Intelligent question answering system and method for cross-domain knowledge reasoning

Abstract

The invention discloses an intelligent question answering system and method for cross-domain knowledge reasoning, and relates to the technical field of intelligent question answering. In order to solve the problem that an existing intelligent question answering system is limited by technical shortages and cannot give satisfactory reply to a user in cross-domain question processing, the adopted scheme comprises the steps that a domain knowledge base construction module collects multi-domain knowledge data, and the multi-domain knowledge data is stored in a corresponding domain vector knowledge base after being preprocessed and vectorized; the user interaction module provides an interface, receives input, interacts and returns a result; the field identification module identifies the field and extracts key information by using a large language model; the knowledge retrieval module retrieves related information accordingly; the cross-domain knowledge reasoning module is used for reasoning a knowledge logic relationship of each domain and guiding a large language model to integrate fused knowledge; and the intelligent question-answering module generates prompt words, inputs the prompt words into the large language model to obtain a final response, and returns each result to the user interaction module for display. According to the method, knowledge from multiple fields can be processed, and the method is suitable for various different types of problems and scenes.

Description

Technical Field

[0001] The present invention relates to the field of intelligent question answering technology, and in particular to an intelligent question answering system and method for cross-domain knowledge reasoning. Background Art

[0002] In the booming development of artificial intelligence, intelligent question-answering systems, as one of the key applications, are tasked with digging deep into structured and unstructured data with the help of natural language processing technology and accurately outputting answers. At present, intelligent question-answering systems mainly rely on three major technical pillars: knowledge graphs, rule matching, and retrieval-based models. In exclusive single-domain scenarios, they perform well and can answer questions more efficiently.

[0003] Take the medical field as an example. If the question is only about common diseases and treatment processes in a certain department, the traditional question-answering system will give a reliable answer with a high probability by relying on the knowledge base of the corresponding department, combined with established rules and graph associations. However, once you jump out of the "comfort zone" of a single field and face complex, cross-domain problems, the existing intelligent question-answering system will be unable to cope with it.

[0004] The first bottleneck is the technical level of domain knowledge retrieval. Traditional knowledge retrieval methods, such as keyword matching, Boolean retrieval, and simple semantic matching, are very capable in a single domain, just like driving smoothly on a familiar one-way street and accurately locating the required knowledge. However, when the problem spans multiple domains and the situation takes a sharp turn for the worse, the existing technology is like a traveler who has lost his navigation and is at a loss. This is because it can only mechanically search each domain separately, and it completely lacks the ability to integrate and relate knowledge from different domains.

[0005] Focusing on the intelligent question-answering system itself, its limitations become increasingly prominent. Most systems are like being tied to a single field by invisible ropes, and knowledge retrieval and reasoning can only be used in familiar territory. They are accustomed to relying on knowledge bases in specific fields, and hastily piece together answers by keyword matching and simple semantic analysis. Imagine a question about health and nutritional diet, which may involve knowledge in multiple fields such as medical care, food nutrition, etc., but traditional question-answering systems are constrained by a single field knowledge base and it is difficult to give comprehensive and integrated suggestions.

[0006] A detailed study of the problems exposed by the cross-field application of existing technologies mainly focuses on three points: The first is the difficulty of cross-domain information integration. The knowledge bases in various fields are like isolated islands, isolated from each other and operating independently. When users raise cross-domain questions, although the system tries to search for information from different "islands", it cannot cleverly build bridges and effectively integrate them. The answers given in the end are fragmented and the accuracy is greatly reduced.

[0007] The second is the weakness of logical reasoning. Faced with complex information interwoven in multiple fields, the system lacks the ability to deeply analyze and unravel the underlying logical context of information in different fields. The generated answers are like jigsaw puzzle pieces with no order, piled up in a mess, lacking coherent logic, leaving users confused.

[0008] Third, the accuracy of knowledge retrieval is poor. The system does not seem to have truly understood the essence of user input, and it is difficult to accurately identify the multiple fields involved in the question. It is like blindly groping in the ocean of knowledge with your eyes covered. The retrieved field knowledge often deviates greatly from the actual needs, which in turn lowers the performance level of the entire question-answering system and fails to meet users' expectations for accurate answers.

[0009] To sum up, although the existing intelligent question-answering systems have made some achievements in a single field, they are still unable to provide users with satisfactory answers when dealing with cross-domain problems due to technical shortcomings. They are in urgent need of technological innovation and breakthroughs to break down domain barriers and achieve knowledge fusion and efficient reasoning. Summary of the invention

[0010] In view of the needs and deficiencies of current technological development, the present invention provides an intelligent question-answering system and method for cross-domain knowledge reasoning.

[0011] In the first aspect, the present invention provides an intelligent question-answering system for cross-domain knowledge reasoning, and the technical solutions adopted to solve the above technical problems are as follows: An intelligent question-answering system for cross-domain knowledge reasoning, comprising: The domain knowledge base construction module is used to collect knowledge data in multiple fields, perform preprocessing and vectorization operations, and store them in the constructed domain vector knowledge base accordingly; The user interaction module provides an interactive interface for receiving user input and interacting with the user to obtain the user's detailed needs, and is also used to return the response results to the user input, the search results in various fields, and the integrated cross-domain fusion knowledge; The domain identification module is used to identify the corresponding domain of the domain vector knowledge base involved in the user input, and use the large language model A to extract key information of the identified domain from the user input; The knowledge retrieval module is used to retrieve information related to the user input from the corresponding domain vector knowledge base based on the key information extracted by the domain identification module; The cross-domain knowledge reasoning module is used to obtain search results from all fields. By prompting the logical relationship between knowledge in various fields, the large language model B is guided to integrate multi-domain knowledge and output logically coherent cross-domain fusion knowledge. The intelligent question-answering module is used to further integrate the fused knowledge with the user input to generate prompt words, input the large language model C to obtain the final response, and return the response generated by the large language model C, the retrieval results of the knowledge retrieval module, and the cross-domain fused knowledge output by the cross-domain knowledge reasoning module to the user interaction module for display.

[0012] Optionally, the domain knowledge base construction module involved collects knowledge data in multiple domains in an offline manner, and then performs preprocessing operations of data cleaning and slicing, and then uses the bge-m3 model to vectorize the preprocessed data, and stores the results in the corresponding domain vector knowledge base. The domain vector knowledge base adopts the Milvus knowledge base, and the data storage format is in the form of (embedding, text), where embedding is vector data and text is the corresponding text data.

[0013] Optionally, the domain identification module identifies the corresponding domain of the domain vector knowledge base involved in the user input, and uses the large language model A to extract key information of the identified domain from the user input. This process specifically includes: Determine the domain covered by the domain vector knowledge base constructed by the domain knowledge base construction module, and use the prompt engineering to generate prompt words for the large language model A; The large language model A first identifies the corresponding domain of the domain vector knowledge base involved in the user input based on the user input, and then extracts the key information of the identified domain from the user input. The key information is output in json format, with the key being the domain and the value being the corresponding key information.

[0014] Optionally, the knowledge retrieval module involved first converts the key information extracted by the domain identification module into a vector, and then uses cosine similarity to calculate the similarity with the knowledge vectors stored in the vector knowledge base of different domains. The calculation formula is as follows: , in, Represents the inner product of vector M and vector N, that is, the result of multiplying and adding corresponding elements; represents the length of vector M, which is obtained by summing the squares of each element in vector M and then taking the square root; Represents the length of vector N, which is obtained by summing the squares of each element in vector N and then taking the square root.

[0015] Optionally, a large language model A based on a Transformer architecture is trained in advance using a data set containing text samples from multiple fields and annotated with corresponding field labels, and the field identification module uses the large language model A to extract key information of the identified field from the user input; A large language model B based on the Transformer architecture is trained in advance using a data set containing knowledge from multiple fields and having logical relationships between the knowledge. The cross-domain knowledge reasoning module guides the large language model B to integrate the knowledge from multiple fields and output logically coherent cross-domain fusion knowledge. A large language model C based on the Transformer architecture is trained in advance using a dataset containing question-answer pairs and contextual information related to the questions. The intelligent question-answering module further integrates the fused knowledge with the user input to generate prompt words, which are input into the large language model C to obtain the final response.

[0016] In the second aspect, the present invention provides an intelligent question-answering method for cross-domain knowledge reasoning, and the technical solution adopted to solve the above technical problems is as follows: An intelligent question answering method for cross-domain knowledge reasoning includes the following implementation steps: S1. Collect knowledge data in multiple fields, perform preprocessing and vectorization operations, and store them in the constructed field vector knowledge base accordingly; S2, providing an interactive interface, on the one hand, receiving user input and interacting with the user to obtain the user's detailed needs, and on the other hand, returning the response results to the user input, the search results in various fields, and the integrated cross-domain fusion knowledge; S3, identifying the corresponding domain of the domain vector knowledge base involved in the user input, and using the large language model A to extract key information of the identified domain from the user input; S4, according to the key information extracted by the domain identification module, retrieve the information related to the user input from the corresponding domain vector knowledge base; S5. Obtain search results from all fields, and guide the large language model B to integrate multi-field knowledge by prompting the logical relationship between knowledge in various fields of engineering reasoning, and output logically coherent cross-field fusion knowledge; S6. Further integrate the fused knowledge with the user input to generate prompt words, input the large language model B to obtain the final response, and return the response generated by the large language model B, the retrieval result of step S4, and the cross-domain fused knowledge output of step S5 to the interactive interface for display.

[0017] Optionally, execute step S1 to collect knowledge data in multiple fields in an offline manner, then perform data cleaning and segmentation preprocessing operations, and then use the bge-m3 model to vectorize the preprocessed data, and store the results in the corresponding domain vector knowledge base. The domain vector knowledge base uses the Milvus knowledge base, and the data storage format is in the form of (embedding, text), where embedding is vector data and text is the corresponding text data.

[0018] Optionally, step S3 specifically includes the following process: Determine the domain covered by the domain vector knowledge base constructed by the domain knowledge base construction module, and use the prompt engineering to generate prompt words for the large language model A; The large language model A first identifies the corresponding domain of the domain vector knowledge base involved in the user input based on the user input, and then extracts the key information of the identified domain from the user input. The key information is output in json format, with the key being the domain and the value being the corresponding key information.

[0019] Optionally, the step S4 specifically includes the following process: First, the key information extracted by the domain identification module is converted into a vector; Then, the cosine similarity method is used to calculate the similarity with the knowledge vectors stored in the vector knowledge base of different fields. The calculation formula is as follows: , in, Represents the inner product of vector M and vector N, that is, the result of multiplying and adding corresponding elements; represents the length of vector M, which is obtained by summing the squares of each element in vector M and then taking the square root; Represents the length of vector N, which is obtained by summing the squares of each element in vector N and then taking the square root.

[0020] Optionally, before executing step S3, a large language model A based on the Transformer architecture is pre-trained using a data set containing text samples from multiple fields and annotated with corresponding field labels; after executing step S3, the corresponding field of the field vector knowledge base involved in the user input is identified, and the key information of the identified field is extracted from the user input using the large language model A; Before executing step S5, a large language model B based on the Transformer architecture is trained in advance using a data set containing knowledge from multiple fields and having logical relationships between these knowledge; executing step S5, the search results of all fields are obtained, and after prompting the engineering reasoning of the logical relationship between the knowledge in each field, the large language model B is guided to integrate the knowledge from multiple fields and output logically coherent cross-field fusion knowledge; Before executing step S6, a large language model C based on the Transformer architecture is trained in advance using a data set containing question-answer pairs and contextual information related to the question; after executing step S6, the fused knowledge is further integrated with the user input to generate prompt words, which are then input into the large language model C to obtain the final response.

[0021] The cross-domain knowledge reasoning intelligent question answering system and method of the present invention has the following beneficial effects compared with the prior art: 1. The present invention can realize cross-domain knowledge integration of complex problems, improve the accuracy and reliability of question-answering systems and methods; can effectively identify and extract key information in various fields, and accurately generate answers through reasoning, ensuring the accuracy and logic of the answers; 2. The present invention can accurately identify the domain to which the user input belongs, extract key information, and retrieve relevant information from the corresponding domain vector knowledge base, which can significantly improve the accuracy and relevance of the answer; it can guide the large language model to integrate knowledge from different fields through prompt engineering, and output logically coherent cross-domain fusion knowledge, which not only helps to solve complex problems, but also promotes the generation of new knowledge and the development of innovative thinking; it provides a friendly interface that allows users to easily input questions and interact, which not only returns direct answers, but also displays retrieval results and cross-domain fusion knowledge, providing users with richer information and deeper understanding. This interactive method can enhance the user's sense of participation and satisfaction; 3. The present invention can process knowledge from multiple fields and adapt to various types of problems and scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Attached Figure 1 is a module connection block diagram of the first embodiment of the present invention; Attached Figure 2 This is a flowchart of the method implementation of the second embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to make the technical solution, the technical problem solved and the technical effect of the present invention more clearly understood, the technical solution of the present invention is clearly and completely described below in conjunction with specific embodiments.

[0024] Embodiment 1: Combining with Figure 1 , this embodiment proposes an intelligent question-answering system for cross-domain knowledge reasoning, which includes: The domain knowledge base construction module is used to collect knowledge data in multiple fields, perform preprocessing and vectorization operations, and store them in the constructed domain vector knowledge base accordingly; The user interaction module provides an interactive interface for receiving user input and interacting with the user to obtain the user's detailed needs, and is also used to return the response results to the user input, the search results in various fields, and the integrated cross-domain fusion knowledge; The domain identification module is used to identify the corresponding domain of the domain vector knowledge base involved in the user input, and use the large language model A to extract key information of the identified domain from the user input; The knowledge retrieval module is used to retrieve information related to the user input from the corresponding domain vector knowledge base based on the key information extracted by the domain identification module; The cross-domain knowledge reasoning module is used to obtain search results from all fields. By prompting the logical relationship between knowledge in various fields, the large language model B is guided to integrate multi-domain knowledge and output logically coherent cross-domain fusion knowledge. The intelligent question-answering module is used to further integrate the fused knowledge with the user input to generate prompt words, input the large language model C to obtain the final response, and return the response generated by the large language model C, the retrieval results of the knowledge retrieval module, and the cross-domain fused knowledge output by the cross-domain knowledge reasoning module to the user interaction module for display.

[0025] In this embodiment, the domain knowledge base construction module involved collects knowledge data in multiple fields in an offline manner, and then performs data cleaning and segmentation preprocessing operations, wherein data cleaning includes removing meaningless text data such as garbled characters and hyperlinks, and deduplicating text data. Data segmentation includes segmenting long texts according to fixed lengths (for example, 250 characters), and then using the bge-m3 model to vectorize the preprocessed data, and storing the results in the corresponding domain vector knowledge base. The domain vector knowledge base adopts the Milvus knowledge base, and the data storage format is in the form of (embedding, text), wherein embedding is vector data, and text is the corresponding text data.

[0026] In this embodiment, the domain identification module identifies the corresponding domain of the domain vector knowledge base involved in the user input, and uses the large language model A to extract key information of the identified domain from the user input. This process specifically includes: Determine the domain covered by the domain vector knowledge base constructed by the domain knowledge base construction module, and use the prompt engineering to generate prompt words for the large language model A; The large language model A first identifies the corresponding domain of the domain vector knowledge base involved in the user input based on the user input, and then extracts the key information of the identified domain from the user input. The key information is output in json format, with the key being the domain and the value being the corresponding key information.

[0027] In this embodiment, the knowledge retrieval module first converts the key information extracted by the domain identification module into a vector, and then uses the cosine similarity method to calculate the similarity with the knowledge vectors stored in the vector knowledge base of different domains. The calculation formula is as follows: , in, Represents the inner product of vector M and vector N, that is, the result of multiplying and adding corresponding elements; represents the length of vector M, which is obtained by summing the squares of each element in vector M and then taking the square root; Represents the length of vector N, which is obtained by summing the squares of each element in vector N and then taking the square root.

[0028] It is necessary to add that: (1) A large language model A based on the Transformer architecture is trained in advance using a dataset containing text samples from multiple fields and annotated with corresponding field labels. The field identification module uses the large language model A to extract key information of the identified field from user input. Specifically, the pre-training process includes: (1.1) Collecting text samples from multiple fields, such as news, academic papers, novels, industry reports, etc., and ensuring that each sample is labeled with the corresponding field label, such as "medicine", "finance", "technology", etc. These sample data can come from public data sets, professional literature, website content, etc.; (1.2) Performing pre-processing operations such as cleaning, word segmentation, and encoding on the text to meet the model input requirements; (1.3) Selecting a Transformer-based architecture, such as BERT, RoBERTa, or ALBERT, these models perform well in text classification and information extraction tasks; (1.4) Fine-tuning the model using a data set with domain labels so that it learns to recognize the field of the text. At the same time, the model is trained to extract key information from the text. This can be achieved through sequence labeling tasks, such as named entity recognition (NER) or keyword extraction; (1.5) Using the validation set to evaluate the performance of the model, adjust the hyperparameters or model structure as needed to improve the accuracy and recall. Through steps (1.1) to (1.5), the large language model A of this embodiment can be obtained and applied to the domain identification module.

[0029] (2) Pre-train a large language model B based on the Transformer architecture using a dataset containing knowledge from multiple fields and logical relationships between these knowledge. The cross-domain knowledge reasoning module guides the large language model B to integrate knowledge from multiple fields and output logically coherent cross-domain fusion knowledge. Specifically, the pre-training process includes: (2.1) Collecting text data containing knowledge from multiple fields and logical relationships between these knowledge, such as research reports on the impact of scientific and technological development on social economy, academic literature in the intersection of medicine and biology, etc.; (2.2) Pre-processing these data, including cleaning, word segmentation, part-of-speech tagging and other operations, so that they can be better input into the model for training; (2.3) Selecting a Transformer-based architecture that can handle complex logical relationships and long-distance dependencies, such as GPT-3 or its variants; (2.4) Converting text data into a vector representation that the model can handle, mapping words to a low-dimensional vector space through methods such as word embedding, and at the same time, using the position encoding in the Transformer architecture to capture the sequential information in the text, so as to better understand the logical relationship between knowledge; (2.5) Given a piece of text in Part of the content, let the model predict the subsequent content and the logical relationship between them, such as causal relationship, progressive relationship, etc. In this way, the model learns the logical connection pattern between knowledge in different fields; (2.6) Provide relevant text fragments in multiple fields, let the model try to integrate these fragments, and generate a logically coherent text containing multi-domain knowledge fusion, which helps the model to exercise the ability of cross-domain knowledge reasoning and integration, so that it can output logically coherent cross-domain fusion knowledge in practical applications; (2.7) In the training process, use a large amount of pre-training data to iterate the model for multiple times. According to the performance of the model on the pre-training task, such as the accuracy of logical relationship prediction, the rationality of cross-domain knowledge integration, etc., the parameters of the model are continuously adjusted through the back propagation algorithm to optimize the performance of the model, so that it can better grasp the logical relationship and integration method between cross-domain knowledge. After steps (2.1)-(2.7), the large language model B of this embodiment can be applied to the cross-domain knowledge reasoning module.

[0030] (3) A large language model C based on the Transformer architecture is trained in advance using a dataset containing question-answer pairs and contextual information related to the question. The intelligent question-answering module further integrates the fused knowledge with the user input to generate prompt words, which are input into the large language model C to obtain the final response. Specifically, the pre-training process includes: (3.1) Collecting a dataset containing question-answer pairs and contextual information related to the questions, such as data from FAQ websites, customer service conversation records, question-answer posts in online forums, etc., to ensure that the data covers various fields and topics; (3.2) Cleaning and screening the data, removing noise and invalid data, improving data quality, and properly organizing and annotating the question-answer pairs and contextual information so that the model can better understand and learn; (3.3) Selecting a Transformer-based architecture that performs well in generation tasks, such as T5 or other variants that focus on text generation; (3.4) Encoding the question-answer pairs and contextual information into a vector form that the model can accept; for questions and answers, they can be encoded separately, and at the same time, the contextual information can be encoded in association with the question or answer so that the model can make full use of the contextual information to generate more accurate answers; (3.5) Given a question and related contextual information information, so that the model can generate the corresponding answer; through a large amount of question-answering data training, the model learns the answering methods and techniques of different types of questions, and how to generate more reasonable and accurate answers based on context information; (3.6) Provide a text containing multiple question-answer pairs and related contexts, so that the model can understand and integrate the information therein, and then generate corresponding answers based on specific questions; this task aims to improve the model's ability to understand and utilize context information, so that it can accurately obtain the required information in a complex text environment and generate coherent answers related to the context; (3.7) During the pre-training process, according to the evaluation indicators of the model in question-answering generation and context understanding and fusion tasks, such as answer accuracy, recall rate, BLEU value, etc., the model parameters are adjusted and optimized; by continuously increasing training data, adjusting hyperparameters, etc., the generalization ability and performance of the model are improved, so that it can better cope with various types of question-answering tasks and generate high-quality final responses. After steps (3.1)-(3.7), the large language model C of this embodiment can be obtained and applied to the intelligent question-answering module.

[0031] Embodiment 2: Combining with Figure 2 This embodiment proposes an intelligent question answering method for cross-domain knowledge reasoning, which includes the following implementation steps: S1. Collect knowledge data in multiple fields, perform preprocessing and vectorization operations, and store them in the constructed field vector knowledge base accordingly.

[0032] Specifically, knowledge data in multiple fields are collected offline, and then preprocessing operations of data cleaning and segmentation are performed, wherein data cleaning includes removing meaningless text data such as garbled characters and hyperlinks, and deduplicating text data; data segmentation includes segmenting long texts according to fixed lengths (e.g., 250 characters), and then the bge-m3 model is used to vectorize the preprocessed data, and the results are stored in the corresponding domain vector knowledge base, which adopts the Milvus knowledge base, and the data storage format is in the form of (embedding, text), wherein embedding is vector data, and text is the corresponding text data.

[0033] S2. Provide an interactive interface that receives user input and interacts with users to obtain detailed user needs, and returns response results to user input, search results in various fields, and integrated cross-domain fusion knowledge.

[0034] S3, identifying the corresponding domain of the domain vector knowledge base involved in the user input, and using the large language model A to extract key information of the identified domain from the user input. This process specifically includes: Determine the domain covered by the domain vector knowledge base constructed by the domain knowledge base construction module, and use the prompt engineering to generate prompt words for the large language model A; The large language model A first identifies the corresponding domain of the domain vector knowledge base involved in the user input based on the user input, and then extracts the key information of the identified domain from the user input. The key information is output in json format, with the key being the domain and the value being the corresponding key information.

[0035] S4. According to the key information extracted by the domain identification module, the information related to the user input is retrieved from the corresponding domain vector knowledge base. In this process, the key information extracted by the domain identification module is first converted into a vector, and then the cosine similarity is used to calculate the similarity with the knowledge vectors stored in different domain vector knowledge bases. The calculation formula is as follows: , in, Represents the inner product of vector M and vector N, that is, the result of multiplying and adding corresponding elements; represents the length of vector M, which is obtained by summing the squares of each element in vector M and then taking the square root; Represents the length of vector N, which is obtained by summing the squares of each element in vector N and then taking the square root.

[0036] S5. Obtain search results from all fields, and guide the large language model B to integrate multi-field knowledge by prompting the logical relationship between knowledge in various fields of engineering reasoning, and output logically coherent cross-field fusion knowledge.

[0037] S6. Further integrate the fused knowledge with the user input to generate prompt words, input the large language model C to obtain the final response, and return the response generated by the large language model C, the retrieval result of step S4, and the cross-domain fused knowledge output of step S5 to the interactive interface for display.

[0038] It is necessary to add that: (1) Before executing step S3, a large language model A based on the Transformer architecture is trained in advance using a data set containing text samples from multiple fields and annotated with corresponding field labels; after executing step S3, the corresponding field of the field vector knowledge base involved in the user input is identified, and the large language model A is used to extract key information of the identified field from the user input. Specifically, the pre-training process includes: (1.1) Collecting text samples from multiple fields, such as news, academic papers, novels, industry reports, etc., and ensuring that each sample is labeled with the corresponding field label, such as "medicine", "finance", "technology", etc. These sample data can come from public data sets, professional literature, website content, etc.; (1.2) Performing pre-processing operations such as cleaning, word segmentation, and encoding on the text to meet the model input requirements; (1.3) Selecting a Transformer-based architecture, such as BERT, RoBERTa, or ALBERT, these models perform well in text classification and information extraction tasks; (1.4) Fine-tuning the model using a data set with domain labels so that it learns to recognize the field of the text. At the same time, the model is trained to extract key information from the text. This can be achieved through sequence labeling tasks, such as named entity recognition (NER) or keyword extraction; (1.5) Using the validation set to evaluate the performance of the model, adjust the hyperparameters or model structure as needed to improve the accuracy and recall. After steps (1.1) to (1.5), the large language model A of this embodiment can be obtained and applied to step S3.

[0039] (2) Before executing step S5, a large language model B based on the Transformer architecture is trained in advance using a data set containing knowledge from multiple fields and having logical relationships between such knowledge; executing step S5, the search results of all fields are obtained, and after prompting the engineering reasoning about the logical relationship between the knowledge in various fields, the large language model B is guided to integrate the knowledge from multiple fields and output logically coherent cross-domain fusion knowledge. Specifically, the pre-training process includes: (2.1) collecting text data that contains knowledge from multiple fields and has logical relationships between knowledge, such as research reports on the impact of scientific and technological development on social economy, academic literature in the intersection of medicine and biology, etc.; (2.2) pre-processing these data, including cleaning, word segmentation, part-of-speech tagging and other operations, so that they can be better input into the model for training; (2.3) selecting a Transformer-based architecture that can handle complex logical relationships and long-distance dependencies, such as GPT-3 or its variants; (2.4) converting text data into a vector representation that the model can handle, mapping words to a low-dimensional vector space through methods such as word embedding, and at the same time, using the position encoding in the Transformer architecture to capture the sequential information in the text, so as to better understand the logical relationship between knowledge; (2.5) given a piece of text Part of the content, let the model predict the subsequent content and the logical relationship between them, such as causal relationship, progressive relationship, etc. In this way, the model learns the logical connection pattern between knowledge in different fields; (2.6) Provide relevant text fragments in multiple fields, let the model try to integrate these fragments, and generate a logically coherent text containing multi-domain knowledge fusion, which helps the model to exercise the ability of cross-domain knowledge reasoning and integration, so that it can output logically coherent cross-domain fusion knowledge in practical applications; (2.7) In the training process, use a large amount of pre-training data to iterate the model for multiple times. According to the performance of the model on the pre-training task, such as the accuracy of logical relationship prediction, the rationality of cross-domain knowledge integration, etc., the parameters of the model are continuously adjusted through the back propagation algorithm to optimize the performance of the model, so that it can better grasp the logical relationship and integration method between cross-domain knowledge. After steps (2.1)-(2.7), the large language model B of this embodiment can be obtained and applied to step S5.

[0040] (3) Before executing step S6, a large language model C based on the Transformer architecture is trained in advance using a dataset containing question-answer pairs and contextual information related to the question; after executing step S6, the fused knowledge is further integrated with the user input to generate a prompt word, which is then input into the large language model C to obtain the final response. Specifically, the pre-training process includes: (3.1) Collecting a dataset containing question-answer pairs and contextual information related to the questions, such as data from FAQ websites, customer service conversation records, question-answer posts in online forums, etc., to ensure that the data covers various fields and topics; (3.2) Cleaning and screening the data, removing noise and invalid data, improving data quality, and properly organizing and annotating the question-answer pairs and contextual information so that the model can better understand and learn; (3.3) Encoding the question-answer pairs and contextual information and converting them into a vector form that the model can accept; for questions and answers, they can be encoded separately, and at the same time, the contextual information can be associated with the question or answer so that the model can make full use of the contextual information to generate more accurate answers; (3.4) Selecting a Transformer-based architecture that performs well in generation tasks, such as T5 or other variants that focus on text generation; (3.5) Given a question and related contextual information information, so that the model can generate the corresponding answer; through a large amount of question-answering data training, the model can learn the answering methods and techniques of different types of questions, as well as how to generate more reasonable and accurate answers based on context information; (3.6) Provide a text containing multiple question-answer pairs and related contexts, so that the model can understand and integrate the information therein, and then generate corresponding answers based on specific questions; this task aims to improve the model's ability to understand and utilize context information, so that it can accurately obtain the required information in a complex text environment and generate coherent answers related to the context; (3.7) During the pre-training process, according to the evaluation indicators of the model in question-answering generation and context understanding and fusion tasks, such as answer accuracy, recall rate, BLEU value, etc., the model parameters are adjusted and optimized; by continuously increasing training data, adjusting hyperparameters, etc., the generalization ability and performance of the model are improved, so that it can better cope with various types of question-answering tasks and generate high-quality final responses. After steps (3.1)-(3.7), the large language model C of this embodiment can be obtained and applied to step S6.

[0041] In summary, the intelligent question-answering system and method using cross-domain knowledge reasoning of the present invention can realize cross-domain knowledge integration of complex problems and improve the accuracy and reliability of questions and answers; it can effectively identify and extract key information in various fields, and accurately generate answers through reasoning, ensuring the accuracy and logic of the answers.

[0042] The above specific examples are used to explain the principles and implementation methods of the present invention in detail. These examples are only used to help understand the core technical content of the present invention. Based on the above specific embodiments of the present invention, any improvements and modifications made by technicians in this technical field without departing from the principles of the present invention should fall within the scope of patent protection of the present invention.

Claims

1. An intelligent question-answering system for cross-domain knowledge reasoning, characterized in that: It includes: The domain knowledge base construction module is used to collect knowledge data in multiple fields, perform preprocessing and vectorization operations, and store them in the constructed domain vector knowledge base accordingly; The user interaction module provides an interactive interface for receiving user input and interacting with the user to obtain the user's detailed needs, and is also used to return the response results to the user input, the search results in various fields, and the integrated cross-domain fusion knowledge; The domain identification module is used to identify the corresponding domain of the domain vector knowledge base involved in the user input, and use the large language model A to extract key information of the identified domain from the user input; The knowledge retrieval module is used to retrieve information related to the user input from the corresponding domain vector knowledge base based on the key information extracted by the domain identification module; The cross-domain knowledge reasoning module is used to obtain search results from all fields. By prompting the logical relationship between knowledge in various fields, the large language model B is guided to integrate multi-domain knowledge and output logically coherent cross-domain fusion knowledge. The intelligent question-answering module is used to further integrate the fused knowledge with the user input to generate prompt words, input the large language model C to obtain the final response, and return the response generated by the large language model C, the retrieval results of the knowledge retrieval module, and the cross-domain fused knowledge output by the cross-domain knowledge reasoning module to the user interaction module for display.

2. The cross-domain knowledge reasoning intelligent question answering system according to claim 1, characterized in that: The domain knowledge base construction module collects knowledge data in multiple fields in an offline manner, then performs preprocessing operations such as data cleaning and slicing, and then uses the bge-m3 model to vectorize the preprocessed data, and stores the results in the corresponding domain vector knowledge base. The domain vector knowledge base uses the Milvus knowledge base, and the data storage format is in the form of (embedding, text), where embedding is vector data and text is the corresponding text data.

3. The cross-domain knowledge reasoning intelligent question answering system according to claim 1, characterized in that: The domain identification module identifies the corresponding domain of the domain vector knowledge base involved in the user input, and uses the large language model A to extract key information of the identified domain from the user input. This process specifically includes: Determine the domain covered by the domain vector knowledge base constructed by the domain knowledge base construction module, and use the prompt engineering to generate prompt words for the large language model A; The large language model A first identifies the corresponding domain of the domain vector knowledge base involved in the user input based on the user input, and then extracts the key information of the identified domain from the user input. The key information is output in json format, with the key being the domain and the value being the corresponding key information.

4. The cross-domain knowledge reasoning intelligent question answering system according to claim 1, characterized in that: The knowledge retrieval module first converts the key information extracted by the domain identification module into a vector, and then uses the cosine similarity method to calculate the similarity with the knowledge vectors stored in the vector knowledge base of different domains. The calculation formula is as follows: ， in, Represents the inner product of vector M and vector N, that is, the result of multiplying and adding corresponding elements; represents the length of vector M, which is obtained by summing the squares of each element in vector M and then taking the square root; Represents the length of vector N, which is obtained by summing the squares of each element in vector N and then taking the square root.

5. The cross-domain knowledge reasoning intelligent question answering system according to claim 1, characterized in that: A large language model A based on the Transformer architecture is trained in advance using a data set containing text samples from multiple fields and annotated with corresponding field labels. The field identification module uses the large language model A to extract key information of the identified field from user input; A large language model B based on the Transformer architecture is trained in advance using a data set containing knowledge from multiple fields and having logical relationships between the knowledge. The cross-domain knowledge reasoning module guides the large language model B to integrate the knowledge from multiple fields and output logically coherent cross-domain fusion knowledge. A large language model C based on the Transformer architecture is trained in advance using a dataset containing question-answer pairs and contextual information related to the questions. The intelligent question-answering module further integrates the fused knowledge with the user input to generate prompt words, which are input into the large language model C to obtain the final response.

6. An intelligent question-answering method for cross-domain knowledge reasoning, characterized in that: It includes the following implementation steps: S1. Collect knowledge data in multiple fields, perform preprocessing and vectorization operations, and store them in the constructed field vector knowledge base accordingly; S2, providing an interactive interface, on the one hand, receiving user input and interacting with the user to obtain the user's detailed needs, and on the other hand, returning the response results to the user input, the search results in various fields, and the integrated cross-domain fusion knowledge; S3, identifying the corresponding domain of the domain vector knowledge base involved in the user input, and using the large language model A to extract key information of the identified domain from the user input; S4, according to the key information extracted by the domain identification module, retrieve the information related to the user input from the corresponding domain vector knowledge base; S5. Obtain search results from all fields, and guide the large language model B to integrate multi-field knowledge by prompting the logical relationship between knowledge in various fields of engineering reasoning, and output logically coherent cross-field fusion knowledge; S6. Further integrate the fused knowledge with the user input to generate prompt words, input the large language model C to obtain the final response, and return the response generated by the large language model C, the retrieval result of step S4, and the cross-domain fused knowledge output of step S5 to the interactive interface for display.

7. The intelligent question-answering method for cross-domain knowledge reasoning according to claim 6, characterized in that: Execute step S1 to collect knowledge data in multiple fields in an offline manner, then perform data cleaning and slicing preprocessing operations, and then use the bge-m3 model to vectorize the preprocessed data, and store the results in the corresponding field vector knowledge base. The field vector knowledge base uses the Milvus knowledge base, and the data storage format is in the form of (embedding, text), where embedding is vector data and text is the corresponding text data.

8. The intelligent question-answering method for cross-domain knowledge reasoning according to claim 6, characterized in that: The step S3 specifically includes the following process: Determine the domain covered by the domain vector knowledge base constructed by the domain knowledge base construction module, and use the prompt engineering to generate prompt words for the large language model A; The large language model A first identifies the corresponding domain of the domain vector knowledge base involved in the user input based on the user input, and then extracts the key information of the identified domain from the user input. The key information is output in json format, with the key being the domain and the value being the corresponding key information.

9. The intelligent question-answering method for cross-domain knowledge reasoning according to claim 6, characterized in that: The step S4 specifically includes the following process: First, the key information extracted by the domain identification module is converted into a vector; Then, the cosine similarity method is used to calculate the similarity with the knowledge vectors stored in the vector knowledge base of different fields. The calculation formula is as follows: ， in, Represents the inner product of vector M and vector N, that is, the result of multiplying and adding corresponding elements; represents the length of vector M, which is obtained by summing the squares of each element in vector M and then taking the square root; Represents the length of vector N, which is obtained by summing the squares of each element in vector N and then taking the square root.

10. The intelligent question-answering method for cross-domain knowledge reasoning according to claim 6, characterized in that: Before executing step S3, a large language model A based on the Transformer architecture is trained in advance using a data set containing text samples from multiple fields and annotated with corresponding field labels; after executing step S3, the corresponding field of the field vector knowledge base involved in the user input is identified, and the large language model A is used to extract key information of the identified field from the user input; Before executing step S5, a large language model B based on the Transformer architecture is trained in advance using a data set containing knowledge from multiple fields and having logical relationships between these knowledge; executing step S5, the search results of all fields are obtained, and after prompting the engineering reasoning of the logical relationship between the knowledge in each field, the large language model B is guided to integrate the knowledge from multiple fields and output logically coherent cross-field fusion knowledge; Before executing step S6, a large language model C based on the Transformer architecture is trained in advance using a data set containing question-answer pairs and contextual information related to the question; after executing step S6, the fused knowledge is further integrated with the user input to generate prompt words, which are then input into the large language model C to obtain the final response.

Images