response interactive text distinction variable is smaller than the text distinction variable setting value, if so, implementing step 620, if not (i.e. the second trusted comparison result reaches the trusted comparison setting value or the initial query response interactive text distinction variable reaches the text distinction variable setting value), implementing step 622.
Step 620, updating the text unit description value of the initial induction query response interactive text in combination with the induction update instruction to obtain an updated initial induction query response interactive text, performing question-answer interaction semantic mining on the updated initial induction query response interactive text to obtain an updated induction question-answer interaction semantic vector, and jumping to step 612.
Step 622, the initial induced query response interaction text is used as the final induced query response interaction text corresponding to the selected query response interaction text.
The intelligent query response processing system can update the text unit description value of the initial induction query response interactive text in combination with the induction update instruction to obtain the optimized initial induction query response interactive text, determine the distinguishing variable between the optimized initial induction query response interactive text and the basic induction query response interactive text corresponding to the initial induction query response interactive text, and obtain the distinguishing variable of the initial query response interactive text.
Step 624, loading the final induction query response interaction text into the AI interaction text processing network to be debugged corresponding to the selected online dialogue topic to obtain the induction credibility coefficient of the final induction query response interaction text corresponding to the selected online dialogue topic.
The final induced query response interaction text may also be referred to as an induced text example.
Step 626, determining a target network cost coefficient by combining the induction credibility coefficient, updating network variables of the AI interactive text processing network by combining the target network cost coefficient to obtain a debugged AI interactive text processing network, and analyzing the query response interactive text by using the debugged AI interactive text processing network.
Wherein, the target network cost coefficient and the induced credibility coefficient have a first quantization relation.
In the embodiment of the invention, the induction text examples in the text detail vector coordinate systems of different layers of the AI interactive text processing network are limited at the same time, namely, the distribution of the induction text examples in the vector coordinate system is limited, so that the induction text examples are fitted in the relation network of the conventional training examples of the selected online dialogue theme, the AI interactive text processing network is debugged by utilizing the induction text examples, the anti-interference performance of the AI interactive text processing network on the induction text examples can be improved, and the stability of the AI interactive text processing network is improved.
In some independent embodiments, after updating the network variables of the AI-interactive text-processing network in conjunction with the target network cost coefficient to obtain a debug-completed AI-interactive text-processing network, the method further includes: loading a query response interaction text to be processed into the AI interaction text processing network completing debugging to obtain a target online dialogue theme of the query response interaction text to be processed, which is output by the AI interaction text processing network completing debugging; acquiring information to be pushed by utilizing the target online dialogue theme; and pushing the information to be pushed to the query client corresponding to the query response interaction text to be processed.
According to the embodiment of the invention, the AI interactive text processing network obtained by countermeasure debugging can accurately mine the target online dialogue topic of the query response interactive text to be processed, namely the online dialogue class of the query response interactive text to be processed, and on the basis, the information to be pushed can be accurately obtained based on the target online dialogue topic to carry out targeted data pushing, so that the utilization rate of limited resources is improved.
In some independent embodiments, the target online dialog topic is utilized to obtain information to be pushed, including steps 902-908.
Step 902, mining target user preference characteristics from the query response interaction text to be processed through the target online dialogue theme; a first query requirement description and a first push preference description of the target user preference feature are obtained, and a second query requirement description and a second push preference description of the reference user preference feature are obtained.
Step 904, determining whether a questioning keyword of the target user preference feature and a questioning keyword of the reference user preference feature are matched according to the first query requirement description and the second query requirement description, and determining whether a desired push mode of the target user preference feature and a desired push mode of the reference user preference feature are matched according to the first push preference description and the second push preference description.
Step 906, if the question key of the target user preference feature matches the question key of the reference user preference feature and the desired push pattern of the target user preference feature matches the desired push pattern of the reference user preference feature, determining that the target user preference feature matches the reference user preference feature.
Step 908, obtaining information to be pushed corresponding to the target user preference feature based on the determination mode of the reference push information corresponding to the reference user preference feature.
In some independent embodiments, determining whether the question key of the target user preference feature and the question key of the reference user preference feature match according to the first query requirement description and the second query requirement description in step 904 includes: determining a first matching weight between the first query requirement description and the second query requirement description; if the first matching weight is greater than a first set weight, determining that the questioning keyword of the target user preference feature is matched with the questioning keyword of the reference user preference feature;
further, determining whether the desired push pattern of the target user preference feature and the desired push pattern of the reference user preference feature match according to the first push preference description and the second push preference description in step 904 includes: determining a second matching weight between the first push preference description and the second push preference description; if the second matching weight is greater than a second set weight, determining that the desired push pattern of the target user preference feature matches the desired push pattern of the reference user preference feature.
Applying the above steps 902-908, determining whether the questioning keyword of the target user preference feature and the questioning keyword of the reference user preference feature match according to the first query requirement description of the target user preference feature and the second query requirement description of the reference user preference feature, and determining whether the desired push mode of the target user preference feature and the desired push mode of the reference user preference feature match according to the first push preference description of the target user preference feature and the second push preference description of the reference user preference feature; if the questioning keyword of the target user preference feature is matched with the questioning keyword of the reference user preference feature and the expected pushing mode of the target user preference feature is matched with the expected pushing mode of the reference user preference feature, the target user preference feature is determined to be matched with the reference user preference feature, so that whether the user preference feature is matched or not is determined by combining the questioning keyword of the user preference feature and the expected pushing mode, the accuracy of the user preference feature matching is improved, and the information to be pushed corresponding to the target user preference feature can be quickly and accurately obtained based on the determination mode of the reference pushing information corresponding to the reference user preference feature.
Based on the same or similar technical ideas described above, the embodiments of the present application also provide a computer-readable storage medium having a computer program stored thereon, the computer program executing an artificial intelligence based intelligent query response mutual information mining method at runtime.
Based on the same or similar technical conception, the embodiment of the application also provides a computer program product, which comprises a computer program or a computer executable instruction, wherein the computer program or the computer executable instruction realizes the intelligent query response interaction information mining method based on artificial intelligence when being executed by a processor.
Those of ordinary skill in the art will appreciate that the elements and algorithm steps described in connection with the embodiments disclosed herein may be embodied in electronic hardware, in computer software, or in a combination of the two, and that the elements and steps of the examples have been generally described in terms of function in the foregoing description to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.
The foregoing disclosure is merely illustrative of the presently preferred embodiments of the present application, and it is to be understood that the scope of the application is not limited thereto, but is intended to cover modifications as fall within the scope of the present application.
Claims (10)
1. An intelligent inquiry response interaction information mining method based on artificial intelligence is characterized by being applied to an intelligent inquiry response processing system, and comprises the following steps:
acquiring a selected query response interaction text of a selected online dialogue topic and an initial induced query response interaction text corresponding to an induced online dialogue topic;
performing question-answer interaction semantic mining on the selected query response interaction text to obtain a selected question-answer interaction semantic vector, and combining the selected question-answer interaction semantic vector to obtain an auxiliary question-answer interaction semantic vector;
carrying out question-answer interaction semantic mining on the initial induced query response interaction text to obtain an induced question-answer interaction semantic vector;
determining semantic difference variables between the induction question-answer interaction semantic vectors and the auxiliary question-answer interaction semantic vectors, and combining the semantic difference variables to obtain a selected interaction text offset index;
the semantic difference variable has a first quantization relation with the selected interactive text offset index;
And obtaining an induction updating instruction by combining the selected interactive text offset index, and updating a text unit description value of the initial induction inquiry response interactive text by combining the induction updating instruction to obtain a final induction inquiry response interactive text corresponding to the selected inquiry response interactive text.
2. The method of claim 1, wherein the plurality of selected query response interaction texts, the performing the question-answer interaction semantic mining on the selected query response interaction texts to obtain selected question-answer interaction semantic vectors, and the combining the selected question-answer interaction semantic vectors to obtain auxiliary question-answer interaction semantic vectors comprises:
performing question-answer interaction semantic mining on each selected query response interaction text to obtain selected question-answer interaction semantic vectors corresponding to each selected query response interaction text;
acquiring an initial semantic thermodynamic relationship network, and determining the initial existence probability of each selected question-answer interaction semantic vector in the initial semantic thermodynamic relationship network;
summarizing the initial existence probabilities corresponding to the selected question-answer interaction semantic vectors to obtain initial probability summarizing results, updating the relation network variables corresponding to the initial semantic thermodynamic relation network to adjust the initial probability summarizing results to obtain a target semantic thermodynamic relation network;
Obtaining a target linear output field corresponding to the target semantic thermodynamic relationship network to obtain the auxiliary question-answer interaction semantic vector;
the initial semantic thermodynamic relationship network comprises a first initial semantic thermodynamic relationship network and a second initial semantic thermodynamic relationship network, and the initial existence probability comprises a first initial existence probability corresponding to the first initial semantic thermodynamic relationship network and a second initial existence probability corresponding to the second initial semantic thermodynamic relationship network; summarizing the initial existence probabilities corresponding to the selected question-answer interaction semantic vectors to obtain initial probability summarizing results, updating the relation network variables corresponding to the initial semantic thermodynamic relation network to adjust the initial probability summarizing results, and obtaining a target semantic thermodynamic relation network comprises the following steps:
acquiring a first initial thermodynamic index corresponding to the first initial semantic thermodynamic relationship net and a second initial thermodynamic index corresponding to the second initial semantic thermodynamic relationship net;
combining the first initial heat index and the first initial existence probability, and calculating the second initial heat index and the second initial existence probability to obtain initial existence probability corresponding to the selected question-answer interaction semantic vector;
Summarizing initial existence probabilities corresponding to the selected question-answer interaction semantic vectors to obtain initial probability summarizing results, and updating relation network variables, the first initial heat index and the second initial heat index corresponding to a first initial semantic heat relation network and a second initial semantic heat relation network to adjust the initial probability summarizing results to obtain a first target semantic heat relation network, a second target semantic heat relation network, a first target heat index and a second target heat index;
the obtaining the target linear output field corresponding to the target semantic thermodynamic relationship network, and the obtaining the auxiliary question-answer interaction semantic vector comprises the following steps:
determining a first target existence probability of the induction question-answer interaction semantic vector in the first target semantic thermodynamic relationship network, and combining the first target thermodynamic index and the first target existence probability to obtain a first correction probability;
determining a second target existence probability of the induction question-answer interaction semantic vector in the second target semantic thermodynamic relationship network, and combining the second target thermodynamic index and the second target existence probability to obtain a second correction probability;
Selecting a semantic thermodynamic relationship network with the maximum probability value from the first target semantic thermodynamic relationship network and the second target semantic thermodynamic relationship network by combining the first correction probability and the second correction probability as a significance semantic thermodynamic relationship network;
and obtaining a target linear output field corresponding to the significance semantic thermodynamic relationship network as the auxiliary question-answer interaction semantic vector.
3. The method of claim 1, wherein said combining the semantic difference variable to obtain a selected interactive text offset index comprises:
combining the semantic difference variable to obtain a first query response interaction text cost coefficient, wherein the semantic difference variable has a first quantization relation with the first query response interaction text cost coefficient;
determining a first credibility coefficient of an initial induced query response interaction text on the selected online dialogue topic by combining the induced question response interaction semantic vector;
combining the first trusted coefficient to obtain a second query response interactive text cost coefficient;
the second query response interaction text cost coefficient is in second quantization relation with the first trusted coefficient;
combining the first query response interaction text cost coefficient and the second query response interaction text cost coefficient to obtain a selected interaction text offset index;
Wherein, the obtaining the second query response interaction text cost coefficient by combining the first trusted coefficient includes:
determining a second credibility coefficient of the initial induction inquiry response interaction text in the induction online dialogue topic by combining the induction inquiry response interaction semantic vector;
combining a first trusted comparison result between the first trusted coefficient and the second trusted coefficient to obtain a second query response interaction text cost coefficient; and a second quantization relation exists between the second query response interaction text cost coefficient and the first trusted comparison result.
4. The method of claim 1, wherein said combining the semantic difference variable to obtain a selected interactive text offset index comprises:
combining the semantic difference variable to obtain a first query response interaction text cost coefficient, wherein the semantic difference variable has a first quantization relation with the first query response interaction text cost coefficient;
determining the priori credibility coefficients of the initial induced query response interaction text on each priori online dialogue topic by combining the induced question response interaction semantic vectors;
selecting the largest trusted coefficient from the prior trusted coefficients as a third trusted coefficient;
Combining a second trusted comparison result between the first trusted coefficient and the third trusted coefficient to obtain a third query response interactive text cost coefficient; the third query response interaction text cost coefficient has a second quantization relation with the second trusted comparison result;
combining the first query response interaction text cost coefficient and the third query response interaction text cost coefficient to obtain a selected interaction text offset index;
the step of updating the text unit description value of the initial induced query response interaction text in combination with the induction updating instruction to obtain the final induced query response interaction text corresponding to the selected query response interaction text comprises the following steps:
when the second trusted comparison result is smaller than the trusted comparison set value, the text unit description value of the initial induced query response interactive text is updated in combination with the induced update instruction, and the updated initial induced query response interactive text is obtained;
and jumping to the step of carrying out question-answer interaction semantic mining on the initial induction query response interaction text to obtain induction question-answer interaction semantic vectors until the second trusted comparison result reaches a trusted comparison set value, and taking the initial induction query response interaction text as a final induction query response interaction text corresponding to the selected query response interaction text.
5. The method of claim 1, wherein the updating the text unit description value of the initial induced query response interaction text in combination with the induction update instruction to obtain the final induced query response interaction text corresponding to the selected query response interaction text comprises:
updating a text unit description value of the initial induction query response interactive text by combining the induction updating instruction to obtain an optimized initial induction query response interactive text;
determining an initial query response interaction text difference variable between the optimized initial induction query response interaction text and a basic induction query response interaction text corresponding to the initial induction query response interaction text, and jumping to a step of carrying out question-answer interaction semantic mining on the initial induction query response interaction text when the initial query response interaction text difference variable is smaller than a text difference variable set value to obtain an induction question-answer interaction semantic vector until the initial query response interaction text difference variable reaches the text difference variable set value, and taking the initial induction query response interaction text as a final induction query response interaction text corresponding to the selected query response interaction text.
6. The method of claim 1, wherein the induction update indication includes description update values corresponding to respective text units, and wherein the updating the text unit description values of the initial induction query response interaction text in combination with the induction update indication to obtain the final induction query response interaction text corresponding to the selected query response interaction text includes:
processing the selected interactive text offset index by combining text unit description values of all text units of the initial induction query response interactive text to obtain description update values corresponding to all the text units in the initial induction query response interactive text;
and updating the text unit description value of the text unit in the initial induced query response interaction text by combining the description update values respectively corresponding to the text units to obtain the final induced query response interaction text corresponding to the selected query response interaction text.
7. The method according to claim 1, wherein the method further comprises:
loading the final induction query response interaction text into an AI interaction text processing network to be debugged corresponding to the selected online dialogue topic to obtain an induction credibility coefficient of the final induction query response interaction text corresponding to the selected online dialogue topic;
Determining a target network cost coefficient by combining the induction credible coefficient; the target network cost coefficient is in first quantization connection with the induction credible coefficient;
and updating network variables of the AI interactive text processing network by combining the target network cost coefficient to obtain the AI interactive text processing network for completing debugging.
8. An intelligent query response processing system, comprising: a processor, a memory, and a network interface;
the processor is connected with the memory and the network interface;
the network interface is configured to provide data communication functionality, the memory is configured to store program code, and the processor is configured to invoke the program code to perform the artificial intelligence based intelligent query response mutual information mining method of any of claims 1-7.
9. A computer readable storage medium, having stored thereon a computer program which, when run, performs the artificial intelligence based intelligent query response mutual information mining method of any of claims 1-7.
10. A computer program product comprising a computer program or computer-executable instructions which, when executed by a processor, implement the artificial intelligence based intelligent query response mutual information mining method of any of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311208533.4A CN117131160A (en) | 2023-09-19 | 2023-09-19 | Intelligent query response interaction information mining method and system based on artificial intelligence |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311208533.4A CN117131160A (en) | 2023-09-19 | 2023-09-19 | Intelligent query response interaction information mining method and system based on artificial intelligence |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117131160A true CN117131160A (en) | 2023-11-28 |
Family
ID=88861185
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311208533.4A Pending CN117131160A (en) | 2023-09-19 | 2023-09-19 | Intelligent query response interaction information mining method and system based on artificial intelligence |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117131160A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118071007A (en) * | 2024-01-23 | 2024-05-24 | 重庆金石智诚科技有限公司 | Method and system for evaluating intelligent chromium plating treatment process parameters of luminous car logo |
-
2023
- 2023-09-19 CN CN202311208533.4A patent/CN117131160A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118071007A (en) * | 2024-01-23 | 2024-05-24 | 重庆金石智诚科技有限公司 | Method and system for evaluating intelligent chromium plating treatment process parameters of luminous car logo |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20230100376A1 (en) | Text sentence processing method and apparatus, computer device, and storage medium | |
| CN109033068B (en) | Method and device for reading and understanding based on attention mechanism and electronic equipment | |
| CN109635273B (en) | Text keyword extraction method, device, equipment and storage medium | |
| CN108920654B (en) | Question and answer text semantic matching method and device | |
| CN110598206A (en) | Text semantic recognition method and device, computer equipment and storage medium | |
| US20170193086A1 (en) | Methods, devices, and systems for constructing intelligent knowledge base | |
| CN114565104A (en) | Language model pre-training method, result recommendation method and related device | |
| CN113268609A (en) | Dialog content recommendation method, device, equipment and medium based on knowledge graph | |
| CN113723070B (en) | Text similarity model training method, text similarity detection method and device | |
| CN112131883B (en) | Language model training method, device, computer equipment and storage medium | |
| CN112085091B (en) | Short text matching method, device, equipment and storage medium based on artificial intelligence | |
| CN113569578B (en) | User intention recognition method and device and computer equipment | |
| CN111400340A (en) | Natural language processing method and device, computer equipment and storage medium | |
| KR20210043283A (en) | System and method for extracting knowledge based on machine reading | |
| CN115456421A (en) | Work order dispatching method and device, processor and electronic equipment | |
| CN111897943A (en) | Session record searching method and device, electronic equipment and storage medium | |
| CN116974554A (en) | Code data processing method, apparatus, computer device and storage medium | |
| CN117972033A (en) | Large model illusion detection method, device, computer equipment and storage medium | |
| CN110489730B (en) | Text processing method, device, terminal and storage medium | |
| CN117131160A (en) | Intelligent query response interaction information mining method and system based on artificial intelligence | |
| CN112084776A (en) | Similar article detection method, device, server and computer storage medium | |
| CN111507108B (en) | Alias generation method and device, electronic equipment and computer readable storage medium | |
| CN114925681B (en) | Knowledge graph question-answering question-sentence entity linking method, device, equipment and medium | |
| CN117828024A (en) | Plug-in retrieval method, device, storage medium and equipment | |
| CN113792131B (en) | Keyword extraction method and device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20231205 Address after: 755099 No. 234 Feng'an East Road, Shapotou District, Zhongwei City, Ningxia Hui Autonomous Region Applicant after: Gao Lei Address before: No. 83 Jiefang East Street, Xingqing District, Yinchuan City, Ningxia Hui Autonomous Region, 750000 Applicant before: Ningxia Xinke Chuangxiang Electronic Technology Co.,Ltd. |