Question and answer model training method oriented to specific field and intelligent question and answer method and device

Abstract

The invention provides a question and answer model training method and device and an intelligent question and answer method and device for a specific field, and relates to technologies such as large models, model training and deep learning in the field of artificial intelligence and the technical fields such as intelligent search and intelligent question and answer. According to the specific implementation scheme, corpus data of a specific field are obtained; meta-knowledge question and answer data are constructed based on the corpus data through a first model, and the meta-knowledge question and answer data comprise meta-questions and answers corresponding to the meta-questions; based on the meta-knowledge question-answer data, domain knowledge training is carried out on the to-be-trained question-answer model; on the basis of the meta-knowledge question and answer data, a semantic-related question set is constructed for the meta-questions, and structured reasoning training data is generated through a second model on the basis of the question set of the meta-questions; and based on the structured reasoning training data, performing reasoning ability enhancement training on the question and answer model to be trained after domain knowledge training. The question and answer performance of the model in the vertical field can be remarkably improved.

Description

Technical Field

[0001] The present disclosure relates to technologies such as large models, model training, deep learning, and other technologies in the field of artificial intelligence, as well as technical fields such as intelligent search and intelligent question-answering, and in particular to a question-answering model training method, an intelligent question-answering method, and an apparatus for a specific field. Background Art

[0002] With the continuous development of artificial intelligence, intelligent question-answering (Q&A) has been widely applied in various fields, such as virtual assistants and voice assistants. Intelligent Q&A uses large models (such as question-answering models) to perform semantic analysis on input questions. It then uses semantic retrieval or dialogue management within a large-scale knowledge base to match relevant questions. Finally, it uses natural language generation (NLG) to generate and respond to answers.

[0003] However, although current large models perform well in general question-answering tasks, they still face problems of knowledge blind spots and insufficient reasoning ability in vertical fields (such as internal corporate fields). Summary of the Invention

[0004] The present disclosure provides a question-answering model training method, an intelligent question-answering method, and a device for a specific field.

[0005] In a first aspect, the present disclosure provides a method for training a question-answering model for a specific field, including:

[0006] Acquiring corpus data in the specific field;

[0007] Constructing meta-knowledge question-answering data based on the corpus data using a first model, wherein the meta-knowledge question-answering data includes meta-questions and answers corresponding to the meta-questions;

[0008] Based on the meta-knowledge question-answering data, performing domain knowledge training on the question-answering model to be trained;

[0009] Based on the meta-knowledge question-answering data, construct a semantically related question set for the meta-question, and generate structured reasoning training data based on the question set of the meta-question through a second model;

[0010] Based on the structured reasoning training data, the question-answering model to be trained after the domain knowledge training is trained to enhance reasoning ability.

[0011] In a second aspect, the embodiments of the present disclosure provide an intelligent question-answering method for a specific field, including:

[0012] Get the input pending issues;

[0013] The question to be processed is input into a pre-trained question-answering model to obtain the answer output by the question-answering model; wherein, the question-answering model is trained based on the method described in the first aspect above.

[0014] In a third aspect, the present disclosure provides a field-specific question-answering model training device, including:

[0015] An acquisition module, used to acquire corpus data in the specific field;

[0016] A first construction module is configured to construct meta-knowledge question-answering data based on the corpus data using a first model, wherein the meta-knowledge question-answering data includes meta-questions and answers corresponding to the meta-questions;

[0017] A first training module is used to perform domain knowledge training on the question-answering model to be trained based on the meta-knowledge question-answering data;

[0018] A second building module is configured to build a semantically related question set for the meta-question based on the meta-knowledge question-answering data;

[0019] a generating module, configured to generate structured reasoning training data based on the question set of the meta-problem by using a second model;

[0020] The second training module is used to perform reasoning ability enhancement training on the question-answering model to be trained after the domain knowledge training based on the structured reasoning training data.

[0021] In a fourth aspect, an embodiment of the present disclosure provides an intelligent question-answering device for a specific field, including:

[0022] The first acquisition module is used to obtain input questions to be processed;

[0023] The second acquisition module is used to input the question to be processed into a pre-trained question-answering model to obtain the answer output by the question-answering model; wherein, the question-answering model is trained based on the method described in the first aspect above.

[0024] In a fifth aspect, an embodiment of the present disclosure provides an intelligent agent, including:

[0025] An input module, used for receiving input information;

[0026] a processing module, configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, and execute the methods described in the first and second aspects above by calling the large model to obtain output information;

[0027] An output module is used to output the output information obtained by the processing module.

[0028] In a sixth aspect, an embodiment of the present disclosure provides an electronic device, including:

[0029] at least one processor; and

[0030] a memory communicatively connected to the at least one processor; wherein,

[0031] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the methods described in the first and second aspects above.

[0032] In a seventh aspect, an embodiment of the present disclosure provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute the methods described in the first and second aspects above.

[0033] In an eighth aspect, an embodiment of the present disclosure provides a computer program product, comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of the method according to the first and second aspects above.

[0034] According to the technical solution disclosed in the present invention, not only can the general knowledge accumulated in the model pre-training be retained, but also the new knowledge or detailed information in the specific field corpus can be ensured to be absorbed. The model's domain knowledge mastery and complex reasoning capabilities can be enhanced at a lower cost, the problem of "hallucination" of external knowledge can be reduced, and structured reasoning can be supported, which can be applied to actual complex problems. It has low external dependence and can be applied to data-sensitive fields. It can significantly improve the model's question-answering performance in vertical fields, and can solve the problem that traditional continuous pre-training technology requires extremely high computing resources and has poor cross-domain adaptability.

[0035] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings are used to better understand the present invention and do not constitute a limitation of the present invention.

[0037] Figure 1 A flowchart of a method for training a question-answering model for a specific field provided in an embodiment of the present disclosure;

[0038] Figure 2 A flowchart of a method for training a question-answering model for a specific field provided in an embodiment of the present disclosure;

[0039] Figure 3 An example diagram of the effect of the sliding window provided in an embodiment of the present disclosure;

[0040] Figure 4 A flowchart of a method for training a question-answering model for a specific field provided in an embodiment of the present disclosure;

[0041] Figure 5 A flowchart of an intelligent question-answering method for a specific field provided by an embodiment of the present disclosure;

[0042] Figure 6 A block diagram of a question-answering model training device for a specific field provided by an embodiment of the present disclosure;

[0043] Figure 7 A block diagram of an intelligent question-answering device for a specific field provided by an embodiment of the present disclosure;

[0044] Figure 8 is a block diagram of an electronic device according to an embodiment of the present disclosure;

[0045] Figure 9 A block diagram of an intelligent agent provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0046] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0047] The embodiments of the present disclosure relate to technologies such as natural language processing, large language models, model training, deep learning, and technical fields such as intelligent search and intelligent question answering.

[0048] Artificial Intelligence (AI) is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence.

[0049] Natural Language Processing (NLP) is a key area of ​​research in computer science and artificial intelligence. It studies the theories and methods that enable effective communication between humans and computers using natural language. It is a discipline that uses computer technology to analyze, understand, and process natural language. This discipline uses computers as powerful tools for language research, quantitatively analyzing linguistic information with computer support and providing linguistic descriptions that can be used jointly by humans and computers.

[0050] Large language models (LLMs, also known as large models) are deep learning models trained using large amounts of text data. They can generate natural language text or understand the meaning of text. Large language models can handle a variety of natural language tasks, such as text classification, question-answering, and conversation, and are an important path to artificial intelligence.

[0051] An agent is an agent that can perceive its environment and take actions to achieve specific goals. It can be software, hardware, or a system, and possesses autonomy, adaptability, and interaction. An agent perceives changes in its environment (e.g., through sensors or data input), makes judgments and decisions based on learned knowledge and algorithms, and then executes actions to influence the environment or achieve a predetermined goal.

[0052] It should be noted that the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in the technical solution of this disclosure are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0053] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in this disclosure are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0054] It is worth noting that in the embodiments of the present disclosure, certain software, components, models, etc. that already exist in the industry may be mentioned. They should be considered as exemplary and their purpose is only to illustrate the feasibility of implementing the technical solution of the present disclosure, but it does not mean that the applicant has or will necessarily use the solution.

[0055] To improve the question-answering performance of models in specific domains, two main techniques are employed: 1) Retrieval-augmented generation (RAG), which provides auxiliary information by retrieving external documents; and 2) continuous pre-training, which retrains the model using massive amounts of domain data. However, RAG suffers from redundant retrieval, response delays, and "hallucination" issues caused by conflicting information. Furthermore, continuous pre-training requires extremely high computational resources and has poor cross-domain adaptability.

[0056] Based on this, the present disclosure aims to solve the problems of performance degradation, poor generalization ability, high training cost, etc. of large language models when handling question-answering tasks in specific fields. Currently, although large language models perform well in general question-answering tasks, they still face problems of knowledge blind spots and insufficient reasoning ability in the internal fields of enterprises. To this end, the embodiments of the present disclosure provide a question-answering model training method, intelligent question-answering method and device for specific fields, which can enhance the model's domain knowledge mastery ability and complex reasoning ability at a relatively low cost, and significantly improve the model's question-answering performance in vertical fields.

[0057] Exemplary products or projects to which the present disclosure is applied include but are not limited to the following 1) to 4):

[0058] 1) Enterprise Internal Knowledge Base Question and Answer System: Within large organizations, knowledge management faces data silos and difficulty accessing data across systems. This disclosure can automatically generate domain-specific QA (Question Answering) data from internal enterprise documents (such as manuals, technical documentation, emails, and announcements), and perform efficient training and reasoning enhancement.

[0059] 2) Financial Policy Q&A: In the financial industry, policies, products, and regulatory knowledge are frequently updated, requiring both ordinary users and financial practitioners to understand and respond to them in real time. This disclosure can transform documents such as financial regulatory announcements, product manuals, and financial reports into Q&A pairs, and design reasoning tasks related to finance and macroeconomic policies to train models with domain explanatory power and reasoning judgment.

[0060] 3) Legal Document Analysis: In the legal services sector, legal texts are complex, and cases rely on logical reasoning and standardized retrieval. When processing legal scenarios, this disclosure can extract structured questions and answers and reasoning links from the Judgment Documents Network and legal databases to train a large legal model capable of citing legal provisions and comparing precedents.

[0061] 4) Medical Assisted Consultation System: In healthcare scenarios, doctors and patients often face highly specialized and complex symptom consultation and case analysis needs. This disclosure integrates structured medical knowledge with the internal pre-training capabilities of large models to extract high-quality question-answer pairs from a large number of electronic medical records, medical textbooks, and guidelines, thereby training a medical assistant model with reasoning capabilities.

[0062] The following describes the field-specific question-answering model training method, intelligent question-answering method, and device of the embodiments of the present disclosure with reference to the accompanying drawings.

[0063] Figure 1Flowchart of a method for training a question-answering model for a specific domain provided by an embodiment of the present disclosure. It should be noted that the execution entity of the method for training a question-answering model for a specific domain in an embodiment of the present disclosure may be a device for training a question-answering model for a specific domain, which may be implemented by software and / or hardware and may be configured in an electronic device, which may include but is not limited to a terminal, a server, etc.

[0064] It is worth noting that the question-answering model training method for a specific field in the embodiment of the present disclosure can be implemented by an autonomous intelligent agent (AutoAgent) based on a large language model, and the autonomous intelligent agent can implement intelligent question-answering for a specific field based on the large model.

[0065] like Figure 1 As shown, the field-specific question-answering model training method may include but is not limited to the following steps.

[0066] In step 101, corpus data of a specific field is obtained.

[0067] In the embodiments of the present disclosure, the corpus data in the specific field can be understood as the corpus in the vertical field. For example, the corpus data in the specific field may include but is not limited to medical literature in the medical field, or the corpus data in the specific field may also include but is not limited to legal precedents, or the corpus data in the specific field may also include but is not limited to financial announcements, but is not limited to this.

[0068] In step 102, meta-knowledge question and answer data is constructed based on the corpus data using a first model. The meta-knowledge question and answer data includes meta-questions and answers corresponding to the meta-questions.

[0069] In some embodiments, the first model can be a large language model (LLM). In an embodiment of the present disclosure, in order to avoid interference from redundant information, the corpus data of a specific field can be cleaned and segmented first, and natural language processing can be performed based on the corpus data through the first model to generate meta-questions covering basic concepts, definitions, mechanisms, etc. The generated meta-questions focus on semantically intensive knowledge points in the corpus data. Based on the corpus data and the generated meta-questions, meta-knowledge question-and-answer data is constructed. For example, after generating the meta-questions, the answers corresponding to the meta-questions can be found from the corpus data, and the meta-knowledge question-and-answer data is constructed based on the meta-questions and their answers.

[0070] In step 103, domain knowledge training is performed on the question-answering model to be trained based on the meta-knowledge question-answering data.

[0071] In the embodiments of the present disclosure, the question-answering model to be trained is different from the first model described above. Exemplarily, the first model can be a large language model with question generation capabilities, which can be a trained model or a model that does not require training. The question-answering model to be trained can be a large model used for training. Optionally, the number of parameters in the question-answering model to be trained is smaller than that of the first model.

[0072] In the embodiment of the present disclosure, when the question-answering model to be trained is trained on domain knowledge, the question-answering model to be trained can be allowed to generate answers under two conditions: the model's built-in knowledge base and the corpus data of the specific domain (relative to the question-answering model to be trained, which can be understood as external knowledge). Based on the answers generated under these two conditions, a fused answer is constructed by self-selection. A loss value is determined based on the fused answer and the answer corresponding to the meta-question in the meta-knowledge question-answering data, and the question-answering model to be trained is trained based on the loss value. In this way, not only can the general knowledge accumulated in the model pre-training (such as term definitions, general judgments, etc.) be retained, but also new knowledge or detailed information in the corpus of the specific domain can be absorbed.

[0073] In step 104, a semantically related question set is constructed for the meta-question based on the meta-knowledge question-answering data, and structured reasoning training data is generated based on the question set of the meta-question through a second model.

[0074] In some embodiments, related questions can be retrieved based on semantic similarity based on meta-questions in the meta-knowledge Q&A data. For example, a semantically related set of questions can be constructed for each meta-question from the meta-knowledge Q&A data using semantic similarity retrieval. This allows for precise matching based on semantic similarity, ensuring logical consistency among the combined questions.

[0075] In an embodiment of the present disclosure, after constructing a set of semantically related questions for a meta-question, combined reasoning data across knowledge points can be constructed based on the question set through a second model, that is, structured reasoning training data can be obtained. The structured reasoning training data can be regarded as a complex question sample and used to train the question-answering model to be trained to enhance the reasoning ability. Exemplarily, the second model can be a large language model, and the second model can be a large language model with the function of generating answers based on questions. Optionally, the number of parameters of the question-answering model to be trained is less than the number of parameters of the second model. Optionally, the second model and the above-mentioned first model can be the same model, or they can be different models.

[0076] In step 105, the question-answering model to be trained, which has been trained with domain knowledge, is trained to enhance its reasoning capability based on the structured reasoning training data.

[0077] In the embodiments of the present disclosure, on the basis of domain knowledge training of the question-answering model to be trained so that the model can master the basic knowledge of the domain, the question-answering model to be trained after domain knowledge training can be trained to enhance reasoning ability based on structured reasoning training data, which can further improve the model's multi-hop reasoning ability in actual complex tasks.

[0078] In the above embodiment, meta-knowledge question-answering data is constructed based on a specific domain corpus by a first model, and domain knowledge training is performed on the question-answering model to be trained based on the meta-knowledge question-answering data. This not only retains the general knowledge accumulated in the model pre-training (such as term definitions, general judgments, etc.), but also ensures that new knowledge or detailed information in the specific domain corpus can be absorbed. Structured reasoning training data is generated based on a set of meta-questions in the meta-knowledge question-answering data by a second model, and reasoning ability enhancement training is performed on the question-answering model to be trained based on the generated structured reasoning training data. The model's domain knowledge mastery ability and complex reasoning ability are enhanced at a lower cost, which can reduce the problem of "hallucination" of external knowledge, support structured reasoning that can be applied to actual complex problems, and have low external dependence and can be applied to data-sensitive fields. It can significantly improve the model's question-answering performance in vertical fields, and can solve the problem that traditional continuous pre-training technology requires extremely high computing resources and has poor cross-domain adaptability.

[0079] In some embodiments, based on any of the above embodiments, Figure 2 As shown, the optional implementation methods of the above-mentioned step of performing domain knowledge training on the question-answering model to be trained based on the meta-knowledge question-answering data may include but are not limited to the following steps.

[0080] In step 201, a meta-question is input into the question-answering model to be trained.

[0081] In an embodiment of the present disclosure, the meta-question generated by the first model can be input into the question-answering model to be trained for text similarity retrieval, so that the question-answering model to be trained generates an answer.

[0082] In step 202, the first candidate answer is obtained by performing a text similarity search based on the meta-question in the built-in knowledge base of the question-answering model to be trained.

[0083] In the embodiments of the present disclosure, the model-built-in knowledge base can be understood as the built-in knowledge base of the question-answering model to be trained. This model-built-in knowledge base can be used to maintain general knowledge, such as term definitions, general judgments, etc. A meta-question can be input into the question-answering model to be trained, so that the question-answering model to be trained generates an answer without context. In other words, the question-answering model to be trained performs a text similarity search in the model-built-in knowledge base based on the meta-question to obtain the first candidate answer to the meta-question.

[0084] In step 203, a second candidate answer is obtained by performing text similarity retrieval on the corpus data based on the meta-question by the question-answering model to be trained.

[0085] In the embodiments of the present disclosure, the corpus data is recorded as domain-specific corpus data. A meta-question can be input into the to-be-trained question model, so that the to-be-trained question model generates an answer under the condition of providing the original corpus context. In other words, the to-be-trained question model performs a text similarity search based on the meta-question in the domain-specific corpus data to obtain a second candidate answer to the meta-question.

[0086] It should be noted that, in some embodiments, the above steps 202 and 203 may be performed in an interchangeable order or simultaneously.

[0087] In step 204, the first candidate answer and the second candidate answer are fused to obtain a fused answer.

[0088] In some embodiments, based on the Token (word) granularity, combined with the generation confidence of the Token at the same position in the first candidate answer and the second candidate answer, the words that meet the fusion conditions can be fused to obtain a fused answer. In the embodiments of the present disclosure, the Token (word) can be understood as the smallest unit in text processing, such as a word, subword or character; the generation confidence of the Token can refer to the quantitative value of the degree of certainty of the model's prediction result when the model generates or predicts the Token, usually expressed as a probability value, that is, it can be understood as the probability of the question-answering model to be trained generating the Token, reflecting the probability value (usually a value between 0 and 1) that the model believes that the Token is the next most reasonable choice based on the existing context.

[0089] Optionally, the above-mentioned fusion conditions can be used to select the best knowledge source from the first candidate answer and the second candidate answer for fusion. Exemplarily, the generation confidence of the tokens at the same position in the first candidate answer and the second candidate answer can be compared at the token granularity to select the best knowledge source for fusion. Exemplarily, for the tokens at the same position in the first candidate answer and the second candidate answer, the token with a larger generation confidence can be selected as a component of the fusion answer. For example, taking the generation confidence of the second token in the first candidate answer as 0.45 and the generation confidence of the token at the same position in the second candidate answer as 0.9 as an example, the generation confidence of the second token in the second candidate answer is larger, then the second token in the second candidate answer can be selected as a component of the fusion answer, and the second token in the first candidate answer can be discarded.

[0090] Optionally, in some embodiments, an adjustable threshold parameter (such as a first threshold) can be introduced to control the degree of trust in external information during fusion. In one possible implementation, the generation confidence of the first Token in the second candidate answer can be compared with the first threshold based on the Token granularity; when the generation confidence of the first Token is greater than or equal to the first threshold, the first Token is selected as a component of the fused answer; when the generation confidence of the first Token is less than the first threshold, the second Token at the same position as the first Token in the first candidate answer is selected, and the second Token is used as a component of the fused answer. Optionally, the first threshold can be a hyperparameter, which can be determined based on actual application conditions. For example, the first threshold can be 0.98, but is not limited to this.

[0091] For example, for the tokens at the same position in the first candidate answer and the second candidate answer, the generation confidence of the token in the second candidate answer can be compared with the first threshold. If the generation confidence of the token in the second candidate answer is greater than or equal to the first threshold, the token in the second candidate answer can be fused as the optimal knowledge source; if the generation confidence of the token in the second candidate answer is less than the first threshold, the token at the same position in the first candidate answer can be fused as the optimal knowledge source.

[0092] For example, if the generation confidence of the second token in the first candidate answer is 0.98 and the generation confidence of the token at the same position in the second candidate answer is 0.99, assuming the first threshold is 0.98 and the generation confidence of the second token in the second candidate answer is greater than the first threshold, the second token in the second candidate answer can be selected as a component of the fused answer, and the second token in the first candidate answer can be discarded. Suppose the generation confidence of the third token in the first candidate answer is 0.98 and the generation confidence of the token at the same position in the second candidate answer is 0.9, since the generation confidence of the token at the same position in the second candidate answer is less than the first threshold, the token in the first candidate answer can be used as the optimal knowledge source for fusion, and so on, until all tokens at the same position in the first and second candidate answers are fused and selected, and the final fused answer can be obtained.

[0093] In order to avoid over-reliance on individual high-confidence words in Token selection and loss of overall semantics, optionally, in some embodiments, a sliding window method can be used to determine whether the fusion conditions are met within a local paragraph. In one possible implementation, a sliding window method can be used to dynamically capture the generation confidence of the same local paragraph in the first candidate answer and the second candidate answer; based on the generation confidence of the local paragraph, the local paragraphs that meet the fusion conditions are fused. The generation confidence of the local paragraph is determined based on the generation confidence of the tokens in the local paragraph. For example, the generation confidence of all tokens in the local paragraph can be multiplied, and the resulting product value can be used as the generation confidence of the local paragraph.

[0094] For example, take the sliding window length as 3 (ie 3 Tokens), as Figure 3 As shown, T 1i Represents the i-th Token in the first candidate answer, T 2i represents the i-th Token in the second candidate answer, and n represents the total number of Tokens. A sliding window of length 3 can be used to perform a window sliding operation in the first candidate answer and the second candidate answer respectively, so that the first partial paragraph obtained by the window sliding operation in the first candidate answer includes T 1(i-1) 、T 1i 、T 1(i+1) , the second partial paragraph obtained by sliding the window in the second candidate answer includes T 2(i-1) 、T 2i 、T 2(i+1) For example, the generation confidence of the first partial paragraph = T 1(i-1) The generated confidence ×T 1i The generated confidence ×T 1(i+1) The generation confidence of the second local paragraph = T 2(i-1) The generated confidence ×T 2i The generated confidence ×T 2(i+1) The generation confidence of the first partial paragraph and the second partial paragraph can be used as a component of the fusion answer. Alternatively, the generation confidence of the second partial paragraph can be compared with a second threshold. If the generation confidence of the second partial paragraph is greater than or equal to the second threshold, the second partial paragraph is selected and used as a component of the fusion answer; if the generation confidence of the second partial paragraph is less than the second threshold, the first partial paragraph is selected and used as a component of the fusion answer.

[0095] In step 205, a loss value is generated based on the answer corresponding to the meta-question and the fused answer, and the question-answering model to be trained is trained according to the loss value.

[0096] In an embodiment of the present disclosure, a loss value can be calculated using a preset loss function based on the answer to the meta-question and the fused answer, and the question-answering model to be trained can be trained based on the loss value. Exemplarily, the loss function can be a cross-entropy loss function, but is not limited thereto, and can be, for example, a mean square error (MSE) or the like.

[0097] Optionally, in order to prioritize the optimization of uncertain or incorrectly predicted knowledge points, a "Selective Supervision Fine-tuning (Selective SFT)" mechanism can be introduced. In some embodiments, the Loss weight of the corresponding Token can be dynamically adjusted based on the prediction uncertainty of each Token in the fused answer by the question-answering model to be trained. The prediction uncertainty can be understood as an entropy value, for example, it can be calculated based on the Token in the answer corresponding to the meta-question and the fused answer. Exemplarily, the greater the prediction uncertainty of the Token, the greater the adjustment range of the Loss weight of the Token, and the smaller the prediction uncertainty of the Token, the smaller the adjustment range of the Loss weight of the Token. Alternatively, the Loss weight of the Token can be adjusted based on the mapping relationship between prediction uncertainty and loss weight. Based on the adjusted loss weight, the loss value between the answer corresponding to the meta-question and the fused answer is calculated, and the question-answering model to be trained is trained according to the loss value. In this way, uncertain or incorrectly predicted knowledge points can be prioritized for optimization, thereby improving training efficiency, avoiding wasting resources on content that the model has already mastered, and significantly improving knowledge transfer capabilities.

[0098] Figure 4 This is a flowchart of the method for training a question-answering model for a specific field provided by an embodiment of the present disclosure. Figure 4 As shown, the field-specific question-answering model training method may include but is not limited to the following steps.

[0099] In step 401, corpus data of a specific domain is obtained.

[0100] Optionally, step 401 may be implemented in any of the implementations in the embodiments of the present disclosure, which is not limited in the embodiments of the present disclosure and will not be described in detail.

[0101] In step 402, meta-knowledge question and answer data is constructed based on the corpus data using a first model. The meta-knowledge question and answer data includes meta-questions and answers corresponding to the meta-questions.

[0102] Optionally, step 402 may be implemented using any implementation method in each embodiment of the present disclosure, which is not limited in the embodiments of the present disclosure and will not be described in detail.

[0103] In step 403, domain knowledge training is performed on the question-answering model to be trained based on the meta-knowledge question-answering data.

[0104] Optionally, step 403 may be implemented in any of the implementation methods in the embodiments of the present disclosure, which is not limited in the embodiments of the present disclosure and will not be described in detail.

[0105] In step 404, a semantically related question set is constructed for the meta-question based on the meta-knowledge question-answering data, and structured reasoning training data is generated based on the question set of the meta-question through a second model.

[0106] In some embodiments, related questions can be retrieved based on semantic similarity based on meta-questions in the meta-knowledge Q&A data. For example, a semantically related set of questions can be constructed for each meta-question from the meta-knowledge Q&A data using semantic similarity retrieval. This allows for precise matching based on semantic similarity, ensuring logical consistency among the combined questions.

[0107] In some embodiments, a corresponding specific prompt word template can be determined according to the reasoning type; and structured reasoning training data matching the reasoning type is generated by a second model based on the specific prompt word template corresponding to the reasoning type and a question set of the meta-question.

[0108] Exemplarily, a structured reasoning data generation process can be introduced, covering multiple categories (such as three categories) of typical reasoning paradigms: deductive reasoning (from rules to conclusions), inductive reasoning (from instances to laws), and case analogy (from historical cases to new situations). For each reasoning type, a specific Prompt template corresponding to the reasoning type can be determined. The specific Prompt template is used to allow the large model to combine multiple knowledge points to generate structured reasoning data. In other words, through the second model based on the specific Prompt template, complex question and answer samples with clear reasoning chains and complete logic can be generated, thereby constructing structured reasoning training data with wide coverage. The structured reasoning training data can be used to enhance the reasoning ability of the question and answer model to be trained.

[0109] In step 405, based on the structured reasoning training data, a reinforcement learning algorithm (such as Group Relative Policy Optimization, GRPO for short) of a value-free evaluation network (such as a Critic network) is used to perform reasoning capability enhancement training on the question-answering model to be trained.

[0110] In the embodiments of the present disclosure, during reasoning ability enhancement training, multiple generated results are sampled for each question in the structured reasoning training data. A reward function is then used to normalize the average and variance of these generated results to calculate a reward advantage value. This reward advantage value is then used to update the model parameters of the question-answering model to be trained. This reward function not only focuses on the accuracy of the final answer but also considers format compliance (e.g., whether it is legal JSON) to optimize the structural compliance and logical coherence of the generated content.

[0111] In the above embodiment, the second model generates structured reasoning training data matching the reasoning type based on specific prompt word templates and meta-questions corresponding to different reasoning types. Based on the structured reasoning training data, a reinforcement learning algorithm of a value-free evaluation network is used to enhance the reasoning ability of the question-answering model to be trained. This allows the model to effectively integrate knowledge chains and follow logical processes during the reasoning training process, thereby achieving robust processing of complex tasks.

[0112] Figure 5 Flowchart of a domain-specific intelligent question-answering method provided in an embodiment of the present disclosure. It should be noted that the execution entity of the domain-specific intelligent question-answering method in an embodiment of the present disclosure may be a domain-specific intelligent question-answering device, which may be implemented by software and / or hardware and may be configured in an electronic device, which may include but is not limited to a terminal, a server, etc.

[0113] It is worth noting that the intelligent question-answering method for a specific field in the embodiment of the present disclosure can be implemented by an autonomous agent (AutoAgent) based on a large language model, and the autonomous agent can implement intelligent question-answering for a specific field based on the large model.

[0114] like Figure 5 As shown, the intelligent question-answering method for a specific field may include but is not limited to the following steps.

[0115] In step 501, an input question to be processed is obtained.

[0116] Optionally, user input information may be obtained, and the problem to be processed may be determined based on the input information. Exemplarily, the input information may be in the form of text, and the input information may be used as the problem to be processed. Exemplarily, the input information may be in the form of voice or audio, and the voice or audio may be converted into text, and the converted text may be used as the problem to be processed. Exemplarily, the input information may be in the form of a video including sign language, and the sign language may be identified from the video and converted into corresponding text, and the converted text may be used as the problem to be processed.

[0117] In step 502, the question to be processed is input into the pre-trained question-answering model to obtain the answer output by the question-answering model.

[0118] Among them, in the embodiments of the present disclosure, the question-answering model can be trained based on the method shown in any of the aforementioned embodiments, which will not be repeated here.

[0119] In the above embodiment, intelligent question and answer in vertical fields can be achieved through question and answer models oriented to specific fields, which can improve the accuracy of intelligent question and answer results.

[0120] Figure 6 This is a block diagram of a question-answering model training device for a specific field provided by an embodiment of the present disclosure. Figure 6 As shown, the question-answering model training device for a specific field may include an acquisition module 601, a first construction module 602, a first training module 603, a second construction module 604, a generation module 605, and a second training module 606.

[0121] The acquisition module 601 is used to acquire corpus data in a specific field.

[0122] The first construction module 602 is used to construct meta-knowledge question and answer data based on the corpus data using a first model, where the meta-knowledge question and answer data includes meta-questions and answers corresponding to the meta-questions.

[0123] The first training module 603 is used to perform domain knowledge training on the question-answering model to be trained based on the meta-knowledge question-answering data.

[0124] The second construction module 604 is used to construct a semantically related question set for the meta-question based on the meta-knowledge question-answering data.

[0125] The generating module 605 is configured to generate structured reasoning training data based on the question set of the meta-question by using the second model.

[0126] The second training module 606 is used to perform reasoning capability enhancement training on the question-answering model to be trained after domain knowledge training based on structured reasoning training data.

[0127] In some embodiments, the first construction module 602 is used to: generate meta-questions based on corpus data through a first model, where the meta-questions focus on semantically intensive knowledge points in the corpus data; and construct meta-knowledge question-answering data based on the corpus data and the meta-questions.

[0128] In some embodiments, the first training module 603 is used to: input the meta-question into the question-answering model to be trained; obtain the first candidate answer obtained by the question-answering model to be trained based on the meta-question by performing text similarity retrieval in the model's built-in knowledge base; obtain the second candidate answer obtained by the question-answering model to be trained based on the meta-question by performing text similarity retrieval in the corpus data; fuse the first candidate answer and the second candidate answer to obtain a fused answer; generate a loss value based on the answer corresponding to the meta-question and the fused answer, and train the question-answering model to be trained according to the loss value.

[0129] In some embodiments, the first training module 603 is used to: based on word-gram granularity, combine the generation confidence of the word-grams at the same position in the first candidate answer and the second candidate answer, and fuse the words that meet the fusion conditions to obtain a fused answer.

[0130] In some embodiments, the first training module 603 is used to: compare the generation confidence of the first word in the second candidate answer with the first threshold based on word granularity; when the generation confidence of the first word is greater than or equal to the first threshold, select the first word as a component of the fused answer; when the generation confidence of the first word is less than the first threshold, select the second word at the same position as the first word in the first candidate answer, and use the second word as a component of the fused answer.

[0131] In some embodiments, the first training module 603 is used to: dynamically capture the generation confidence of the same local paragraph in the first candidate answer and the second candidate answer using a sliding window; the generation confidence of the local paragraph is determined based on the generation confidence of the word units in the local paragraph; based on the generation confidence of the local paragraph, the local paragraphs that meet the fusion conditions are fused.

[0132] In some embodiments, the first training module 603 is further used to dynamically adjust the loss weight of the corresponding word based on the prediction uncertainty of each word in the fused answer by the question-answering model to be trained.

[0133] In some embodiments, the generation module 605 is used to: determine the corresponding specific prompt word template according to the reasoning type; and generate structured reasoning training data matching the reasoning type based on the specific prompt word template corresponding to the reasoning type and the question set of the meta-question through the second model.

[0134] In some embodiments, the second training module 606 is used to: based on structured reasoning training data, use a reinforcement learning algorithm with a value-free evaluation network to perform reasoning ability enhancement training on the question-answering model to be trained; wherein, in the reasoning ability enhancement training, multiple generation results are sampled for each question in the structured reasoning training data, and a reward function is used to calculate a reward advantage value by normalizing the average and variance of the multiple generation results, and the reward advantage value is used to update the model parameters of the question-answering model to be trained.

[0135] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0136] Figure 7 This is a block diagram of an intelligent question-answering device for a specific field provided by an embodiment of the present disclosure. Figure 7 As shown, the intelligent question-answering device for a specific field may include: a first acquisition module 701 and a second acquisition module 702.

[0137] The first acquisition module 701 is used to acquire input questions to be processed.

[0138] The second acquisition module 702 is used to input the question to be processed into a pre-trained question-answering model to obtain the answer output by the question-answering model; wherein the question-answering model is trained based on the question-answering model training method for a specific field shown in any of the aforementioned embodiments.

[0139] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0140] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device and a readable storage medium.

[0141] like Figure 8 , is a block diagram of an electronic device according to an embodiment of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0142] like Figure 8As shown, the electronic device includes: one or more processors 801, a memory 802, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. The various components are connected to each other using different buses and can be installed on a common mainboard or installed in other ways as needed. The processor can process instructions executed in the electronic device, including instructions stored in or on the memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In other embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple electronic devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 8 A processor 801 is taken as an example.

[0143] Memory 802 is a non-transitory computer-readable storage medium provided by the present disclosure. The memory stores instructions executable by at least one processor to cause the at least one processor to execute the field-specific question-answering model training method or field-specific intelligent question-answering method provided by the present disclosure. The non-transitory computer-readable storage medium of the present disclosure stores computer instructions for causing a computer to execute the field-specific question-answering model training method or field-specific intelligent question-answering method provided by the present disclosure.

[0144] The memory 802 is a non-transient computer-readable storage medium that can be used to store non-transient software programs, non-transient computer executable programs and modules, such as the program instructions / modules corresponding to the question-answering model training method for a specific field or the intelligent question-answering method for a specific field in the embodiments of the present disclosure (for example, the attached Figure 6 The acquisition module 601, the first construction module 602, the first training module 603, the second construction module 604, the generation module 605, the second training module 606, or the attached Figure 7 The processor 801 executes the non-transient software programs, instructions, and modules stored in the memory 802 to execute various functional applications and data processing of the server, that is, to implement the field-specific question-answering model training method or the field-specific intelligent question-answering method in the above method embodiment.

[0145] The memory 802 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device, etc. In addition, the memory 802 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory 802 may optionally include a memory remotely located relative to the processor 801, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0146] The electronic device may further include: an input device 803 and an output device 804. The processor 801, the memory 802, the input device 803 and the output device 804 may be connected via a bus or other means. Figure 8 The bus connection is taken as an example.

[0147] The input device 803 can receive input digital or character information and generate key signal input related to user settings and function control of the electronic device, such as input devices such as a touch screen, a keypad, a mouse, a trackpad, a touch pad, an indicator stick, one or more mouse buttons, a trackball, and a joystick. The output device 804 may include a display device, an auxiliary lighting device (e.g., an LED), and a tactile feedback device (e.g., a vibration motor). The display device may include, but is not limited to, a liquid crystal display (LCD), a light emitting diode (LED) display, and a plasma display. In some embodiments, the display device may be a touch screen.

[0148] Various implementations of the systems and techniques described herein can be realized in digital electronic circuit systems, integrated circuit systems, dedicated ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0149] These computer programs (also referred to as programs, software, software applications, or code) include machine instructions for a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., a magnetic disk, an optical disk, a memory, a programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0150] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0151] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), the Internet, and a blockchain network.

[0152] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact via a communication network. This client-server relationship is established by computer programs running on the respective computers and establishing a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, a host product within a cloud computing service ecosystem that addresses the management difficulties and limited business scalability of traditional physical hosts and VPS services ("Virtual Private Servers" or simply "VPS"). The server may also be a server in a distributed system or a server integrated with blockchain.

[0153] Figure 9 A block diagram of an intelligent agent provided in an embodiment of the present disclosure. Figure 9 As shown, the intelligent agent may include an input module 901, a processing module 902, and an output module 903. The input module 901 is configured to receive input information; the processing module 902 is configured to determine a target task based on the input information received by the input module 901, determine a large model based on the target task, and execute the domain-specific question-answering model training method or domain-specific intelligent question-answering method in the above-mentioned method embodiment by calling the large model to obtain output information; and the output module 903 is configured to output the output information obtained by the processing module 902.

[0154] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0155] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A method for training a question-answering model for a specific domain, comprising: Acquiring corpus data in the specific field; Constructing meta-knowledge question-answering data based on the corpus data using a first model, wherein the meta-knowledge question-answering data includes meta-questions and answers corresponding to the meta-questions; Based on the meta-knowledge question-answering data, performing domain knowledge training on the question-answering model to be trained; Based on the meta-knowledge question-answering data, construct a semantically related question set for the meta-question, and generate structured reasoning training data based on the question set of the meta-question through a second model; Based on the structured reasoning training data, the question-answering model to be trained after the domain knowledge training is trained to enhance reasoning ability.

2. The method according to claim 1, wherein The step of constructing meta-knowledge question-answering data based on the corpus data by using the first model includes: Generate a meta-question based on the corpus data using the first model, where the meta-question focuses on semantically intensive knowledge points in the corpus data; The meta-knowledge question-answering data is constructed based on the corpus data and the meta-question.

3. The method according to claim 1, wherein The step of performing domain knowledge training on the question-answering model to be trained based on the meta-knowledge question-answering data includes: Inputting the meta-question into the question-answering model to be trained; Obtaining a first candidate answer obtained by the question-answering model to be trained based on the meta-question by performing text similarity retrieval in the model's built-in knowledge base; Obtain a second candidate answer obtained by the question-answering model to be trained through text similarity retrieval in the corpus data based on the meta-question; Fusing the first candidate answer and the second candidate answer to obtain a fused answer; Based on the answer corresponding to the meta-question and the fused answer, a loss value is generated, and the question-answering model to be trained is trained according to the loss value.

4. The method according to claim 3, wherein: The fusing the first candidate answer and the second candidate answer to obtain a fused answer includes: Based on the word-gram granularity and in combination with the generation confidence of the word-grams at the same position in the first candidate answer and the second candidate answer, the words that meet the fusion conditions are fused to obtain a fused answer.

5. The method according to claim 4, wherein: The step of fusing words that meet the fusion condition based on word-unit granularity and combining the generation confidence of word-units at the same position in the first candidate answer and the second candidate answer includes: Based on the word-unit granularity, comparing the generation confidence of the first word-unit in the second candidate answer with a first threshold; When the generation confidence of the first word-gram is greater than or equal to the first threshold, selecting the first word-gram as a component of the fused answer; When the generation confidence of the first word-gram is less than the first threshold, a second word-gram at the same position as the first word-gram in the first candidate answer is selected, and the second word-gram is used as a component of the fused answer.

6. The method of claim 4, wherein: The step of fusing words that meet the fusion condition based on word-unit granularity and combining the generation confidence of word-units at the same position in the first candidate answer and the second candidate answer includes: Dynamically capturing the generation confidence of the same local paragraph in the first candidate answer and the second candidate answer using a sliding window; the generation confidence of the local paragraph is determined based on the generation confidence of the word unit in the local paragraph; Based on the generation confidence of the local paragraphs, the local paragraphs that meet the fusion conditions are fused.

7. The method according to any one of claims 3 to 6, further comprising: According to the prediction uncertainty of each word in the fused answer by the question-answering model to be trained, the loss weight of the corresponding word is dynamically adjusted.

8. The method of claim 1, wherein: Generating structured reasoning training data based on the meta-problem question set by the second model includes: According to the reasoning type, determine the corresponding specific prompt word template; The second model generates structured reasoning training data matching the reasoning type based on the specific prompt word template corresponding to the reasoning type and the question set of the meta-question.

9. The method of claim 1, wherein: The reasoning ability enhancement training of the question-answering model to be trained after the domain knowledge training based on the structured reasoning training data includes: Based on the structured reasoning training data, a reinforcement learning algorithm of a value-free evaluation network is used to perform reasoning ability enhancement training on the question-answering model to be trained; wherein, in the reasoning ability enhancement training, multiple generation results are sampled for each question in the structured reasoning training data, and a reward function is used to calculate a reward advantage value by normalizing the average and variance of the multiple generation results, and the reward advantage value is used to update the model parameters of the question-answering model to be trained.

10. An intelligent question-answering method for a specific domain, comprising: Get the input pending issues; The question to be processed is input into a pre-trained question-answering model to obtain an answer output by the question-answering model; wherein the question-answering model is trained based on the method according to any one of claims 1 to 9.

11. A question-answering model training device for a specific field, comprising: An acquisition module, used to acquire corpus data in the specific field; A first construction module is configured to construct meta-knowledge question-answering data based on the corpus data using a first model, wherein the meta-knowledge question-answering data includes meta-questions and answers corresponding to the meta-questions; A first training module is used to perform domain knowledge training on the question-answering model to be trained based on the meta-knowledge question-answering data; A second building module is configured to build a semantically related question set for the meta-question based on the meta-knowledge question-answering data; a generating module, configured to generate structured reasoning training data based on the question set of the meta-problem by using a second model; The second training module is used to perform reasoning ability enhancement training on the question-answering model to be trained after the domain knowledge training based on the structured reasoning training data.

12. An intelligent question-answering device for a specific field, comprising: The first acquisition module is used to obtain input questions to be processed; The second acquisition module is used to input the question to be processed into a pre-trained question-answering model to obtain the answer output by the question-answering model; wherein, the question-answering model is trained based on the method as described in any one of claims 1 to 9.

13. An intelligent agent comprising: An input module, used for receiving input information; a processing module, configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, and execute the method according to any one of claims 1 to 9 and 10 by calling the large model to obtain output information; An output module is used to output the output information obtained by the processing module.

14. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 9 and 10.

15. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to enable the computer to execute the method according to any one of claims 1 to 9 and 10.

16. A computer program product comprising a computer program, wherein When the computer program is executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 9 and 10.

Images