CN116775846A - Domain knowledge question and answer method, system, equipment and medium - Google Patents

Abstract

The invention provides a domain knowledge question and answer method, system, equipment and medium, which belongs to the technical field of natural language processing. The method includes: obtaining a query statement; inputting the query statement into a semantic similarity model to obtain a query vector output by the semantic similarity model. ; The semantic similarity model is obtained after training based on domain knowledge; based on the query vector and at least one question vector in the knowledge base, the similarity between the query vector and each question vector is determined; based on the similarity between the query vector and each question vector, from At least one target vector is determined in each question vector; based on the answer corresponding to each target vector, the answer corresponding to the query statement is determined, which improves the accuracy of the matching results and the scalability of the question and answer system. Through vector matching, the question and answer system is improved response speed.

Description

Domain knowledge question and answer methods, systems, equipment and media

Technical field

The present invention relates to the technical field of natural language processing, and in particular to a domain knowledge question and answer method, system, equipment and medium.

Background technique

In intelligent operation and maintenance scenarios, operation and maintenance personnel often need to query operation and maintenance knowledge in their daily work, such as querying a certain operation and maintenance process, querying a certain operation and maintenance command, etc. Operation and maintenance robots with text matching functions can easily implement this function. The operation and maintenance personnel store the knowledge to be queried in the knowledge base of the operation and maintenance robot, and input the question statement to be queried through the dialogue interface. The robot analyzes and queries through a series of natural language algorithms (NLP), and provides the most suitable answer. Return to user.

Existing text matching technology can use text classification algorithms or text similarity matching algorithms. In terms of text classification algorithms, traditional classification algorithms include support vector machines and logistic regression; deep neural network algorithms include LSTM, TextCNN, etc. Text similarity matching algorithms are generally interactive.

The interactive text matching algorithm is very time-consuming to calculate. The classification algorithm is a supervised algorithm. Each training model needs to be labeled, which makes it difficult to operate and maintain the robot's knowledge expansion and real-time query. At the same time, the low quality of the training samples affects the question and answer matching. accuracy. Therefore, existing domain knowledge problem systems have problems such as slow response speed, low matching accuracy, and difficulty in expansion.

Contents of the invention

The present invention provides a field knowledge question and answer method, system, equipment and medium to solve the shortcomings of slow response speed, low matching accuracy and difficulty in expansion in the knowledge question system in the prior art, and achieve comprehensive improvement of the performance of the question and answer system. promote.

The present invention provides a domain knowledge question and answer method, including:

Get the query statement;

Input the query statement into the semantic similarity model to obtain the query vector output by the semantic similarity model; the semantic similarity model is obtained after training based on domain knowledge;

Based on the query vector and at least one question vector in the knowledge base, determine the similarity between the query vector and each of the question vectors;

Determine at least one target vector from each of the question vectors based on the similarity between the query vector and each of the question vectors;

Based on the answer corresponding to each target vector, the answer corresponding to the query statement is determined.

According to the domain knowledge question and answer method provided by the present invention, the semantic similarity model is trained based on the following steps:

Based on the domain knowledge dictionary, the Word2Vec model is weighted and trained through the neural network algorithm to obtain domain knowledge sentence vectors based on inverse document frequency weighting;

Expand based on the domain knowledge sentence vector to obtain domain knowledge sentence pairs;

The semantic similarity model is trained based on the domain knowledge sentence pairs.

According to the domain knowledge question and answer method provided by the present invention, the method further includes:

Perform word segmentation processing on at least one domain knowledge, and calculate the inverse document frequency weighted value of each domain knowledge after word segmentation;

The domain knowledge dictionary is constructed based on the inverse document frequency weighted value of each domain knowledge after word segmentation.

According to the domain knowledge question and answer method provided by the present invention, the domain knowledge sentence pairs are obtained by expanding based on the domain knowledge sentence vector, including:

Generate a sentence vector corresponding to the sentence vector through synonym replacement and/or sentence repetition;

Based on the sentence vector and the sentence vector, determine the domain knowledge sentence pair;

Calculate the similarity between the sentence vector and the sentence vector;

Based on the similarity, the domain knowledge sentence pairs are marked.

According to the domain knowledge question and answer method provided by the present invention, labeling the domain knowledge sentence pairs based on the similarity includes:

If the similarity is greater than or equal to the first preset threshold, mark the domain knowledge sentence pair as similar;

Alternatively, if the similarity is less than or equal to the second preset threshold, the domain knowledge sentence pair is marked as dissimilar.

According to the domain knowledge question and answer method provided by the present invention, before training the semantic similarity model based on the domain knowledge sentence pairs, the method further includes:

Based on Chinese prior knowledge, the semantic similarity model is preprocessed and trained.

According to the domain knowledge question and answer method provided by the present invention, the method further includes:

A real-time distributed search analysis engine component is used to perform distributed storage and retrieval of each of the question vectors in the knowledge base.

The present invention also provides a domain knowledge question and answer system, including:

Query statement acquisition module, used to obtain query statements;

A query vector acquisition module is used to input the query statement into the semantic similarity model to obtain the query vector output by the semantic similarity model; the semantic similarity model is obtained after training based on domain knowledge;

A similarity determination module, configured to determine the similarity between the query vector and each of the question vectors based on the query vector and at least one question vector in the knowledge base;

A target vector determination module, configured to determine at least one target vector from each of the question vectors based on the similarity between the query vector and each of the question vectors;

An answer determination module is used to determine the answer corresponding to the query statement based on the answer corresponding to each of the target vectors.

The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it realizes any of the above-mentioned domain knowledge. Question and answer method.

The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the field knowledge question and answer method as described in any of the above-mentioned methods is implemented.

The invention provides a domain knowledge question and answer method, system, equipment and medium. The obtained query statement is input into the semantic similarity model obtained after training based on domain knowledge, and the query vector corresponding to the query statement is obtained. Several pre-stored query vectors are obtained. Matching is performed in the knowledge base of question vectors, and several question vectors with the highest similarity to the query vector are selected as target vectors. The answer corresponding to the query statement is determined based on the target vector. The semantic similarity model obtained after training based on domain knowledge realizes the training Optimization of samples, the answer corresponding to the target vector obtained by matching the query statement and the question statement, using the semantic similarity model for domain knowledge question and answer, can improve the response speed of the question and answer system, because the training samples of the semantic similarity model are based on domain knowledge Optimized training was performed, thus improving the accuracy of the matching results of the question and answer system.

Description of drawings

In order to explain the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are of the present invention. For some embodiments of the invention, those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting creative efforts.

Figure 1 is one of the flow diagrams of the domain knowledge question and answer method provided by the present invention;

Figure 2 is the second schematic flow chart of the domain knowledge question and answer method provided by the present invention;

Figure 3 is one of the training schematic diagrams of the domain knowledge question and answer method provided by the present invention;

Figure 4 is the second training schematic diagram of the domain knowledge question and answer method provided by the present invention;

Figure 5 is the third schematic diagram of training of the domain knowledge question and answer method provided by the present invention;

Figure 6 is the third schematic flow chart of the domain knowledge question and answer method provided by the present invention;

Figure 7 is a schematic diagram of preprocessing of the domain knowledge question and answer method provided by the present invention;

Figure 8 is the fourth schematic flow chart of the domain knowledge question and answer method provided by the present invention;

Figure 9 is a schematic structural diagram of the domain knowledge question and answer system provided by the present invention;

Figure 10 is a schematic structural diagram of the electronic device provided by the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the technical solutions in the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

The following describes the domain knowledge question and answer method, system, equipment and medium of the present invention with reference to Figures 1-10.

Figure 1 is one of the flow diagrams of the domain knowledge question and answer method provided by the present invention. As shown in Figure 1, the method includes the following steps S1-S5:

S1. Get the query statement;

Optionally, in this step, the application scenario of the knowledge question and answer system in this field is the field of intelligent operation and maintenance. The user can input query statements on the interactive page of the front-end of the system, and obtain the query statements of the front-end page through the operation and maintenance robot equipment. The knowledge in this field The question and answer system is integrated into the knowledge base of the operation and maintenance robot equipment. Query statements generally include keywords in the field of operation and maintenance, which can be to query the process of a certain operation and maintenance operation, the definition of a certain operation and maintenance term, or to query a certain operation and maintenance command, such as "What is the execution command of oos-job?" ", "What are the common commands of redis?".

S2. Input the query statement into the semantic similarity model and obtain the query vector output by the semantic similarity model;

The semantic similarity model is obtained after training based on domain knowledge;

Optionally, in this embodiment, the semantic similarity model is obtained through pre-training and fine-tuning, and adopts a supervised training method. The supervised training method means that the training samples have settings In this embodiment, the labels are divided into "0" and "1", which represent whether the similarity between the input training samples reaches the set threshold. The model is trained and fine-tuned based on domain knowledge, making the model's vector transformation of operation and maintenance domain knowledge more accurate.

S3. Based on the query vector and at least one question vector in the knowledge base, determine the similarity between the query vector and each question vector;

Optionally, in this step, several question vectors are stored in the knowledge base. These question vectors are the question vectors corresponding to the question statements obtained after passing through the above semantic similarity model and stored in the knowledge base. After receiving the query statement input by the user, the query statement will go through the semantic similarity model to obtain the query vector corresponding to the query statement, and then compare the query vector with several question vectors in the knowledge base one by one. The comparison method is calculation The similarity between the query vector and each question vector can be calculated using cosine similarity, Jaccard similarity coefficient, Pearson correlation coefficient, etc.

S4. Based on the similarity between the query vector and each question vector, determine at least one target vector from each question vector;

Optionally, in this step, by calculating the similarity between the query vector and each question vector in the knowledge base respectively, matching and outputting TOPN target vectors according to the similarity from high to low; calculating the similarity between the query vector and each question vector in the knowledge base The similarity is sorted from high to low, and those greater than or equal to the similarity setting threshold are identified as target vectors. In this embodiment, the similarity setting threshold can be set to 80%. Use TOPN question vectors as target vectors. If there is only one question vector that has a similarity greater than or equal to the set threshold with the query vector after matching in the knowledge base, only one target vector will be output. In specific implementation, if the query vector corresponding to the question statement input by the user matches the question vector stored in the knowledge base with a very low similarity, that is, the similarity is lower than the set similarity threshold, then no The target vector is returned, and "No answer to the question you queried was found" is displayed on the Q&A interface.

S5. Based on the answers corresponding to each target vector, determine the answer corresponding to the query statement.

In the knowledge base, the answer string corresponding to the question statement is also stored in the form of a vector, which is a one-to-one mapping relationship. After the query vector and the question vector are successfully matched, the system will query the answers corresponding to the question vector in a distributed manner in the knowledge base. , return it to the operation and maintenance robot and input it to the front-end interactive page for highlighting.

The domain knowledge question and answer method provided in this embodiment inputs the obtained query statement into the semantic similarity model obtained after training based on domain knowledge, and obtains the query vector corresponding to the query statement. In the knowledge base that stores several question vectors in advance, Matching is performed, and several question vectors with the highest similarity to the query vector are selected as target vectors, and the answer corresponding to the query statement is determined based on the target vector. The semantic similarity model obtained after training based on domain knowledge realizes the optimization of the training samples, and the knowledge base After passing through the semantic similarity model, the knowledge question statements in the database and the client's query statements will get their corresponding vectors. The query statements and question statements are matched through vector similarity calculation, and the semantic similarity model is used for domain knowledge question and answer. It can improve the response speed of the question and answer system. Since the training samples of the semantic similarity model are optimized and trained based on domain knowledge, the accuracy of the matching results of the question and answer system is improved.

Based on the above embodiment, further, the semantic similarity model in the above step S2 is trained based on the following steps. Figure 2 is the second flow diagram of the domain knowledge question and answer method provided by the present invention. As shown in Figure 2, the above semantics The training steps of the similarity model include the following steps S21-S23:

S21. Based on the domain knowledge dictionary, the Word2Vec model is weighted and trained through the neural network algorithm to obtain the domain knowledge sentence vector based on inverse document frequency weighting;

Optionally, in this embodiment, the Word2vec model is a related model used to generate word vectors. The Word2Vec model can be used to map each word to a vector, which can be used to express the relationship between words and represent distributed word vectors. The Word2Vec model can be based on the characteristics of the language that "a word can be described by the words around it". Study large data sets to get a comprehensive description of the word.

In this embodiment, Word2Vec is trained through a neural network algorithm to obtain a vector representation corresponding to each word or a single word. After addition and averaging, the sentence vector of this embodiment is obtained.

Figure 3 is one of the training schematic diagrams of the domain knowledge question and answer method provided by the present invention. As shown in Figure 3, in this embodiment, the Word2Vec model is trained using the Continuous Bag of Words (CBOW) model. The training features are trained using a common method based on words and characters. The vector length of each word or character is set to 300 dimensions. The input of the CBOW model is a string of text. Usually the length of the text is set to 5. It can also be based on Changes will be made on a case-by-case basis. As shown in Figure 3, w(t) represents the word or character in each sample. The CBOW type relies on the context w(t-2), w(t-1), w(t+1), w(t+2) calculates the word or word w(t). The result of this calculation is that similar words in the training sample will have similar representations in the vector. The basic idea of CBOW is to predict the word itself given the context of a word. For example, for the sentence "Thecat climbed up the tree", if the text length is 5, then when the center word is "climbed", the context words are "The", "cat", "up" and "the". The CBOW model requires calculating the probability distribution of "climbed" based on these four context words. Compared with traditional counting-based or matrix decomposition-based methods, the CBOW model can use large-scale corpora for training, and can learn high-quality and low-dimensional dense vectors, usually between 50 and 300, saving storage space and computing resources. , and can capture complex relationships between words, such as synonyms, antonyms, analogy relationships, etc.

Figure 4 is the second training schematic diagram of the domain knowledge question and answer method provided by the present invention. As shown in Figure 4, when calculating the sentence vector representation, the corresponding word vectors can be directly obtained for the words existing in the domain knowledge dictionary. For the domain knowledge dictionary For words that are not included in , the sentence vector is obtained by obtaining the vector of each word of the word, adding and averaging. As shown in Figure 4, the length of the vector corresponding to each knowledge is 300 dimensions. After the step of multiplying with the corresponding IDF value and the addition and averaging step, the vector length of the obtained sentence vector is also 300 dimensions.

If the word vectors generated by Word2Vec are added and averaged to obtain a sentence vector, it means that the algorithm considers each word in the sentence to have equal weight, which will lead to poor matching between the query vector and the question vector. For example, if the user inputs a query sentence of, "What are the common commands of redis?", if each word in the sentence has equal weight, then in addition to the keyword "redis" in the operation and maintenance field, the remaining problem statement sentence "What are the common commands of redis?" The similarity of ",,Commonly Used,Commands" is 100%. The algorithm will match "What common webgate commands are there" in the first position, and "What are the redis commands" will be matched in the second position. At this time, the Inverse Document Frequency (IDF) value is introduced to increase the weight of keywords in the operation and maintenance field such as "redis" and "webgate", and reduce the weight of the words "commonly used" and "command" to optimize the sorting results. It should be noted that when in actual use, if a word is not in the domain knowledge dictionary, we count the first four digits of all inverse document frequency (IDF) values in the dictionary as the inverse document frequency (IDF) value of the unincluded word, that is By default, uncollected words appear less frequently and their weight should be slightly larger.

S22. Expand based on the domain knowledge sentence vector to obtain domain knowledge sentence pairs;

In this step, based on the domain knowledge sentence vector obtained in step S21, a domain knowledge sentence pair is created to prepare for subsequent model training.

Optionally, in this step, the acquired domain knowledge sentence vector is expanded, which actually creates a sentence similar to the domain knowledge sentence vector, which can also be called a sentence vector. The sentence vector and the expanded sentence vector together constitute Domain knowledge sentence pairs.

S23. Based on domain knowledge sentence pairs, train the semantic similarity model.

In this embodiment, the semantic similarity model is trained using pre-training and fine-tuning. The basic model uses the SBert model (Siamese network Bert) with a twin network architecture. The subnetworks of the SBert model all use the BERT model, and the two BERT models share parameters. When comparing the similarity of two sentences, in this embodiment, the input domain knowledge sentence pairs are input into the BERT network respectively, and the output is two sets of vectors representing the sentences. A supervised training method is used. The supervised training method here means that the input training samples are labeled.

Figure 5 is the third training schematic diagram of the domain knowledge question and answer method provided by the present invention. As shown in Figure 5, the twin network (SBert) architecture has two input channels, and the input is the domain knowledge sentence pair obtained in step S22, that is, the sentence Vectors and similarly expanded sentence vectors. The two channels share the same semantic representation model (BERT). After processing by the pooling layer, the vectors u and v corresponding to the sentence vector and the expanded sentence vector are output, respectively, and then cosine similarity is performed. calculate. The loss function during the training process uses COS similarity or L2 distance algorithm.

The domain knowledge question and answer method provided in this embodiment trains the semantic similarity model used. Based on the domain knowledge dictionary, the Word2Vec model is weighted and trained through the neural network algorithm to obtain the domain knowledge sentence vector based on inverse document frequency weighting, which is convenient and effective. It solves the problem of low matching accuracy caused by the equal weight of words in Word2Vec, and improves the representation effect of sentence vectors.

Based on the above embodiment, further, the steps for constructing the domain knowledge dictionary in the above step S21 are as follows. Figure 6 is the third flow diagram of the domain knowledge question and answer method provided by the present invention. As shown in Figure 6, the steps for constructing the domain knowledge dictionary are as follows. Including the following steps S211-S212:

S211. Perform word segmentation processing on at least one domain knowledge, and calculate the inverse document frequency weighted value of each domain knowledge after word segmentation;

The constructed domain knowledge dictionary is the operation and maintenance domain knowledge dictionary. The construction of a knowledge dictionary in this field is to prepare for more accurate word segmentation in the future. Most of the contents in the dictionary belong to the computer category, such as "hbase", "hive", "html5", "css", etc. Other types in the dictionary belong to the dictionary of operation and maintenance proper nouns. After segmenting all operation and maintenance knowledge into words, calculate the inverse document frequency (IDF) value of each word. For example, the operation and maintenance knowledge "What is the execution command of oos-job", after word segmentation, it shows "oos-job, the execution command, is, what". At this time, because "oos-job" exists in the dictionary, the word segmenter can segment the word completely and accurately. Then calculate the inverse document frequency (IDF) value of all words, the calculation formula is as formula (1):

N represents the total number of knowledge in all fields in the knowledge base, DF represents the number of knowledge containing the word, and the function of adding 1 to the denominator is to prevent the situation where DF=0. It is easy to see that the more times the word appears, the larger the denominator is, and the smaller the value after taking the logarithm is, indicating that the weight of this word in all knowledge can be understood as the smaller the importance. For example, in "oos-job, execution, command, is, what", oos-job has fewer knowledge entries than other words, so its IDF value will be larger than other words. The purpose of obtaining the value of IDF is to assign different weights to the word vectors of different words in the above step S21.

S212. Construct a domain knowledge dictionary based on the inverse document frequency weighted values of each domain knowledge after word segmentation.

Figure 7 is a preprocessing schematic diagram of the domain knowledge question and answer method provided by the present invention. As shown in Figure 7, it is a constructed domain knowledge dictionary, including the domain knowledge after word segmentation and the corresponding IDF value.

The domain knowledge question and answer method provided in this embodiment improves the quality of the semantic similarity model training input samples in subsequent steps by constructing an operation and maintenance domain knowledge dictionary, including the domain knowledge after word segmentation and the corresponding IDF value, and improves the matching of the question and answer system. accuracy.

Based on the above embodiment, further, FIG. 8 is a flow diagram 4 of the domain knowledge question and answer method provided by the present invention. As shown in FIG. 8, the execution steps of the above step S22 include the following steps S221-S224:

S221. Generate a sentence vector corresponding to the sentence vector through synonym replacement and/or sentence repetition;

In this step, the obtained sentence vector is expanded by synonym replacement and/or sentence repetition, which is completed automatically using thesaurus and python script. The corresponding sentence vector is generated from the sentence vector as shown in Table 1:

Table 1

sentence vector What are the common commands of redis? synonym replacement What are the common commands for redis? What are the common commands for redis? word repetition What are the commonly used commands for redis? What are the commonly used commands for redis?

Synonym replacement and statement repetition can be performed at the same time, or one of them can be performed. For example, in the example, the original sentence vector "What are the common commands of redis" is expanded, and "which" is replaced with "what" through the synonym library. , or replace "commonly used" with "idiomatic". The meaning of the sentence vector represented in this way is the same as the original sentence vector, and the keyword "redis" in the operation and maintenance field has not been replaced.

S222. Based on sentence vectors and sentence vectors, determine domain knowledge sentence pairs;

S223. Calculate the similarity between the sentence vector and the sentence vector;

Optionally, in this step, COS cosine similarity is used to calculate the similarity between the sentence vector and the sentence vector.

S224. Based on the similarity, label the domain knowledge sentence pairs.

The domain knowledge question and answer method provided in this embodiment can automatically generate high-quality extended operation and maintenance field questions in batches by using the thesaurus and python scripts to automatically generate sentence pairs through synonym replacement and/or sentence repetition for the obtained sentence vectors. Statements, in the subsequent optimization of the model, you only need to pay attention to the keyword updates in the operation and maintenance domain knowledge dictionary. The original domain knowledge statements are automatically expanded, improving the scalability and usability of the question and answer system.

Based on the above embodiment, further, the above step S224 also includes:

If the similarity is greater than or equal to the first preset threshold, the domain knowledge sentence pair is marked as similar;

Alternatively, if the similarity is less than or equal to the second preset threshold, mark the domain knowledge sentence pair as dissimilar.

Optionally, in this embodiment, the sentence pairs are marked for supervised training. The COS similarity calculates the similarity between the sentence pairs. If the similarity is greater than or equal to the first preset threshold, this implementation In the example, the first preset threshold is 95%, then the domain knowledge sentence pairs are marked as similar, and the label can be set to 1; if the similarity is less than or equal to the second preset threshold, the second preset threshold in this example is 30 %, then mark the domain knowledge sentence pairs as dissimilar, set the label to 0, and the domain knowledge sentence pairs after labeling are shown in Table 2:

Table 2

Label sentence vector statement vector Similar(1) What are the common commands of redis? What are the commands of redis? Not similar(0) What are the common commands of redis? What is the execution command of oos-job?

The domain knowledge question and answer method provided in this embodiment facilitates the supervised training method in subsequent model training by labeling domain knowledge sentence pairs. The input of domain knowledge sentence pairs with similarity labels improves the semantic similarity model training input samples. The quality improves the matching accuracy of the question and answer system.

Based on the above embodiment, further, before step S23, it also includes:

Based on Chinese prior knowledge, the semantic similarity model is preprocessed and trained.

This step is the pre-training stage of the semantic similarity model. First, Chinese semantic text similarity STS (Semantic Textual Similarity) data sets in various fields are selected for training. This data set covers many fields, including gossip, finance, novels, sports, etc. The purpose of pre-training is to hope that the SBert model can first learn extensive semantic knowledge in the entire Chinese field. This pre-training stage will include sentence pairs and the Pearson correlation coefficient between sentence pairs. The Pearson correlation coefficient value generally ranges from 0-5. , the larger the score, the more similar the two statements are. For example, the Pearson correlation coefficient between the statement pair "a child is riding a horse" and "the child is riding a horse" is 4, while the statement pair "a woman plays the guitar" and " The Pearson correlation coefficient between "a person reading a book" is 2. The pre-training of the semantic similarity model is helpful for the subsequent fine-tuning of knowledge in the operation and maintenance field. After multiple model training and testing, when the data set selects 100,000 statement pairs and the Bert model parameter dropout_rate=0.2, the optimal model can be obtained. During training, we freeze several initial network layers of Bert to retain the pre-training to the greatest extent. Stage prior knowledge of Chinese.

The domain knowledge question and answer method provided in this embodiment optimizes the training basis of the model by pre-training the semantic similarity model based on Chinese prior knowledge, improves the effect of subsequent operation and maintenance domain knowledge fine-tuning, and improves the final semantic similarity model. output accuracy.

Based on the above embodiment, further, the domain knowledge question and answer method also includes:

A real-time distributed search analysis engine component is used to perform distributed storage and retrieval of each of the question vectors in the knowledge base.

The domain knowledge question and answer method provided in this embodiment uses the time-distributed search analysis engine (ElasticSearch) component to build a high-performance vector retrieval library. After obtaining the SBert model, we obtain the vectors of all knowledge question statements through the model and store them in In ES, ES introduced the dense_vector index type in version 7.3 or above, which supports the calculation of COS similarity between vectors. The domain knowledge question and answer method in this embodiment tested the vector retrieval performance of single-node ES and cluster ES. In this embodiment, a single node and a vector length of 300 dimensions are used as an example for testing. In the specific implementation, the 3-node optimized ES cluster can retrieve similar operation and maintenance knowledge from 1 million pieces of data in 100 milliseconds. On the basis of ensuring the text matching accuracy of the model, it improves the rapid response of the question and answer system. Retrieval capabilities.

The present invention also provides a domain knowledge question and answer system. Figure 9 is a schematic structural diagram of the domain knowledge question and answer system provided by the present invention. As shown in Figure 9, the domain knowledge question and answer system includes:

Query statement acquisition module 91, used to obtain query statements;

The query vector acquisition module 92 is used to input the query statement into the semantic similarity model to obtain the query vector output by the semantic similarity model; the semantic similarity model is obtained after training based on domain knowledge;

The similarity determination module 93 is configured to determine the similarity between the query vector and each of the question vectors based on the query vector and at least one question vector in the knowledge base;

A target vector determination module 94, configured to determine at least one target vector from each of the question vectors based on the similarity between the query vector and each of the question vectors;

The answer determination module 95 is used to determine the answer corresponding to the query statement based on the answer corresponding to each of the target vectors.

In the domain knowledge question and answer system of this embodiment, through the cooperation between various modules, the query statement is obtained through the query statement acquisition module, input into the semantic similarity model obtained after training based on domain knowledge, and the query statement is obtained through the query vector acquisition module The corresponding query vectors are matched in a knowledge base that stores several question vectors in advance, and the question vectors with the highest similarity to the query vectors are selected as target vectors. Based on the target vectors, the answers corresponding to the query statements are determined. After training based on domain knowledge The obtained semantic similarity model optimizes the training samples. After passing through the semantic similarity model, the knowledge question statements in the knowledge base and the client query statements will get their corresponding vectors. The query statements are compared with each other through vector similarity calculation. Matching question statements and using the semantic similarity model for domain knowledge question and answer can improve the response speed of the question and answer system. Since the training samples of the semantic similarity model are optimized and trained based on domain knowledge, the matching results of the question and answer system are improved. Accuracy.

Figure 10 is a schematic structural diagram of an electronic device provided by the present invention. As shown in Figure 10, the electronic device may include: a processor (processor) 1010, a communications interface (Communications Interface) 1020, a memory (memory) 1030 and a communication bus 1040. Among them, the processor 1010, the communication interface 1020, and the memory 1030 complete communication with each other through the communication bus 1040. The processor 1010 can call logical instructions in the memory 1030 to execute a domain knowledge question and answer method. The method includes: obtaining a query statement; inputting the query statement into a semantic similarity model to obtain a query vector output by the semantic similarity model. ; The semantic similarity model is obtained after training based on domain knowledge; Based on the query vector and at least one question vector in the knowledge base, determine the similarity between the query vector and each of the question vectors based on the query vector Based on the similarity with each of the question vectors, at least one target vector is determined from each of the question vectors; based on the answer corresponding to each of the target vectors, the answer corresponding to the query statement is determined. In addition, the above-mentioned logical instructions in the memory 1030 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code. .

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. The computer program is implemented when executed by a processor to execute the domain knowledge question and answer method provided by each of the above methods. The method includes : Obtain the query statement; input the query statement into the semantic similarity model to obtain the query vector output by the semantic similarity model; the semantic similarity model is obtained after training based on domain knowledge; based on the query vector and at least one question vector in the knowledge base, determining the similarity between the query vector and each of the question vectors; based on the similarity between the query vector and each of the question vectors, determining at least one target from each of the question vectors Vector; determine the answer corresponding to the query statement based on the answer corresponding to each target vector.

The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in One location, or it can be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Persons of ordinary skill in the art can understand and implement the method without any creative effort.

Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the part of the above technical solution that essentially contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., including a number of instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be used Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent substitutions are made to some of the technical features; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A domain knowledge question and answer method, characterized by including: Get the query statement; Input the query statement into the semantic similarity model to obtain the query vector output by the semantic similarity model; the semantic similarity model is obtained after training based on domain knowledge; Based on the query vector and at least one question vector in the knowledge base, determine the similarity between the query vector and each of the question vectors; Determine at least one target vector from each of the question vectors based on the similarity between the query vector and each of the question vectors; Based on the answer corresponding to each target vector, the answer corresponding to the query statement is determined. 2. The domain knowledge question and answer method according to claim 1, characterized in that the semantic similarity model is trained based on the following steps: Based on the domain knowledge dictionary, the Word2Vec model is weighted and trained through the neural network algorithm to obtain domain knowledge sentence vectors based on inverse document frequency weighting; Expand based on the domain knowledge sentence vector to obtain domain knowledge sentence pairs; The semantic similarity model is trained based on the domain knowledge sentence pairs. 3. The domain knowledge question and answer method according to claim 2, characterized in that the method further includes: Perform word segmentation processing on at least one domain knowledge, and calculate the inverse document frequency weighted value of each domain knowledge after word segmentation; The domain knowledge dictionary is constructed based on the inverse document frequency weighted value of each domain knowledge after word segmentation. 4. The domain knowledge question and answer method according to claim 2, characterized in that the domain knowledge sentence pair is obtained by expanding based on the domain knowledge sentence vector, including: Generate a sentence vector corresponding to the sentence vector through synonym replacement and/or sentence repetition; Based on the sentence vector and the sentence vector, determine the domain knowledge sentence pair; Calculate the similarity between the sentence vector and the sentence vector; Based on the similarity, the domain knowledge sentence pairs are marked. 5. The domain knowledge question and answer method according to claim 4, characterized in that, based on the similarity, labeling the domain knowledge sentence pairs includes: If the similarity is greater than or equal to the first preset threshold, mark the domain knowledge sentence pair as similar; Alternatively, if the similarity is less than or equal to the second preset threshold, the domain knowledge sentence pair is marked as dissimilar. 6. The domain knowledge question and answer method according to claim 2, characterized in that, before training the semantic similarity model based on the domain knowledge sentence pairs, the method further includes: Based on Chinese prior knowledge, the semantic similarity model is preprocessed and trained. 7. The domain knowledge question and answer method according to claim 2, characterized in that the method further includes: A real-time distributed search analysis engine component is used to perform distributed storage and retrieval of each of the question vectors in the knowledge base. 8. A domain knowledge question and answer system, characterized by including: Query statement acquisition module, used to obtain query statements; A query vector acquisition module is used to input the query statement into the semantic similarity model to obtain the query vector output by the semantic similarity model; the semantic similarity model is obtained after training based on domain knowledge; A similarity determination module, configured to determine the similarity between the query vector and each of the question vectors based on the query vector and at least one question vector in the knowledge base; A target vector determination module, configured to determine at least one target vector from each of the question vectors based on the similarity between the query vector and each of the question vectors; An answer determination module is used to determine the answer corresponding to the query statement based on the answer corresponding to each of the target vectors. 9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that when the processor executes the program, it implements claim 1 The domain knowledge question and answer method described in any one of to 7. 10. A non-transitory computer-readable storage medium with a computer program stored thereon, characterized in that when the computer program is executed by a processor, the field knowledge question and answer method as claimed in any one of claims 1 to 7 is implemented. .