Intelligent dialogue method and system, computer equipment and storage medium

Abstract

An intelligent dialogue method comprises the following steps: establishing a dialogue according to a dialogue request of a user, and generating or calling a user portrait; inputting the user portrait into a pre-trained user personalized retrieval model, and determining a retrieval strategy; receiving a user question, and performing content optimization on the user question by using a pre-trained semantic recognition model to obtain a complete question; retrieving the complete question in a pre-constructed knowledge graph model based on a retrieval strategy to obtain a reference text; performing consistency and correlation evaluation on the reference text and the complete question, generating a final answer if the evaluation is passed, otherwise, performing retrieval again; and feeding back the final answer to the user. The invention provides an intelligent dialogue method and system, computer equipment and a storage medium, which can more accurately answer a question of a user, consider the personalized demand of the user and optimize the user experience.

Description

Technical Field

[0001] The present invention relates to the field of intelligent dialogue technology, and in particular to an intelligent dialogue method, system, computer equipment and storage medium. Background Art

[0002] With the rapid development of artificial intelligence technology, intelligent customer service systems have made significant progress in recent years. These technological innovations have promoted the gradual maturity of intelligent customer service systems based on conversational artificial intelligence, solving many bottlenecks in traditional customer service processes, such as high labor costs and insufficient service standardization. The optimization of intelligent customer service systems in traditional application scenarios such as service and marketing mainly focuses on the following four aspects: human-machine collaboration, automated question-and-answer, manual assistance services, and monitoring and management of data and service operations.

[0003] For traditional customer service robots based on keywords or simple semantic matching, such systems have significant deficiencies in processing complex natural languages ​​due to their reliance on predefined keywords and rules. They can only match based on keywords, but cannot deeply understand the semantics and context of user input. This makes them prone to incorrect answers when faced with long sentences, polysemous words or language ambiguity, especially when user input does not completely match the predefined keywords, and may return irrelevant or wrong answers. In addition, such systems have poor scalability and are difficult to adapt to increasingly complex and diverse user needs.

[0004] For customer service robots based on large models (such as deep learning models and language models), although these models have made great progress in natural language understanding and generation, they still have significant limitations. First, the "hallucination" problem of large models, that is, the answers generated by the model may not be based on facts or accurate information, but are generated by relying on the patterns learned during its pre-training process. This phenomenon is particularly serious when dealing with enterprise-level applications, because the pre-training data of large models is mostly public data in general fields, lacking enterprise-private knowledge and industry-specific information, making it difficult to provide personalized and accurate business answers.

[0005] In addition, although the knowledge blind spots of large models can be compensated by plug-in knowledge bases, this also brings new challenges. Plug-in knowledge bases usually compare user input with knowledge base content through semantic matching technology, but the accuracy of semantic matching is often limited by the performance of the matching algorithm. When the knowledge matching module cannot accurately capture the semantics of user input or its subtle changes, the system may still provide inaccurate or irrelevant answers. Especially when it comes to multi-level semantic understanding or complex business scenarios, it is difficult for existing matching technologies to guarantee high accuracy.

[0006] At the same time, the training and fine-tuning costs of large models are high, especially for fine-tuning of enterprise private data, which faces limitations in data security, privacy protection, and computing resources. This also makes it difficult for many companies to easily customize and optimize their systems to adapt to their own business needs when using general large models, thus affecting the actual application effect of intelligent customer service robots. Summary of the invention

[0007] In order to solve the shortcomings of poor accuracy and poor user experience in the prior art intelligent dialogue technology, the present invention provides an intelligent dialogue method, system, computer device and storage medium, which can answer user questions more accurately, take into account the user's personalized needs, and optimize the user experience.

[0008] In order to achieve the above object, the specific scheme adopted by the present invention is: an intelligent dialogue method, comprising the following steps: Establish a dialogue based on the user's dialogue request and generate or call a user profile; Input the user profile into the pre-trained user personalized retrieval model to determine the retrieval strategy; Receive user questions and use the pre-trained semantic recognition model to optimize the content of user questions to obtain complete questions; Based on the retrieval strategy, the complete question is searched in the pre-built knowledge graph model to obtain the reference text; Evaluate the consistency and relevance of the reference text and the complete question. If the evaluation is passed, the final answer will be generated. Otherwise, the search will be repeated. Feedback the final answer to the user.

[0009] As a further optimization of the above-mentioned intelligent dialogue method: after establishing a dialogue according to the user's dialogue request, the user is identified. If the user is identified as a new user, a user portrait is generated based on the user's user information. If the user is identified as an old user, a user portrait matching the user is called.

[0010] As a further optimization of the above intelligent dialogue method: the method for generating user portraits includes: Obtain the user's basic information, which is used to record the user's identity; Get extended information about the user, including conversation history, user preferences, and user reviews; When at least one extended information is obtained, a user portrait is generated based on the basic information and the extended information; otherwise, a user portrait is generated based on the basic information.

[0011] As a further optimization of the above intelligent dialogue method: the method of optimizing the content of user questions by using a pre-trained semantic recognition model includes: Search the user's conversation history based on the user's question to obtain context fragments associated with the user's question; Perform reference disambiguation on user questions based on context fragments. If unsuccessful, generate reference information to be verified. Complete the omitted information in the user's question based on the context fragment. If it fails, generate the omitted information to be verified. If the reference information to be verified and / or the omitted information to be verified is generated, further generating the information to be verified; Use the knowledge graph model to retrieve the information to be verified, and generate verification questions based on the search results; Feedback verification issues to users and wait for user verification.

[0012] As a further optimization of the above-mentioned intelligent dialogue method: after the user verifies the verification question, the user's verification result is used to perform reference disambiguation and / or complete omitted information in the user's question to obtain a complete question.

[0013] As a further optimization of the above intelligent dialogue method: the method of constructing the knowledge graph model includes: Determine the target industry and obtain industry standard dialogue data and industry standard knowledge base data based on the target industry; Obtain actual industry conversation data and actual industry knowledge base data; After data cleaning of all the acquired receipts, basic conversation data and basic knowledge base data are generated; Annotate the basic dialogue data and basic knowledge base data, and integrate the annotated basic dialogue data and basic knowledge base data into the basic model; Perform multi-level feature extraction on the basic model and generate index rules; Integrate the basic model and index rules to obtain the knowledge graph model.

[0014] As a further optimization of the above intelligent dialogue method: the method of retrieving the complete question in the pre-built knowledge graph model includes: Perform word-by-word search on the complete question to obtain word-by-word search results; Perform word segmentation search on the complete question to obtain word segmentation search results; Perform semantic similarity retrieval on the complete question to obtain semantic retrieval results; Re-rank the word segmentation retrieval results, word segmentation retrieval structure and semantic retrieval results, and score them based on the cross entropy loss function; Generate reference text based on the highest scoring re-ranking results.

[0015] An intelligent dialogue system is used to implement the above-mentioned intelligent dialogue method, and the system includes: The interactive module is used to receive user questions and provide the final answer to the user; The strategy optimization module is used to input the user profile into the pre-trained user personalized retrieval model to determine the retrieval strategy; The question optimization module is used to optimize the content of user questions using a pre-trained semantic recognition model to obtain complete questions; The answer retrieval module is used to retrieve the complete question in the pre-built knowledge graph model based on the retrieval strategy to obtain the reference text and generate the final answer.

[0016] Computer equipment, including: Memory for storing computer programs; The processor is used to execute the computer program to implement the above-mentioned intelligent dialogue method.

[0017] The storage medium is used to store a computer program, which implements the above-mentioned intelligent dialogue method when executed.

[0018] Beneficial effects: The present invention first disambiguates concise references in multiple rounds of user conversations, and supplements the information that the user may have omitted according to the context, to ensure a complete understanding of the user's needs before searching; then, the present invention introduces a deep search technology that combines semantic understanding with knowledge graphs, which can more deeply understand the semantic connotations of user needs and achieve more accurate searches. Moreover, by utilizing the structured information of the knowledge graph, the breadth and depth of searches can be expanded to provide users with more comprehensive and relevant information; then, the system adopts a personalized search optimization strategy to perform personalized sorting and screening of search results according to the user's historical behavior, preferences and specific scenarios, so as to meet the user's personalized needs and improve the user experience; finally, the present invention introduces a reflection check mechanism to perform deep semantic analysis on the search results to ensure that the search results are highly consistent with the actual semantics of the user input, thereby significantly improving the accuracy of the search and reducing the illusion problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, an intelligent dialogue method includes S1 to S6.

[0022] S1. Establish a dialogue according to the user's dialogue request, and generate or call a user portrait. The user portrait is used to reflect the user's characteristics and their personalized needs for the dialogue process. Based on the user portrait, a better intelligent dialogue experience can be provided to the user. Further, after establishing a dialogue according to the user's dialogue request, the user is identified. If the user is identified as a new user, a user portrait is generated based on the user's user information. If the user is identified as an old user, a user portrait matching the user is called. For new users, a basic user portrait can be generated based on the information filled in during registration, and the user portrait is continuously optimized during the subsequent intelligent dialogue process so that the user portrait can more accurately reflect the user's personalized needs. Considering that some new users may not have a dialogue for a long time after registration, after the user successfully registers, the user portrait can be generated in real time when the user establishes a dialogue, thereby saving hardware resources. The method for generating a user portrait is as follows.

[0023] First, obtain the user's basic information, which is used to record the user's identity. Basic information may include gender, age, and occupation. Different genders, ages, and occupations may have different focuses, so basic information can reflect some basic personalized needs of users.

[0024] After that, the user's extended information is obtained, including conversation history, user preferences, and user evaluation. Among them, conversation history and user evaluation can only be generated after the user has had a conversation. Conversation history records may include question content, browsing history, and click behavior, etc. User preferences may include areas of interest and commonly used query methods, etc. For new users, their conversation history and user evaluation cannot be obtained, and user preferences need to be entered or selected by the user. However, some users refuse to provide user preferences for reasons such as privacy protection, so user preferences may not be obtained smoothly. In general, extended information cannot be obtained stably.

[0025] Finally, when at least one extended information is obtained, a user portrait is generated based on the basic information and the extended information, otherwise a user portrait is generated based on the basic information. Because the extended information cannot be obtained stably, when the extended information cannot be obtained at all, the user portrait is generated based on the basic information, and when at least one extended information can be obtained, the user portrait can be generated based on the basic information and the obtained extended information. The more information is used to generate the user portrait, the more accurate the generated user portrait is. It should also be noted that the method of generating a user portrait based on basic information and extended information belongs to the prior art, and those skilled in the art can choose according to actual needs, which will not be repeated here.

[0026] S2. Input the user portrait into the pre-trained user personalized retrieval model to determine the retrieval strategy. The user personalized retrieval model can be constructed based on the existing neural network model. For example, a convolutional neural network (CNN) can be used, with the user portrait as the input variable, the relevance of the retrieval results and the user satisfaction as the target variable, and the model is trained by supervised learning to obtain a user personalized retrieval model. The output of the personalized retrieval model is the retrieval strategy. After the personalized retrieval model is trained, the user portrait is input into the personalized retrieval model to output the optimal retrieval strategy. The retrieval strategy may include parameters related to the dialogue, such as the ranking weight and screening conditions of the retrieval results, which are conventional choices in this field and will not be repeated here.

[0027] S3, receiving user questions, and optimizing the content of user questions using a pre-trained semantic recognition model to obtain complete questions. Because when asking questions, users may not be able to clearly and completely express their true doubts, or because of language habits and other reasons, the user questions may be incomplete, which will lead to huge errors in the final search results, so it is necessary to optimize the user questions first. The method of optimizing the content of user questions using a pre-trained semantic recognition model includes S31 to S37.

[0028] S31. Search the user's conversation history records according to the user's question to obtain context fragments associated with the user's question.

[0029] S32. Perform reference disambiguation on the user question based on the context fragments. If unsuccessful, generate reference information to be verified. Reference disambiguation is used to clarify personal pronouns such as he, she, and it, as well as thing pronouns such as this, that, these, and those in the user question, so that the user question can clearly show the target object. Because in a conversation, when a user uses pronouns to refer to some objects, most of the time it occurs in several consecutive sentences or several sentences with a short interval. Therefore, using context fragments, reference disambiguation can be achieved more accurately.

[0030] S33. Complete the omitted information in the user's question based on the context fragment. If unsuccessful, generate the omitted information to be verified. In many cases, the user's question may omit important information, such as a specific time, place, or specific object, so it is necessary to complete the omitted information first. For common omitted information, it can be completed based on the context fragment. For example, if the user mentioned a product in the previous question, and the product name was omitted in the subsequent question, the system will automatically supplement the product information by analyzing the historical conversation records.

[0031] S34: If the reference information to be verified and / or the omitted information to be verified is generated, further generating information to be verified.

[0032] S35. Use the knowledge graph model to retrieve the information to be verified, and generate verification questions based on the search results.

[0033] S36: Feedback verification questions to the user and wait for user verification. The verification questions can be asked in the form of selection questions, that is, several options are provided to the user for selection, which can improve efficiency on the one hand and avoid more unclear content when the user edits the answer content on the other hand.

[0034] S37. After the user verifies the verification question, the user's verification result is used to perform reference disambiguation and / or complete omitted information on the user's question to obtain a complete question.

[0035] In actual situations, some user questions raised by users cannot be resolved with reference or omitted information even with the use of context fragments. In this case, reference information to be verified and / or omitted information to be verified can be sent to the user, that is, a counter-question can be asked to the user, and the user can determine and answer it, thereby achieving full optimization of the user's question and obtaining a complete question.

[0036] S4. Based on the search strategy, the complete question is searched in the pre-built knowledge graph model to obtain the reference text. In order to be able to retrieve a sufficiently accurate answer, the knowledge graph model needs to have sufficient accuracy. The method for constructing the knowledge graph model in the present invention is as follows.

[0037] The first step is to determine the target industry, and obtain industry standard dialogue data and industry standard knowledge base data based on the target industry. Because the knowledge involved in different industries and the way users ask questions may be completely different, if a knowledge graph model is constructed for the entire industry, the scale of the required data is unpredictable and difficult to achieve. Therefore, a knowledge graph model can be constructed for a specific industry, thereby reducing the amount of data required, reducing costs, and improving feasibility. For example, in one embodiment of the present invention, taking the human resources and social security industry as the target, it is first necessary to obtain the human resources and social security industry standard dialogue data and human resources and social security industry standard knowledge base data.

[0038] The second step is to obtain industry actual dialogue data and industry actual knowledge base data. After obtaining industry standard dialogue data and industry standard knowledge base data, because these standard data may be greatly different from the data generated in the practice process, for example, users will not refer to industry standard dialogue data when asking questions, and it is impossible to ask questions completely based on industry standard dialogue data. Therefore, it is necessary to obtain industry actual dialogue data and industry actual knowledge base data to complement the industry standard dialogue data and industry standard knowledge base data.

[0039] The third step is to clean all the acquired receipts to generate basic conversation data and basic knowledge base data. The data cleaning methods mainly include deduplication, typo correction, and format unification. After data cleaning, the data has higher accuracy and consistency, and basic conversation data and basic knowledge base data can be sorted out.

[0040] The fourth step is to annotate the basic dialogue data and basic knowledge base data, and integrate the annotated basic dialogue data and basic knowledge base data into the basic model. In the annotation process, focus on annotating concise references, omitted information, key entities and relationships, etc. On the basis of sufficient annotation, in S3, when performing reference disambiguation and omitted information completion for user questions, user questions can be compared with the basic dialogue data and basic knowledge base data to form options for users to select and verify.

[0041] Furthermore, a basic model in the form of a knowledge graph can be constructed based on basic conversation data and basic knowledge base data. The methods used may include entity extraction, relationship extraction, attribute extraction, entity alignment and conflict resolution. Among them, entity extraction can be carried out by combining rule-based and machine learning methods; relationship extraction can use the relationship extraction model in the deep learning algorithm to analyze the text in the knowledge base, and identify the relationship between entities through understanding the sentence structure and semantics; attribute extraction can define the attributes of various entities according to the characteristics of the knowledge base and application requirements, and extract the attribute information of entities from the text through regular expression matching, semantic analysis and other methods; conflict resolution can be judged and selected by formulating conflict resolution rules based on factors such as data reliability and authority. These methods are all common methods for constructing knowledge graphs, and will not be repeated here. On this basis, Neo4J is selected as the graph database to construct a knowledge graph. Entities can be used as nodes, and the relationship between entities can be used as edges. Both nodes and edges can contain corresponding attribute information.

[0042] The fifth step is to extract multi-level features from the basic model and generate index rules. Index rules can include indexes based on entity attributes and indexes based on relationship types. Based on index rules, it is helpful to quickly retrieve content that is highly relevant to the user's question from the knowledge graph during the conversation.

[0043] The sixth step is to integrate the basic model and index rules to obtain the knowledge graph model.

[0044] After constructing the knowledge graph model, you can first prepare the knowledge graph model for retrieval. Specifically, it includes character retrieval, word segmentation retrieval and meaning similarity retrieval for knowledge items. The method of character retrieval is: preprocess the knowledge base content, perform character segmentation on the knowledge items, split each item into a separate word sequence, and then build an inverted index based on the BM25 algorithm, and store the character-segmented items in the character-segmentation retrieval index. BM25 calculates relevance through weighted scores of word frequency, document frequency and item length. The method of word segmentation retrieval is: perform word segmentation on the knowledge base items. Use the jieba library as a word segmentation tool to split the items into multiple words or phrases, and then establish a BM25 retrieval index based on the word-segmented items to store the frequency information of the item words and phrases. The method of semantic similarity retrieval is: use the text2vec-base-chinese-sentence model to vectorize all items in the knowledge base, generate a semantic vector for each item, and determine whether the semantics are similar based on the pre-similarity between the vectors. Furthermore, the BCEreranker model is used to rerank the search results obtained through BM25 (character segmentation) and sentence-embedding. The BCEreranker model re-scores the candidate entries through the Binary Cross Entropy (BCE) loss function to evaluate their relevance to the user input. The BCEreranker model uses positive and negative sample pairs during the training phase. The positive samples are search results that users consider valid, and the negative samples are invalid search results. Through this supervised learning method, the reranker model can more accurately capture the subtle semantic differences between entries and queries.

[0045] On the basis of the completion of the construction of the above-mentioned knowledge graph model, the method for searching for a complete question in the pre-constructed knowledge graph model includes S41 to S44.

[0046] S41, performing a word-by-word search on the complete question to obtain a word-by-word search result.

[0047] S42. Perform a word segmentation search on the complete question to obtain a word segmentation search result.

[0048] S43. Perform semantic similarity retrieval on the complete question to obtain semantic retrieval results.

[0049] S44, reorder the word segmentation retrieval results, word segmentation retrieval structure and semantic retrieval results, and score them based on the cross entropy loss function.

[0050] After optimizing the user question to obtain the complete question, the method for retrieving the complete question is similar to the method for retrieval preparation, including character segmentation retrieval, word segmentation retrieval and semantic similarity retrieval.

[0051] Furthermore, a reference text is generated based on the reranking result with the highest score. The reranking tool may use the BCEreranker model, which belongs to the prior art and will not be described in detail here.

[0052] S5. Evaluate the consistency of the reference text and the complete question. If the evaluation is passed, the final answer is generated. Otherwise, the search is repeated. The consistency evaluation is used to determine whether the reference text and the complete question match. If they match, the final answer can be generated. Otherwise, the search is repeated to avoid giving the wrong answer to the user. If the complete question involves multiple aspects or contains multiple sub-questions, a comprehensive answer containing detailed answers to each question can be generated through a multi-level analysis of the knowledge base content, so that the conversation can be completed more efficiently and the user experience can be optimized.

[0053] S6. Feedback the final answer to the user.

[0054] like Figure 2 As shown, the present invention further provides an intelligent dialogue system for implementing the above-mentioned intelligent dialogue method, wherein the system includes an interaction module, a strategy optimization module, a question optimization module and an answer retrieval module.

[0055] The interactive module is used to receive user questions and provide the final answer to the user.

[0056] The strategy optimization module is used to input the user profile into the pre-trained user personalized retrieval model to determine the retrieval strategy.

[0057] The question optimization module is used to optimize the content of user questions using a pre-trained semantic recognition model to obtain complete questions.

[0058] The answer retrieval module is used to retrieve the complete question in the pre-built knowledge graph model based on the retrieval strategy to obtain the reference text and generate the final answer.

[0059] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0060] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0061] The present invention also provides a computer device, including a memory and a processor.

[0062] Memory, used to store computer programs.

[0063] The processor is used to execute the computer program to implement the above-mentioned intelligent dialogue method.

[0064] The present invention finally provides a storage medium for storing a computer program, which implements the above-mentioned intelligent dialogue method when executed.

[0065] Based on the storage medium, the technical solution of the present disclosure is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, and other media that can store program codes. The above description of the disclosed embodiments enables professionals and technicians in this field to implement or use the present invention. Various modifications to these embodiments will be obvious to professionals and technicians in this field, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent dialogue method, characterized in that: The steps include: Establish a dialogue based on the user's dialogue request and generate or call a user profile; Input the user profile into the pre-trained user personalized retrieval model to determine the retrieval strategy; Receive user questions and use the pre-trained semantic recognition model to optimize the content of user questions to obtain complete questions; Based on the retrieval strategy, the complete question is searched in the pre-built knowledge graph model to obtain the reference text; Evaluate the consistency and relevance of the reference text and the complete question. If the evaluation is passed, the final answer will be generated. Otherwise, the search will be repeated. Feedback the final answer to the user.

2. An intelligent dialogue method as claimed in claim 1, characterized in that: After establishing a conversation according to the user's conversation request, the user is identified. If the user is identified as a new user, a user profile is generated based on the user's user information. If the user is identified as an old user, a user profile matching the user is called.

3. An intelligent dialogue method as claimed in claim 1, characterized in that: The methods for generating user portraits include: Obtain the user's basic information, which is used to record the user's identity; Get extended information about the user, including conversation history, user preferences, and user reviews; When at least one extended information is obtained, a user portrait is generated based on the basic information and the extended information; otherwise, a user portrait is generated based on the basic information.

4. An intelligent dialogue method as claimed in claim 1, characterized in that: Methods for optimizing content for user questions using pre-trained semantic recognition models include: Search the user's conversation history based on the user's question to obtain context fragments associated with the user's question; Perform reference disambiguation on user questions based on context fragments. If unsuccessful, generate reference information to be verified. Complete the omitted information in the user's question based on the context fragment. If it fails, generate the omitted information to be verified. If the reference information to be verified and / or the omitted information to be verified is generated, further generating the information to be verified; Use the knowledge graph model to retrieve the information to be verified, and generate verification questions based on the search results; Feedback verification issues to users and wait for user verification.

5. An intelligent dialogue method as claimed in claim 4, characterized in that: After the user verifies the verification question, the user's verification result is used to perform reference disambiguation and / or complete omitted information on the user's question to obtain a complete question.

6. An intelligent dialogue method as claimed in claim 1, characterized in that: Methods for building knowledge graph models include: Determine the target industry and obtain industry standard dialogue data and industry standard knowledge base data based on the target industry; Obtain actual industry conversation data and actual industry knowledge base data; After cleaning all acquired data, basic conversation data and basic knowledge base data are generated; Annotate the basic dialogue data and basic knowledge base data, and integrate the annotated basic dialogue data and basic knowledge base data into the basic model; Perform multi-level feature extraction on the basic model and generate index rules; Integrate the basic model and index rules to obtain the knowledge graph model.

7. An intelligent dialogue method as claimed in claim 1, characterized in that: Methods for retrieving complete questions in pre-built knowledge graph models include: Perform word-by-word search on the complete question to obtain word-by-word search results; Perform word segmentation search on the complete question to obtain word segmentation search results; Perform semantic similarity retrieval on the complete question to obtain semantic retrieval results; Re-rank the word segmentation retrieval results, word segmentation retrieval structure and semantic retrieval results, and score them based on the cross entropy loss function; Generate reference text based on the highest scoring re-ranking results.

8. Intelligent dialogue system, characterized in that: Used to implement an intelligent dialogue method as described in any one of claims 1 to 7, the system comprises: The interactive module is used to receive user questions and provide the final answer to the user; The strategy optimization module is used to input the user profile into the pre-trained user personalized retrieval model to determine the retrieval strategy; The question optimization module is used to optimize the content of user questions using a pre-trained semantic recognition model to obtain complete questions; The answer retrieval module is used to retrieve the complete question in the pre-built knowledge graph model based on the retrieval strategy to obtain the reference text and generate the final answer.

9. Computer device, characterized in that include: Memory for storing computer programs; A processor, configured to execute the computer program to implement an intelligent dialogue method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that Used to store a computer program, which, when executed, implements an intelligent dialogue method as described in any one of claims 1-7.

Images