An intelligent question-answering system that is adaptive to knowledge graphs in different fields and its construction method

Abstract

The present invention discloses a method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields. The method analyzes user questions by constructing a user question analysis module; constructs an entity linker and a pattern linker to optimize the user question analysis results, and generates a candidate question combination feature list by arranging and combining graph data and graph patterns; constructs a three-level knowledge retriever to obtain candidate answers corresponding to candidate questions from a graph database through the three-level knowledge retriever; constructs a candidate answer processing module, and the candidate answer processing module statistically converts, sorts and naturalizes the candidate answers through a statistical intent classification model, an answer sorting model and an answer natural language model, and generates a final answer to send to the user. The invention generates a question combination feature list by analyzing user questions, and the three-level knowledge retriever retrieves the corresponding candidate answers, thereby improving the ability of the intelligent question-answering system to adapt to different fields.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and relates to information retrieval, knowledge graphs, natural language processing and question-answering technology. Specifically, it is a method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields and an intelligent question-answering system. Background Art

[0002] Information retrieval is an indispensable way to obtain information and news, from ancient book information retrieval to search engines, and then to the current natural language question-answering system. The early information retrieval method was based on text or string matching, including the use of wildcards such as "?" and "*" in regular expressions to match strings, and "like" or "ilike" expressions in the database query language SQL. With the development of the Internet, search engines represented by Google and Baidu have provided new information retrieval methods to obtain information in the form of keywords.

[0003] With the development of information extraction, knowledge graph and related technologies, the intelligent question answering technology based on knowledge graph (KBQA) can provide users with more accurate and concise answers to questions. Current research on KBQA is attached to specific knowledge graphs, such as DBPedia. In real applications, an organization or enterprise often has multiple knowledge graphs suitable for their respective businesses, such as business knowledge base graphs that meet customer service scenarios, human resource graphs that meet human resource-related application scenarios, and customer relationship graphs for departments such as marketing, sales or customer service. This question-and-answer system that relies on a specific graph has various shortcomings for the needs of multiple graphs, such as high construction cost, inability to switch freely, high maintenance cost, and unfriendly to new businesses and new scenarios.

[0004] Patent No. CN108182262A discloses a method and system for constructing an intelligent question-answering system based on deep learning and knowledge graphs. A crawler is used to obtain a medical consultation data set from the Internet, and data preprocessing is performed to obtain a labeled data set; a word segmentation dictionary based on the medical field is constructed in combination with the hospital electronic medical records, and is merged with the medical dictionary as the system's word segmentation dictionary; a knowledge graph associated with diseases and symptoms is constructed, and disease entity alignment and symptom entity alignment are performed; a labeled data set is obtained based on the disease entity alignment; a language model based on deep learning is constructed; a query optimization algorithm based on a knowledge graph combined with user context information is constructed; a training data set that fuses a language model and a knowledge graph is constructed and model fusion training is performed to obtain a pre-diagnosis fusion model based on a language model and a knowledge graph. This invention is mainly applicable to question-answering methods based on knowledge graphs in the field of medical diagnosis, and is not applicable to question-answering knowledge graphs in different fields.

[0005] From the perspective of user questions, the questions users input often cover a very wide range, and there may be various noises at the same time, different ways of expressing the same question, frequent spelling errors, or even a randomly input text. At the same time, the knowledge for these questions may come from multiple different graphs, and even the knowledge of multiple graphs may be required to answer the questions. At present, a question-answering system that meets this real-world scenario cannot be achieved by a single method, and there is little related research, which makes it difficult for intelligent question-answering technology based on knowledge graphs to achieve large-scale and rapid commercial application value. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for constructing an intelligent question and answer system that is adaptive to knowledge graphs in different fields. The method of the present invention analyzes user questions to generate a question combination feature list, retrieves candidate answers from the graph database through a three-level knowledge retriever, and performs final processing on the candidate answers through a candidate answer processing module and sends the final answer to the user, thereby improving the ability of knowledge graphs in different fields to effectively and accurately answer questions without or with very little secondary development workload, thereby improving the field adaptability of the intelligent knowledge question and answer system.

[0007] In order to achieve the purpose of the above invention, the technical solution provided by the patent of this invention is as follows:

[0008] A method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields, the method specifically comprising the following steps:

[0009] S1, analyze user questions by using named entity recognition technology, stuttering word segmentation technology, full-text retrieval technology, semantic matching technology and synonym strategy;

[0010] S2, obtaining graph data and graph patterns in the knowledge graph, wherein the graph data includes entity values ​​and attribute values, and generating a candidate question combination feature list by arranging and combining the graph data and graph patterns based on the user question analysis results obtained in S1;

[0011] S3, the three-level knowledge retriever includes the first retriever, the second retriever and the third retriever. In the knowledge retrieval stage, based on the question combination feature list generated in S2, the candidate answers corresponding to the candidate questions are obtained from the graph database through the three-level knowledge retriever;

[0012] S4, the candidate answer processing module performs statistical conversion, sorting and natural language conversion on the candidate answers generated in S3 through the statistical intent classification model, answer sorting model and answer natural language conversion model, and generates the final answer to send to the user.

[0013] The construction of the user question analysis module in S1 above specifically includes the following steps:

[0014] S11, by using named entity recognition technology to build a named entity recognition model to perform entity recognition on the graph data of the user's question, using the sentences where the entities, relationships and attribute-related words in the annotated knowledge graph and their annotated values ​​and constructing the knowledge graph<Subject, Predict,Object> The training model corpus is constructed by triples, and the named entity recognition model effect is optimized through the training corpus. The triples include <entity, entity attribute name, entity attribute value> and <entity, relationship, entity>;

[0015] S12, using the Jieba word segmentation technology to import the graph pattern into the word segmentation dictionary, and using the word segmentation dictionary to extract words related to the graph pattern from the user's question, the graph pattern includes entity type names, entity attribute names, relationship type names and relationship attribute names;

[0016] S13, using full-text search technology to import the graph data into the dictionary of the full-text search engine, using the Elasticsearch tool to extract the vocabulary and attribute values ​​in the form of enumerated strings related to the graph data from the user's question, the full-text search technology will also be used to recall the candidate entity vocabulary related to named entity recognition, and semantic matching technology will be used to perform entity linking to obtain accurate entity vocabulary related to the graph data;

[0017] S14, construct the configured entity synonym dictionary and graph pattern synonym dictionary, expand the detected entity vocabulary and graph pattern vocabulary through the entity synonym dictionary and graph pattern synonym dictionary, perform synonym replacement through the synonym strategy, and find the entity vocabulary and graph pattern vocabulary that actually exist in the graph data.

[0018] The question combination feature list in S2 above arranges and combines the retrieved graph data vocabulary and graph pattern vocabulary to obtain all question combinations.

[0019] The first-level knowledge retriever in the above S3 is a rule engine constructed by creating a retrieval question template. It serves as the first-level knowledge retriever in the knowledge retrieval stage. In the knowledge retrieval stage, it quickly configures the question templates and target retrieval statements for intention judgment, and supports any combination of "and, or and conditions" between different rules to realize the reuse of the created question templates.

[0020] The second-level knowledge retriever in S3 above is a second-level knowledge retriever established based on intent analysis. In the knowledge retrieval stage, when the first-level knowledge retriever cannot retrieve a valid candidate answer, the second-level knowledge retriever is run. The second-level knowledge retriever builds an intent classification model, which is trained based on the collected open source corpus and the newly annotated corpus, and performs a secondary search based on the graph data and graph patterns obtained from the knowledge graph to obtain candidate answers.

[0021] The third-level knowledge retriever in the above S3 is the last-level backup retriever in the knowledge retrieval stage. In the knowledge retrieval stage, when the first-level knowledge retriever and the second-level knowledge retriever cannot retrieve valid candidate answers, the third-level knowledge retriever is run. The third-level knowledge retriever uses the question combination feature list detected in the question analysis stage to directly perform semantic matching retrieval with the knowledge graph triple data, and finally obtains the candidate answers to the user's questions.

[0022] The statistical intent classification model in S4 above classifies the generated candidate answers into six categories: “total,” “yes or no,” “enumeration,” “sequence,” “comparison,” and “others.” This classification guides subsequent answer conversion behavior.

[0023] The answer ranking model in S4 above includes two answer types: "sequence type" and "other type". The answer ranking model uses the Bert-based semantic similarity model and the LambdaRank ranking model to re-rank the results and sort the candidate answers according to the probability.

[0024] The natural language model in S4 above converts the final answer into a natural response sentence that can be understood by the user by utilizing a general answer template and a customized answer template.

[0025] An intelligent question-answering system that is adaptive to knowledge graphs in different fields is implemented based on any of the above-mentioned methods for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields, and includes the following:

[0026] In the first step, the user enters his or her own question statement;

[0027] The second step is the user question analysis module, which is used to pre-process and analyze the question sentences input by the user;

[0028] The third step is to use the graph data linker and graph pattern linker to obtain the graph data and graph patterns in the knowledge graph, and generate a candidate question combination feature list by arranging and combining the graph data and graph patterns.

[0029] In the fourth step, the three-level knowledge retriever performs knowledge retrieval based on the candidate question combination feature list generated in the second step to obtain candidate answers corresponding to the candidate questions;

[0030] The fifth step is the candidate answer processing module, which is used to process the candidate answers. The generated candidate answers are classified through the statistical intent classification model. The natural language model converts the final answer into a natural response sentence that the user can understand by using the general answer template and the customized answer template. The answer ranking model uses the Bert-based semantic similarity model and the LambdaRank ranking model to re-rank the generated candidate answers, and sorts the candidate answers according to the probability to generate the final answer.

[0031] Step 6: Feedback the final answer generated in step 4 to the user.

[0032] The method of the present invention has achieved the following positive and beneficial effects through practice:

[0033] 1. The present invention achieves better named entity recognition effect and improves the accuracy of named entity recognition by constructing training samples using data in the knowledge graph.

[0034] 2. The present invention solves the word segmentation performance problem under large-scale graph data by utilizing Elasticsearch's word segmentation technology, thereby improving word segmentation efficiency.

[0035] 3. The present invention arranges and combines entity values, attribute values, and graph pattern data to form a question combination feature list, thereby providing preliminary knowledge for subsequent knowledge retrievers to retrieve answers and improving retrieval accuracy.

[0036] 4. The present invention satisfies the basic question-answering effects under knowledge graphs in different fields by constructing a three-level knowledge retriever, and supports convenient and flexible rule configuration optimization effects. It is suitable for retrieval of different knowledge graphs, thereby improving the applicability of the knowledge retriever. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of constructing an intelligent question and answer system in a method for constructing an intelligent question and answer system that is adaptive to knowledge graphs in different fields according to the present invention.

[0038] Figure 2 It is a flowchart of entity information extraction in a question analysis module in a method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is described below by the specific examples shown in the accompanying drawings. However, it should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0040] like Figure 1 As shown, a method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields, the method specifically includes the following steps:

[0041] S1, analyze user questions by using named entity recognition technology, stuttering word segmentation technology, full-text retrieval technology, semantic matching technology and synonym strategy;

[0042] S2, obtaining graph data and graph patterns in the knowledge graph, wherein the graph data includes entity values ​​and attribute values, and generating a candidate question combination feature list by arranging and combining the graph data and graph patterns based on the user question analysis results obtained in S1;

[0043] S3, the three-level knowledge retriever includes the first retriever, the second retriever and the third retriever. In the knowledge retrieval stage, based on the question combination feature list generated in S2, the candidate answers corresponding to the candidate questions are obtained from the graph database through the three-level knowledge retriever;

[0044] S4, the candidate answer processing module performs statistical conversion, sorting and natural language conversion on the candidate answers generated in S3 through the statistical intent classification model, answer sorting model and answer natural language conversion model, and generates the final answer to send to the user.

[0045] like Figure 2 As shown in , the construction of the user question analysis module in S1 above specifically includes the following steps:

[0046] S11, by using named entity recognition technology to build a named entity recognition model to perform entity recognition on the graph data of the user's question, using the sentences where the entities, relationships and attribute-related words in the annotated knowledge graph and their annotated values ​​and constructing the knowledge graph<Subject, Predict,Object> The training model corpus is constructed by triples, and the named entity recognition model effect is optimized through the training corpus. Among them, the triples include <entity, entity attribute name, entity attribute value> and <entity, relationship, entity>. The training corpus is constructed based on the knowledge graph, which improves the generalization ability of the named entity recognition model and the accuracy of entity recognition.

[0047] S12, using the Jieba word segmentation technology to import the graph pattern into the word segmentation dictionary, and using the word segmentation dictionary to extract words related to the graph pattern from the user's question, the graph pattern includes entity type names, entity attribute names, relationship type names and relationship attribute names;

[0048] S13, using full-text search technology to import the graph data into the dictionary of the full-text search engine, using the Elasticsearch tool to extract the vocabulary and attribute values ​​in the form of enumerated strings related to the graph data from the user's question, the full-text search technology is based on the candidate entity vocabulary related to the recall of named entity recognition, and the semantic matching technology is used to link entities to obtain accurate entity vocabulary related to the graph data; the semantic matching technology is an existing technology;

[0049] S14, construct the configured entity synonym dictionary and graph pattern synonym dictionary, expand the detected entity vocabulary and graph pattern vocabulary through the entity synonym dictionary and graph pattern synonym dictionary, perform synonym replacement through the synonym strategy, and find the entity vocabulary and graph pattern vocabulary that actually exist in the graph data.

[0050] The question combination feature list in S2 above arranges and combines the retrieved graph data vocabulary and graph pattern vocabulary to obtain all question combinations.

[0051] The first-level knowledge retriever in the above S3 is a rule engine constructed by creating a retrieval question template. It serves as the first-level knowledge retriever in the knowledge retrieval stage. In the knowledge retrieval stage, it quickly configures the question templates and target retrieval statements for intention judgment, and supports any combination of "and, or and conditions" between different rules, thereby realizing the reuse of the created question templates and improving the development and use efficiency of question templates.

[0052] The second-level knowledge retriever in S3 above is a second-level knowledge retriever established based on intent analysis. In the knowledge retrieval stage, when the first-level knowledge retriever cannot retrieve a valid candidate answer, the second-level knowledge retriever is run. The second-level knowledge retriever builds an intent classification model, which is trained based on the collected open source corpus and the newly annotated corpus, and performs a secondary search based on the graph data and graph patterns obtained from the knowledge graph to obtain candidate answers; it is used to make up for the problem of insufficient generalization ability of the first-level knowledge retriever engine for intent judgment.

[0053] The third-level knowledge retriever in the above S3 is the last-level backup retriever in the knowledge retrieval stage. In the knowledge retrieval stage, when the first-level knowledge retriever and the second-level knowledge retriever cannot retrieve valid candidate answers, the third-level knowledge retriever is run. The third-level knowledge retriever uses the question combination feature list detected in the question analysis stage to directly perform semantic matching retrieval with the knowledge graph triple data, and finally obtains the candidate answers to the user's questions.

[0054] The statistical intent classification model in S4 above classifies the generated candidate answers into six categories: “total,” “yes or no,” “enumeration,” “sequence,” “comparison,” and “others.” This classification guides subsequent answer conversion behavior.

[0055] The answer ranking model in S4 above includes two answer types: "sequence type" and "other type". The answer ranking model uses the Bert-based semantic similarity model and the LambdaRank ranking model to re-rank the results and sort the candidate answers according to the probability.

[0056] The natural language model in S4 above converts the final answer into a natural response sentence that can be understood by the user by utilizing a general answer template and a customized answer template.

[0057] An intelligent question-answering system that is adaptive to knowledge graphs in different fields is implemented based on any of the above-mentioned methods for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields, and includes the following:

[0058] In the first step, the user enters his or her own question statement;

[0059] The second step is the user question analysis module, which is used to pre-process and analyze the question sentences input by the user;

[0060] The third step is to use the graph data linker and graph pattern linker to obtain the graph data and graph patterns in the knowledge graph, and generate a candidate question combination feature list by arranging and combining the graph data and graph patterns.

[0061] In the fourth step, the three-level knowledge retriever performs knowledge retrieval based on the candidate question combination feature list generated in the second step to obtain candidate answers corresponding to the candidate questions;

[0062] The fifth step is the candidate answer processing module, which is used to process the candidate answers. The generated candidate answers are classified through the statistical intent classification model. The natural language model converts the final answer into a natural response sentence that the user can understand by using the general answer template and the customized answer template. The answer ranking model uses the Bert-based semantic similarity model and the LambdaRank ranking model to re-rank the generated candidate answers, and sorts the candidate answers according to the probability to generate the final answer.

[0063] Step 6: Feedback the final answer generated in step 4 to the user.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, ordinary technicians in the field should understand that the specific implementation methods of the invention can still be modified or some technical features can be replaced by equivalents without departing from the spirit of the technical solution of the present invention, which should be included in the scope of the technical solution for protection of the present invention.

Claims

1. A method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields, characterized in that: The method specifically comprises the following steps: S1, analyze user questions by using named entity recognition technology, stuttering word segmentation technology, full-text retrieval technology, semantic matching technology and synonym strategy; The construction of the user question analysis module in S1 specifically includes the following steps: S11, by using named entity recognition technology to build a named entity recognition model to perform entity recognition on the graph data of the user's question, using the sentences where the entities, relationships and attribute-related words in the annotated knowledge graph and their annotated values ​​and constructing the knowledge graph<Subject, Predict, Object> The training model corpus is constructed by triples, and the named entity recognition model effect is optimized through the training corpus. The triples include <entity, entity attribute name, entity attribute value> and <entity, relationship, entity>; S12, using the Jieba word segmentation technology to import the graph pattern into the word segmentation dictionary, and using the word segmentation dictionary to extract words related to the graph pattern from the user's question, the graph pattern includes entity type names, entity attribute names, relationship type names and relationship attribute names; S13, using full-text search technology to import the graph data into the dictionary of the full-text search engine, using the Elasticsearch tool to extract the vocabulary and attribute values ​​in the form of enumerated strings related to the graph data from the user's question, the full-text search technology will also be used to recall the candidate entity vocabulary related to named entity recognition, and semantic matching technology will be used to perform entity linking to obtain accurate entity vocabulary related to the graph data; S14, constructing a configured entity synonym dictionary and graph pattern synonym dictionary, expanding the detected entity vocabulary and graph pattern vocabulary through the entity synonym dictionary and graph pattern synonym dictionary, performing synonym replacement through the synonym strategy, and finding the entity vocabulary that really exists in the graph data and the vocabulary that really exists in the graph pattern; S2, obtaining graph data and graph patterns in the knowledge graph, wherein the graph data includes entity values ​​and attribute values, and generating a candidate question combination feature list by arranging and combining the graph data and graph patterns based on the user question analysis results obtained in S1; S3, the three-level knowledge retriever includes the first retriever, the second retriever and the third retriever. In the knowledge retrieval stage, based on the question combination feature list generated in S2, the candidate answers corresponding to the candidate questions are obtained from the graph database through the three-level knowledge retriever; The first-level knowledge retriever in S3 is a rule engine constructed by creating a search question template. In the knowledge search stage, it acts as the first-level knowledge retriever. In the knowledge search stage, it quickly configures the question template and target search statement for intention determination, and supports any combination of "and, or and conditions" between different rules to achieve the reuse of the created question template. The second-level knowledge retriever in S3 is a second-level knowledge retriever established based on intent analysis. In the knowledge retrieval stage, when the first-level knowledge retriever cannot retrieve a valid candidate answer, the second-level knowledge retriever is run. The second-level knowledge retriever builds an intent classification model, which is trained based on the collected open source corpus and the newly annotated corpus, and performs a secondary search based on the graph data and graph patterns obtained from the knowledge graph to obtain the candidate answer; The third-level knowledge retriever in S3 is the last-level backup retriever in the knowledge retrieval stage. In the knowledge retrieval stage, when the first-level knowledge retriever and the second-level knowledge retriever cannot retrieve valid candidate answers, the third-level knowledge retriever is run. The third-level knowledge retriever uses the question combination feature list detected in the question analysis stage to directly perform semantic matching retrieval with the knowledge graph triple data, and finally obtains the candidate answer to the user's question; S4, the candidate answer processing module performs statistical conversion, sorting and natural language conversion on the candidate answers generated in S3 through the statistical intent classification model, answer sorting model and answer natural language conversion model, and generates the final answer to send to the user.

2. According to claim 1, a method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields is characterized in that: The question combination feature list in S2 arranges and combines the retrieved graph data vocabulary and graph pattern vocabulary to obtain all question combination situations.

3. According to claim 1, a method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields is characterized in that: The statistical intent classification model in S4 classifies the generated candidate answers into six categories: "total category", "yes or no category", "enumeration category", "sequence category", "comparison category" and "other category". This classification guides the subsequent answer conversion behavior.

4. According to claim 1, a method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields is characterized in that: The answer ranking model in S4 includes two answer types: "sequence type" and "other type". The answer ranking model uses the Bert-based semantic similarity model and LambdaRank ranking model to re-rank the results and sort the candidate answers according to the probability size.

5. According to claim 1, a method for constructing an intelligent question-answering system that is adaptive to knowledge graphs in different fields is characterized in that: The natural language model in S4 converts the final answer into a natural response sentence that can be understood by the user by utilizing a general answer template and a customized answer template.

6. An intelligent question-answering system that is adaptive to knowledge graphs in different fields, characterized in that: The method for constructing an intelligent question-answering system based on any one of claims 1 to 5 that is adaptive to knowledge graphs in different fields is implemented as follows: In the first step, the user enters his or her own question statement; The second step is the user question analysis module, which is used to pre-process and analyze the question sentences input by the user; The third step is to use the graph data linker and graph pattern linker to obtain the graph data and graph patterns in the knowledge graph, and generate a candidate question combination feature list by arranging and combining the graph data and graph patterns. The third step is to use a three-level knowledge retriever to perform knowledge retrieval based on the candidate question combination feature list generated in the second step to obtain candidate answers corresponding to the candidate questions. The fourth step is the candidate answer processing module, which is used to process the candidate answers. The generated candidate answers are classified through the statistical intent classification model. The natural language model converts the final answer into a natural response sentence that the user can understand by using the general answer template and the customized answer template. The answer ranking model uses the Bert-based semantic similarity model and the LambdaRank ranking model to re-rank the generated candidate answers, and sorts the candidate answers according to the probability to generate the final answer. The fifth step is the answer sending module, which is used to feed back the sorted final answers to the user.

Images