CN116910189A - Atlas question-answering method, apparatus, system and medium - Google Patents

Abstract

The present invention provides a graph question and answer method, device, system and medium. The method includes: calculating the similarity between a question vector and a candidate vector to determine a target vector in the candidate vector, wherein the question vector is A vector obtained by vector conversion based on user questions, the candidate vector is a vector set obtained by vector conversion based on a cut triple set, and the similarity between the target vector and the question vector is greater than other candidate vectors. The similarity between the candidate vector and the question vector; answering the user question according to the triplet corresponding to the target vector. In this way, when calculating the similarity, the interference of the answer part in the triplet on the question and answer is reduced, and the accuracy of the graph question and answer system in answering user questions is improved.

Description

A graph question and answer method, device, system and medium

Technical field

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

Background technique

In intelligent question answering, methods based on semantic analysis (Semantic Parser, SP) and methods based on information retrieval (Information Retrieval, IR) are usually used.

However, methods based on semantic analysis rely on predefined rule templates, require a large amount of manual annotation data, and have low question and answer accuracy without training data; methods based on information retrieval require syntactic analysis and entity extraction, among which , Existing pre-trained entity extraction models only support the extraction of a few fixed types of entities, such as names of people, place names, places, dates, etc., and the extraction effect is not good in small samples of vertical domain knowledge.

It can be seen that the graph question answering system in the existing technology has a problem of low question answering accuracy.

Contents of the invention

Embodiments of the present invention provide a graph question and answer method, device, system and medium to solve the problem of low question and answer accuracy in the prior art.

In a first aspect, embodiments of the present invention provide a graph question and answer method, which method includes:

Calculate the similarity between the question vector and the candidate vector to determine the target vector in the candidate vector, wherein the question vector is a vector obtained by vector conversion based on the user question, and the candidate vector is based on the cut three-dimensional vector. A set of vectors obtained by vector conversion of a tuple set, where the similarity between the target vector and the problem vector is greater than the similarity between other candidate vectors in the candidate vectors and the problem vector;

Answer the user question according to the triplet corresponding to the target vector.

Optionally, before calculating the similarity between the question vector and the candidate vector, the method further includes:

Construct a knowledge graph based on the user graph data to obtain a triplet set, where each triplet in the triplet set includes three elements;

Cut each triplet in the triplet set into three tuples to obtain a cut triplet set, and each triplet in the cut triplet set includes two elements;

The cut triple set is vector converted to obtain the candidate vector.

Optionally, performing vector conversion on the cut triple set to obtain the candidate vector includes:

The cut triple set is vector converted and then clustered to obtain index vectors. Each index vector is the clustering center of the candidate vectors it contains. The number of candidate vectors is greater than the number of the index vectors. quantity;

The calculation of the similarity between the question vector and the candidate vector to determine the target vector in the candidate vector includes:

Calculate the similarity between the question vector and the index vector to determine a target index vector in the index vector, the similarity between the target index vector and the question vector is greater than that of other index vectors in the index vector and the question Similarity of vectors;

Among the candidate vectors included in the target index vector, the similarity between the question vector and the candidate vector is calculated to determine the target vector.

Optionally, before calculating the similarity between the question vector and the candidate vector, performing vector conversion on the cut triple set, and after obtaining the candidate vector, the method further includes:

The candidate vectors are pre-stored locally or in a database, the database is associated with a computing device, and the computing device is used to calculate the similarity between the question vector and the candidate vector.

Optionally, before calculating the similarity between the question vector and the candidate vector, the method further includes:

Train a vector conversion model, which is used to perform vector conversion based on user questions to obtain question vectors, and perform vector conversion based on the cut triple set to obtain candidate vectors;

Wherein, the vector conversion model is a pre-trained language model RoFormer-Sim, and the normalization algorithm in the RoFormer-Sim is changed to the root mean square value RMS algorithm.

In a second aspect, embodiments of the present invention provide a graph question and answer device, including:

The calculation module is used to calculate the similarity between the question vector and the candidate vector to determine the target vector in the candidate vector, wherein the question vector is a vector obtained by vector conversion based on the user question, and the candidate vector is A vector set obtained by performing vector conversion based on the cut triple set, where the similarity between the target vector and the problem vector is greater than the similarity between other candidate vectors in the candidate vectors and the problem vector;

An answering module is used to answer the user question according to the triplet corresponding to the target vector.

Optionally, the device also includes:

A building module, used to construct a knowledge graph based on user graph data to obtain a triplet set, where each triplet in the triplet set includes three elements;

A cutting module, used to cut each triplet in the triplet set into three tuples to obtain a cut triplet set. Each triplet in the cut triplet set is Includes two elements;

A conversion module, configured to perform vector conversion on the cut triple set to obtain the candidate vector.

Optionally, the conversion module includes:

The clustering submodule is used to perform vector conversion on the cut triple set and perform clustering to obtain index vectors. Each index vector is the clustering center of the candidate vectors it contains. The number is greater than the number of index vectors;

The computing module includes:

The first calculation sub-module is used to calculate the similarity between the question vector and the index vector to determine the target index vector in the index vector, and the similarity between the target index vector and the question vector is greater than the index vector The similarity between other index vectors in and the problem vector;

The second calculation submodule is used to calculate the similarity between the question vector and the candidate vector among the candidate vectors included in the target index vector, and determine the target vector.

In a third aspect, embodiments of the present invention provide a graph question and answer system, including:

at least one processor; and

a memory communicatively connected to the at least one processor; wherein,

The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

In a fourth aspect, embodiments of the present invention provide a non-transitory computer-readable storage medium storing computer instructions, the computer instructions being used to cause the computer to execute the method described in the first aspect.

In the embodiment of the present invention, the similarity between the question vector obtained by vector conversion based on the user's question and the candidate vector obtained by vector conversion based on the cut triple set is calculated to determine the target vector, and then based on the target vector corresponding The missing part in the cut triplet or the spliced sentence of the triplet corresponding to the target vector is output as the answer to the user's question. Before calculating the similarity between the question vector and the candidate vector, each triplet in the triplet set is cut to obtain the cut triplet set. In this way, the triplet set is reduced when calculating the similarity. The interference of the answer part in the group to the question and answer improves the accuracy of the graph question and answer system in answering user questions.

Description of the drawings

In order to explain the technical solutions of the embodiments of the present invention 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 only some of the drawings of the present invention. Embodiments, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without exerting creative efforts.

Figure 1 is one of the flow charts of the graph question and answer method provided by the embodiment of the present invention;

Figure 2 is the second flow chart of the graph question and answer method provided by the embodiment of the present invention;

Figure 3 is a structural diagram of a graph question and answer device provided by an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of the present invention.

Referring to Figure 1, Figure 1 is one of the flow charts of the graph question and answer method provided by an embodiment of the present invention. As shown in Figure 1, the method includes the following steps:

Step 101: Calculate the similarity between the question vector and the candidate vector to determine the target vector in the candidate vector, wherein the question vector is a vector obtained by vector conversion based on the user question, and the candidate vector is based on cutting A vector set obtained by performing vector conversion on the final set of triples, where the similarity between the target vector and the problem vector is greater than the similarity between other candidate vectors in the candidate vectors and the problem vector;

The triple set includes multiple triples. The triples can be knowledge in the form of <head node-relationship-tail node> or <node-attribute-attribute value> in the knowledge graph. The knowledge triples can be Combine into sentence form. The following takes <head node-relationship-tail node> as an example for illustrative explanation. It should be understood that knowledge in the form of <node-attribute-attribute value> can also achieve the same technical effect, and will not be discussed here. Repeat.

In an example, the head node and the tail node can be entities, that is, <head node-relationship-tail node> can be expressed as <entity-relationship-entity>, for example, <Zhang San-wife-Li Si>. In the question and answer process, the interaction is generally based on sentences, so you can use three vertical lines to separate triples (such as: Zhang San | | | Wife | | | Li Si), you can also use other symbols or use "", " "Yes" and other characters are connected (for example: Zhang San's wife is Li Si), or triples can be directly spliced (for example: Zhang San's wife is Li Si), so that the sentence obtained based on the triples includes the three-tuples. all content.

In single-hop graph Q&A, user questions generally only include head nodes and relationships and ask about tail nodes; or only include relationships and tail nodes and ask about head nodes; or they only include head nodes and tail nodes and ask about relationships. For example, the user question can be "Who is Zhang San's wife?", or "What is the relationship between Zhang San and Li Si?" etc. The existing graph question and answer system believes that "Who is the child of Zhang San and Li Si?" compared with "Who is Zhang San's wife?" and the spliced sentence of the triplet <Zhang San-wife-Li Si> "Zhang San||| Wife|||李思" has a higher similarity. Since the user question "Who is the child of Zhang San and Li Si?" mentions the two entities "Zhang San" and "Li Si", the similarity with "Zhang San|||wife|||Li Si" is even greater. High, it is easier to hit the triple <Zhang San-Wife-Li Si>, resulting in a lower question answering accuracy of the existing graph question and answer system. In fact, the spliced sentence "Zhang San|||wife|||Li Si" of the triplet <Zhang San-wife-Li Si> needs to have a higher degree of similarity with the user question "Who is Zhang San's wife?" Therefore, in the present invention, each triplet in the triplet set is cut to obtain a cut triplet set, so as to reduce the interference of the answer part in the triplet set to the question and answer.

Among them, the similarity between the spliced sentence of triples and the user's question can be realized based on calculating the distance between vectors. The question vector is obtained by vector conversion of the user question, and the candidate vector is obtained by vector conversion of the cut triple set; the similarity between the question vector and the candidate vector is calculated to determine the distance from the question vector in the candidate vector The shortest target vector, that is, the target vector and the question vector in the candidate vectors have a high degree of similarity, so that the user question is hit to the triplet corresponding to the target vector. Further answer user questions through step 102.

Step 102: Answer the user question according to the triplet corresponding to the target vector.

The candidate vector is a vector set obtained by vector conversion based on the cut triple set. In other words, the target vector is determined among the candidate vectors to be a certain vector in the vector set obtained by vector conversion of the cut triple set. According to The target vector can determine its corresponding cut triplet, and output the missing part of its corresponding cut triplet as the answer to the user's question, or it can also use the spliced sentence of the triplet corresponding to the target vector as the user The answer to the question is output. The graph question and answer method provided by the present invention can be applied in the graph question and answer system to improve the accuracy of the graph question and answer system in answering user questions.

In an example, the triplet <Zhang San-wife-Li Si> can be cut into <Zhang San-wife>, <wife-Li Si>, <Zhang San-Li Si>, and the user question can be "Zhang San's Who is the wife?", thus, based on the calculation of the similarity between the question vector and the candidate vector, the cut triplet corresponding to the determined target vector can be <Zhang San-wife>. The missing part in the cut triple <Zhang San-wife> is "Li Si", so "Li Si" can be output as the answer to the user's question "Who is Zhang San's wife?"; or, the target vector corresponding to The spliced sentence "Zhang San's wife is Li Si" of the triplet <Zhang San-wife-Li Si> is output as the answer to the user's question "Who is Zhang San's wife?", which improves the accuracy of question and answer.

In the embodiment of the present invention, the similarity between the question vector obtained by vector conversion based on the user's question and the candidate vector obtained by vector conversion based on the cut triple set is calculated to determine the target vector, and then based on the target vector corresponding The missing part in the cut triplet or the spliced sentence of the triplet corresponding to the target vector is output as the answer to the user's question. Before calculating the similarity between the question vector and the candidate vector, each triplet in the triplet set is cut to obtain the cut triplet set. In this way, the triplet set is reduced when calculating the similarity. The interference of the answer part in the group to the question and answer improves the accuracy of the graph question and answer system in answering user questions.

In some optional implementations, before calculating the similarity between the question vector and the candidate vector in step 101, the method further includes:

Construct a knowledge graph based on the user graph data to obtain a triplet set, where each triplet in the triplet set includes three elements;

Cut each triplet in the triplet set into three tuples to obtain a cut triplet set, and each triplet in the cut triplet set includes two elements;

The cut triple set is vector converted to obtain the candidate vector.

The user map data may be structured map data provided by the user. For example, the user map data contains a spreadsheet (excel) file of each node data (including attribute information of all nodes) and relationship data (including relationship attribute information). The extracted data in excel can be stored in the graph database (neo4j) to facilitate subsequent use.

Faced with small and medium-sized domain maps without a large amount of annotated data, general models are not effective in identifying entities and entity links. They are no longer able to obtain the correct content when finding candidate relationships or candidate subgraphs, which ultimately leads to the overall question and answer effect. Poor. Therefore, in the knowledge graph in the small sample field, the present invention constructs a knowledge graph according to the user graph data to obtain a triplet set. The triplet set can include multiple triples, and the triplet can be a triplet set in the knowledge graph. Knowledge in the form of elements <head node-relationship-tail node> or <node-attribute-attribute value>. By cutting each triple in the triple set into three tuples, the correlation between the candidate tuples and the question is calculated, avoiding the step of entity recognition, and cutting the triple into tuples. Finally, it avoids the situation where the answer entities in the triplet interfere with the similarity calculation and improves the accuracy of question and answer.

In this embodiment, as shown in Figure 2, the set of triples may include n triples, for example, <entity-relationship-entity>1...<entity-relationship-entity>i...<entity-relationship- Entity>n; cut each triplet in the triplet set into three tuples. For example, you can cut the triplet <entity-relationship-entity>1 to get three tuples including Two elements of <entity-relationship>1, <relationship-entity>1 and <entity-entity>1, similarly, cut <entity-relationship-entity>i to get <entity-relationship>i, <relationship- Entity>i and <entity-entity>i, cut <entity-relationship-entity>n to obtain <entity-relationship>n, <relationship-entity>n and <entity-entity>n. That is, the cut triple set can include: <entity-relationship>1, <relationship-entity>1 and <entity-entity>1...<entity-relationship>i, <relationship-entity>i and <entity- Entity>i...<entity-relationship>n, <relationship-entity>n and <entity-entity>n. Among them, i and n are both positive integers, and n is greater than i.

Then, the cut triple set is vector transformed to obtain candidate vectors. For example, perform vector conversion on <entity-relation>1, <relationship-entity>1 and <entity-entity>1 respectively to obtain candidate vectors <vector (vector)1>1, <vector2>1 and <vector3>1; respectively Perform vector conversion on <entity-relationship>i, <relationship-entity>i and <entity-entity>i to obtain candidate vectors <vector1>i, <vector2>i and <vector3>i; respectively, <entity-relationship>n , <relation-entity>n and <entity-entity>n are vector converted to obtain candidate vectors <vector1>n, <vector2>n and <vector3>n. In this way, when calculating the similarity between the question vector and the candidate vector, the interference of the answer part in the triplet on the question and answer is reduced, and the accuracy of the graph question and answer system in answering user questions is improved. In practice, this solution has better results. For general models that have not been trained for the domain map, after using the split triple combination method to eliminate the interference of answer nodes, the results on the knowledge base question set in the knowledge field are better. The answer accuracy rate can even increase from 60% to more than 90%.

Compared with traditional information retrieval methods, the method proposed by the present invention bypasses the step of entity recognition, avoids the fact that the domain map has a lot of professional content, resulting in low accuracy of entity recognition, and there is no annotated data for small and medium sample data to improve entity recognition. Accuracy issue. Compared with the method of directly using all triples as candidate subgraphs to calculate the similarity with user questions, this system uses a split triple combination method. When calculating the similarity, it excludes the answer part from interfering with the similarity calculation and causing the correct answer. The answer is not the one with the highest similarity among the triples.

In some optional implementations, performing vector conversion on the cut triple set to obtain the candidate vector includes:

The cut triple set is vector converted and then clustered to obtain index vectors. Each index vector is the clustering center of the candidate vectors it contains. The number of candidate vectors is greater than the number of the index vectors. quantity;

The calculation of the similarity between the question vector and the candidate vector to determine the target vector in the candidate vector includes:

Calculate the similarity between the question vector and the index vector to determine a target index vector in the index vector, the similarity between the target index vector and the question vector is greater than that of other index vectors in the index vector and the question Similarity of vectors;

Among the candidate vectors included in the target index vector, the similarity between the question vector and the candidate vector is calculated to determine the target vector.

In an example, each triple in the triple set is cut into three triples, and the cut triples are converted into vectors to obtain candidate vectors, and then the correlation between the candidate vectors and the problem vector is calculated. When the problem is solved, the problem vector and all candidate vectors can be traversed and calculated to determine the target vector among the candidate vectors. Although the accuracy of semantic similarity is effectively improved, since each triple is cut into three binary group, the number of spliced sentences corresponding to the candidate vectors has increased three times compared to the number of spliced sentences of the pre-cut triplets, that is, the number of spliced sentences that need to be compared has increased three times, and the calculation time has also increased three times. Moreover, when the amount of knowledge is large, the calculation time itself is also long. Therefore, in other embodiments, the index calculation method can be introduced to reduce the calculation time.

In another embodiment, each triple in the triple set is cut into three triples, and the cut triples are converted into vectors to obtain candidate vectors, and then the relationship between the candidate vector and the problem vector is calculated. When the correlation is high, the method of traversing all candidate vectors can be changed to the method of calculating the index vector.

Specifically, the cut triple set (for example: <entity-relationship>1, <relationship-entity>1 and <entity-entity>1...<entity-relationship>i, <relationship-entity>i and <entity-entity>i...<entity-relationship>n, <relationship-entity>n and <entity-entity>n) perform vector conversion to obtain the cut triplet vector set (for example: <vector1>1, <vector2>1 and <vector3>1...<vector1>i, <vector2>i and <vector3>i...<vector1>n, <vector2>n and <vector3>n).

Then, all candidate vectors <vector> in the cut triple vector set are clustered, and the center point of the cluster is selected as the representative index vector of this type of <vector>, that is, each index vector contains The cluster center of the candidate vector. For example, after clustering all candidate vectors <vector> in the cut triple vector set, the obtained index vectors can include <index vector>1...<index vector>j...<index vector>m,j and m are both positive integers, and j is less than m. Among them, the number of candidate vectors is greater than the number of index vectors, that is, n is greater than m, to improve calculation speed. Among them, <index vector>1 can include Nuogan <vector>, and <index vector>1 is the cluster center of the Noogan <vector> it contains.

Among them, the number of index vectors for each cluster can be set at about 1,000-10,000. Depending on the size of the knowledge base and the practical effect, different numbers of clusters can be set. The number of candidate vectors and the number of index levels in each cluster are not limited here. They should be determined according to the actual situation, so that when calculating the similarity, the most similar vector can be accurately obtained without making mistakes. If the database content is large and the number of indexes is too large, the indexes can be clustered again, that is, setting up a second-level index or a multi-level index to reduce the calculation time.

In this way, when calculating the similarity, first calculate the similarity between the question vector and the index vector to determine the target index vector in the index vector. The similarity between the target index vector and the question vector is greater than the similarity between other index vectors and the question vector in the index vector. similarity; after selecting the most similar target index vector, calculate the similarity between the question vector and the candidate vector among the candidate vectors included in the target index vector to determine the target vector. Then the knowledge combination sentence corresponding to the target vector is obtained, and the answer is obtained. After using index calculation, compared with traditional methods, the accuracy of question and answer is improved and the time consumption is shorter. While improving accuracy, it also improves efficiency.

Among them, it is assumed that the number of triples in the knowledge base is N, assuming that the entities of the triples are not repeated, and the number of entities is 2N. After identifying the entity in the user question, this entity needs to be compared with the 2N entities in the database. Make comparisons. In the present invention, a triplet will be divided and combined to generate 3N candidate sentences. Temporarily regarded as information retrieval, the time consuming to find alternative subgraphs and select the part of the subgraph where the answer is located is equivalent to the time consuming of the present invention to find the corresponding answer based on the selected sentence. The time consuming of the present invention is 3N/2N of the information retrieval method. times, that is, 1.5 times. After constructing an index for the candidate vectors, the question and answer time of the present invention decreases exponentially, and the final time-consuming will be equivalent to, or even shorter than, the time-consuming of information retrieval.

In the experiment, in the process of answering user questions, the first aspect that takes a long time is to calculate the similarity between the question vector and the candidate vector. The time consumption here is related to the size of the knowledge base content, that is, the number of combined sentences. There is a positive correlation. After using the segmented sentence combination method, the time spent here is three times that of the common combination method. Regarding the first aspect, the present invention performs clustering by setting index vectors, thereby shortening the time required for question and answer. The second aspect that takes a long time is the process of vector conversion of the cut triple set to obtain candidate vectors.

Therefore, in some optional implementations, before calculating the similarity between the problem vector and the candidate vector, the cut triple set is vector transformed, and after the candidate vector is obtained, The method also includes:

The candidate vectors are pre-stored locally or in a database, the database is associated with a computing device, and the computing device is used to calculate the similarity between the question vector and the candidate vector.

In this implementation, for the vector conversion part of the cut triple set, after the knowledge graph is constructed and the content is stabilized, the contents of the combined sentence vectors mapped according to the content of the knowledge triples in the graph are fixed. , so the conversion part of the vector does not need to be placed in the question and answer part, but is completed in advance during map construction or training, and the candidate vectors are pre-stored locally or in a database. The database is associated with the computing device, and the computing device is used for calculation The similarity between the question vector and the candidate vector. In this way, when answering questions, you only need to vector-convert user questions to obtain question vectors, saving the time of vector-converting the cut triple set.

In some optional implementations, before calculating the similarity between the question vector and the candidate vector in step 101, the method further includes:

Train a vector conversion model, which is used to perform vector conversion based on user questions to obtain question vectors, and perform vector conversion based on the cut triple set to obtain candidate vectors;

Wherein, the vector conversion model is a pre-trained language model RoFormer-Sim, and the normalization algorithm in the RoFormer-Sim is changed to the root mean square value RMS algorithm.

In this embodiment, a method using similar text semantics can be used to provide graph question and answer services. After constructing the map and obtaining the triplet set, the triplet set is cut and combined, and then the cut triplet set is converted into a candidate vector through the vector conversion model. The vector conversion model in the present invention combines the improved model of Roformer-sim and RoformerV2. The improved model maintains the effect while optimizing the training method and model structure, effectively reducing the training time.

Commonly used semantic similarity models include traditional statistical models (for example, weighted technology for information retrieval and data mining (term frequency–inverse document frequency, TF-IDF), simHash, etc.), deep learning models (for example, deep structured Semantic model (Deep Structured Semantic Model, DSSM), Bidirectional Encoder Representation from Transformers (BERT) and its variant models, etc.). The currently popular similar models are BERT and its variant models. However, when using the basic BERT model to perform similar matching for graph question answering, good results cannot be obtained.

In order to improve the effect of question and answer and similar comparison, the present invention uses an improved model that combines RoFormer-Sim and RoFormerV2, striving to improve training efficiency as much as possible while maintaining accuracy.

Among them, the RoFormer-Sim model is an upgraded version of the effective BERT optimization model SimBERT, which can be popularly called "SimBERTv2" or RoFormer-Sim. From the outside, except that the basic architecture is replaced by RoFormer, there is no obvious difference between RoFormer-Sim and SimBERT. In fact, their main difference lies in the details of training. Two formulas can be used for comparison:

Formula 1: SimBERT=BERT+UniLM+Contrastive Learning

Formula 2: RoFormer-Sim=RoFormer+UniLM+Contrastive Learning+BART+Distillation

In addition, RoFormer-Sim uses more training data and extends to general sentence patterns. In other words, unlike SimBERT which is only limited to interrogative sentences, RoFormer-Sim can be used to generate similar sentences for general sentences. It is suitable for The scene is bigger. Compared with the original BERT model and the improved SimBERT model, the RoFormer-Sim model is more suitable for the current task.

The RoFormerV2 model is an optimization of the RoFormer model structure. Structurally, RoFormerV2 mainly removes all Bias items of the model, replaces the Layer Norm with a simple RMS Norm, and removes the gamma parameter of the RMS Norm. After optimizing these structures, the training speed of the model is increased to 1.3 times that of the baseline model. Change the normalization algorithm in RoFormer-Sim to the root mean square value RMS algorithm to further optimize the training speed.

RoFormer-Sim follows the Transformer classic normalization algorithm (Layer Normalization, LN). For the hidden layer of layer l, let a ^l represent the input vector (the weighted vector output by the front layer network). The statistics of its mean and variance can be yes:

Among them, μ represents the mean, σ represents the variance, and H represents the number of hidden layer units.

Then the normalized value of a ^l of the LN algorithm

where ∈ represents a very small number to prevent the denominator from being 0.

In order to ensure that the normalization operation will not destroy the previous information, a new set of parameters, gain g and bias b, are added. The final output of LN is:

Since LN calculation is relatively complex, in order to speed up the training, a simpler RMS can be used for normalization operation, and the gamma parameter of RMS Norm is removed. The output of the i-th layer of the RMS algorithm used in the present invention is:

Modify the model RoFormer-Sim and change the normalization algorithm of the model to RMS. When calculating the similarity, use the modified RoFormer-Sim model to convert the user question and the spliced sentence of triples, calculate the similarity between the question vector and the candidate vector, and determine the target vector among the candidate vectors. After experiments, when using this model for mapping compared to using Bert for mapping, the performance is better for some domain knowledge with several different words. After changing the model, the accuracy of identifying professional words with small differences in words has been greatly improved.

Referring to Figure 3, Figure 3 is a structural diagram of a graph question and answer device provided by an embodiment of the present invention. As shown in Figure 3, the graph question and answer device 300 includes:

The calculation module 301 is used to calculate the similarity between a question vector and a candidate vector to determine a target vector in the candidate vector, wherein the question vector is a vector obtained by vector conversion based on the user question, and the candidate vector It is a vector set obtained by vector conversion based on the cut triple set, and the similarity between the target vector and the problem vector is greater than the similarity between other candidate vectors in the candidate vectors and the problem vector;

The answering module 302 is used to answer the user question according to the triplet corresponding to the target vector.

Optionally, the graph question and answer device 300 also includes:

A building module, used to construct a knowledge graph based on user graph data to obtain a triplet set, where each triplet in the triplet set includes three elements;

A cutting module, used to cut each triplet in the triplet set into three tuples to obtain a cut triplet set. Each triplet in the cut triplet set is Includes two elements;

A conversion module, configured to perform vector conversion on the cut triple set to obtain the candidate vector.

Optionally, the conversion module includes:

The clustering submodule is used to perform vector conversion on the cut triple set and perform clustering to obtain index vectors. Each index vector is the clustering center of the candidate vectors it contains. The number is greater than the number of index vectors;

The calculation module 301 includes:

The first calculation sub-module is used to calculate the similarity between the question vector and the index vector to determine the target index vector in the index vector, and the similarity between the target index vector and the question vector is greater than the index vector The similarity between other index vectors in and the problem vector;

The second calculation submodule is used to calculate the similarity between the question vector and the candidate vector among the candidate vectors included in the target index vector, and determine the target vector.

Optionally, the graph question and answer device 300 also includes:

A storage module, configured to pre-store the candidate vector locally or in a database, where the database is associated with a computing device, and the computing device is configured to calculate the similarity between the question vector and the candidate vector.

Optionally, the graph question and answer device 300 also includes:

A training module, used to train a vector conversion model, which is used to perform vector conversion based on user questions to obtain question vectors, and perform vector conversion based on the cut triple set to obtain candidate vectors;

Wherein, the vector conversion model is a pre-trained language model RoFormer-Sim, and the normalization algorithm in the RoFormer-Sim is changed to the root mean square value RMS algorithm.

It should be noted that the graph question and answer device 300 can implement each process in the method embodiment shown in Figure 1. To avoid duplication, the details will not be described here.

Embodiments of the present invention also provide a graph question and answer system, including:

at least one processor; and

a memory communicatively connected to the at least one processor; wherein,

The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor, so that the at least one processor can perform various processes such as the above graph question and answer method embodiment, And can achieve the same technical effect, so to avoid repetition, they will not be described again here.

Embodiments of the present invention also provide a non-transient computer-readable storage medium storing computer instructions. The computer instructions are used to cause the computer to execute each process of the above-mentioned graph question and answer method embodiment, and can achieve the same To avoid repetition, the technical effects will not be repeated here. Wherein, the computer-readable storage medium is such as ROM, RAM, magnetic disk or optical disk.

It should be noted that, in this document, the terms "comprising", "comprises" or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or device that includes a series of elements not only includes those elements, It also includes other elements not expressly listed or inherent in the process, method, article or apparatus. Without further limitation, an element defined by the statement "comprises a..." does not exclude the presence of additional identical elements in a process, method, article or apparatus that includes that element. In addition, it should be pointed out that the scope of the methods and apparatuses in the embodiments of the present invention is not limited to performing the functions in the order discussed, but may also include performing the functions in a substantially simultaneous manner or in the reverse order according to the involved functions. For example, the described methods may be performed in an order different from that described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus the necessary general hardware platform. Of course, it can also be implemented by hardware, but in many cases the former is better. implementation. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product in essence or the part that contributes to the existing technology. The computer software product is stored in a storage medium (such as ROM/RAM, disk, CD), including several instructions to cause a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in various embodiments of the present invention.

The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above-mentioned specific implementations. The above-mentioned specific implementations are only illustrative and not restrictive. Those of ordinary skill in the art will Under the inspiration of the present invention, many forms can be made without departing from the spirit of the present invention and the scope protected by the claims, all of which fall within the protection of the present invention.

Claims (10)

1. A graph question and answer method, characterized in that the method includes: Calculate the similarity between the question vector and the candidate vector to determine the target vector in the candidate vector, wherein the question vector is a vector obtained by vector conversion based on the user question, and the candidate vector is based on the cut three-dimensional vector. A set of vectors obtained by vector conversion of a tuple set, where the similarity between the target vector and the problem vector is greater than the similarity between other candidate vectors in the candidate vectors and the problem vector; Answer the user question according to the triplet corresponding to the target vector. 2. The method according to claim 1, characterized in that, before calculating the similarity between the problem vector and the candidate vector, the method further includes: Construct a knowledge graph based on the user graph data to obtain a triplet set, where each triplet in the triplet set includes three elements; Cut each triplet in the triplet set into three tuples to obtain a cut triplet set, and each triplet in the cut triplet set includes two elements; The cut triple set is vector converted to obtain the candidate vector. 3. The method according to claim 2, characterized in that said converting the cut triple set into vectors to obtain the candidate vectors includes: The cut triple set is vector converted and then clustered to obtain index vectors. Each index vector is the clustering center of the candidate vectors it contains. The number of candidate vectors is greater than the number of the index vectors. quantity; The calculation of the similarity between the question vector and the candidate vector to determine the target vector in the candidate vector includes: Calculate the similarity between the question vector and the index vector to determine a target index vector in the index vector, the similarity between the target index vector and the question vector is greater than that of other index vectors in the index vector and the question Similarity of vectors; Among the candidate vectors included in the target index vector, the similarity between the question vector and the candidate vector is calculated to determine the target vector. 4. The method according to claim 2, characterized in that, before calculating the similarity between the problem vector and the candidate vector, the cut triple set is vector transformed to obtain the After the candidate vectors, the method also includes: The candidate vectors are pre-stored locally or in a database, the database is associated with a computing device, and the computing device is used to calculate the similarity between the question vector and the candidate vector. 5. The method according to claim 1, characterized in that, before calculating the similarity between the question vector and the candidate vector, the method further includes: Train a vector conversion model, which is used to perform vector conversion based on user questions to obtain question vectors, and perform vector conversion based on the cut triple set to obtain candidate vectors; Wherein, the vector conversion model is a pre-trained language model RoFormer-Sim, and the normalization algorithm in the RoFormer-Sim is changed to the root mean square value RMS algorithm. 6. A graph question and answer device, characterized in that it includes: The calculation module is used to calculate the similarity between the question vector and the candidate vector to determine the target vector in the candidate vector, wherein the question vector is a vector obtained by vector conversion based on the user question, and the candidate vector is A vector set obtained by performing vector conversion based on the cut triple set, where the similarity between the target vector and the problem vector is greater than the similarity between other candidate vectors in the candidate vectors and the problem vector; An answering module is used to answer the user question according to the triplet corresponding to the target vector. 7. The device of claim 6, further comprising: A building module, used to construct a knowledge graph based on user graph data to obtain a triplet set, where each triplet in the triplet set includes three elements; A cutting module, used to cut each triplet in the triplet set into three tuples to obtain a cut triplet set. Each triplet in the cut triplet set is Includes two elements; A conversion module, configured to perform vector conversion on the cut triple set to obtain the candidate vector. 8. The device according to claim 7, wherein the conversion module includes: The clustering submodule is used to perform vector conversion on the cut triple set and perform clustering to obtain index vectors. Each index vector is the clustering center of the candidate vectors it contains. The number is greater than the number of index vectors; The computing module includes: The first calculation sub-module is used to calculate the similarity between the question vector and the index vector to determine the target index vector in the index vector, and the similarity between the target index vector and the question vector is greater than the index vector The similarity between other index vectors in and the problem vector; The second calculation submodule is used to calculate the similarity between the question vector and the candidate vector among the candidate vectors included in the target index vector, and determine the target vector. 9. A graph question and answer system, characterized by including: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor, so that the at least one processor can perform as claimed in any one of claims 1 to 5 method described. 10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that the computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 5.