A training method for an intelligent question-answering system based on machine learning

Abstract

This invention discloses a training method for an intelligent question-answering system based on machine learning. The method includes optimizing knowledge learning samples to obtain optimized knowledge, determining knowledge weights and constructing an industry knowledge base, fusing multimodal question content to obtain fused question content, obtaining industry description information and user profiles based on the fused question content and user information, inputting the fused question content, the industry description information, and the user profiles into a question-answering reinforcement learning model to obtain answer content, obtaining a first reward signal, a second reward signal, and dynamic reward weights based on the answer content and user question response information, adjusting the question-answering reinforcement learning model, inputting the multimodal question content into the adjusted question-answering reinforcement learning model to obtain optimized answer content, and updating the industry knowledge base using user interaction logs and the knowledge weights. This method can accurately and efficiently train intelligent question-answering systems and provides strong support for intelligent question-answering services in various industries.

Description

Technical Field

[0001] This invention relates to the field of intelligent question answering technology, and in particular to a training method for an intelligent question answering system based on machine learning. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent question-answering systems, as the core carrier of human-computer interaction, play a key role in fields such as customer service, smart education, and healthcare. They can automatically understand user questions and output accurate answers. Therefore, building a highly accurate and adaptable intelligent question-answering system is not only an important direction for the practical application of artificial intelligence technology, but also a key support for meeting the diverse needs of users.

[0003] However, traditional intelligent question-answering system training techniques suffer from significant bottlenecks: First, most solutions lack the ability to integrate multi-source heterogeneous knowledge and lack a state determination mechanism for knowledge weights, resulting in limited accuracy of industry knowledge bases. Second, their ability to integrate and process multi-modal question content in multi-modal interaction scenarios is weak, making it difficult to deeply explore users' real needs. Finally, their model optimization mechanisms are simplistic and lack dynamic adjustment strategies for reward signals, limiting the fine-grained training of models. To address these challenges, this invention proposes a machine learning-based intelligent question-answering system training method. This method acquires knowledge learning samples through dual knowledge channels, combines multi-modal fusion to generate industry description information and user profiles, and achieves model optimization based on multiple reward signals and dynamic reward weights through the synergy of a question-answering reinforcement learning model and a question-answering memory adjustment model. Furthermore, it utilizes user interaction logs to update the industry knowledge base, effectively overcoming the shortcomings of traditional techniques. This method not only effectively solves the problems of inefficient knowledge integration and coarse model training in traditional techniques but also significantly improves the industry adaptability, answer accuracy, and self-evolution capabilities of intelligent question-answering systems, holding significant practical importance for promoting the deep application of intelligent question-answering technology in vertical fields. Summary of the Invention

[0004] The purpose of this invention is to provide a training method for an intelligent question-answering system based on machine learning.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0006] This invention includes the following steps:

[0007] Obtain knowledge learning samples, optimize and classify the knowledge learning samples to obtain optimized knowledge, determine knowledge weights based on the sources of the optimized knowledge, and construct an industry knowledge base based on the optimized knowledge and the knowledge weights;

[0008] Obtain user information and multimodal question content, perform multimodal fusion on the multimodal question content to obtain fused question content, and obtain industry description information and user profile based on the fused question content and the user information;

[0009] The fused question content, the industry description information, and the user profile are input into the question-answering reinforcement learning model to obtain the answer content. The answer content is subjected to repetition detection to obtain a first reward signal. User question response information is obtained, and a second reward signal is determined based on the user question response information.

[0010] The answer content and user question response information are input into the question-and-answer memory adjustment model to obtain dynamic reward weights;

[0011] The question-answering reinforcement learning model is adjusted based on the first reward signal, the second reward signal, and the dynamic reward weight. The multimodal question content is input into the adjusted question-answering reinforcement learning model to obtain optimized answer content.

[0012] A user interaction log is generated based on the fused question content and the optimized answer content, and the industry knowledge base is updated using the user interaction log and the knowledge weight.

[0013] Furthermore, the method for obtaining optimization knowledge includes:

[0014] A dual knowledge channel is constructed to acquire knowledge learning samples; the dual knowledge channel includes a static knowledge channel and a dynamic knowledge channel.

[0015] The static knowledge channel injects and preprocesses industry ontology information; the industry ontology information includes an industry ontology library, industry standards, an industry structured knowledge base, and an industry language model.

[0016] The dynamic knowledge channel injects and preprocesses real-time industry data; the real-time industry data includes industry news streams, user interaction logs, and sensor data streams; the update condition for the dynamic knowledge channel is triggered when both the new knowledge confidence score and the graph connectivity are greater than the corresponding update threshold; the expression for calculating the new knowledge confidence score is:

[0017] ,

[0018] in for t Time information e confidence level Weighting based on the authority of the source, SA To ensure the authority of the information source, For decay rate, For the current time, This refers to the last time the information was updated.

[0019] A support vector machine classifier is used to classify knowledge learning samples according to industry, constructing basic knowledge graphs for different industries. Dynamic pruning is then applied to these basic knowledge graphs, with the following pruning conditions:

[0020] ,

[0021] ,

[0022] ,

[0023] in For nodes A function to determine whether the pruning conditions are met. For the current time, For nodes timestamp, T As a time threshold, For nodes Similarity to the corresponding cluster centers The similarity threshold for the current time. For nodes Importance rating The threshold for node importance. Based on the basic similarity threshold, The update frequency for the current time. K The update frequency for the current time. For nodes Degree centrality, For degree centrality weights, For nodes betweenness centrality, For betweenness centrality weights, For nodes PageRank value, As the PageRank value weight;

[0024] Based on the basic knowledge graph, knowledge distillation is performed on the corresponding knowledge learning samples to obtain knowledge distillation samples. Based on the knowledge distillation samples, the corresponding basic knowledge graph is incrementally updated. The knowledge distillation and incremental update steps are repeated until the graph stability index reaches the stability threshold. Optimized knowledge and optimized knowledge graph are then output.

[0025] The knowledge distillation is achieved through a teacher-student model, a target distillation loss function, and dynamic curriculum learning. The teacher-student model includes a teacher model and a student model. The teacher model outputs the entity relationship probability distribution based on a relational graph convolutional network of the basic knowledge graph. The student model outputs high-confidence prediction results based on a lightweight GraphSAGE model. The target distillation loss function is expressed as follows:

[0026] ,

[0027] ,

[0028] ,

[0029] in Let the target distillation loss function be... For the knowledge distillation loss function, , , These are the weighting coefficients. For Kullback-Leibler divergence, , Output probabilities for the teacher and student models. , The adjacency matrix for the teacher model and the student model. For Jensen-Shannon divergence, , Embed matrices for teacher and student model nodes. To balance the intensity of the adversarial training, Let the loss function be that of the discriminator. To distinguish the loss term output by the teacher model, To differentiate the loss term output by the student model, D For discriminator, As input knowledge learning samples, Output for the teacher model, Output for the student model. To maximize the student model loss.

[0030] Furthermore, the method for constructing an industry knowledge base includes the following steps:

[0031] The authority of optimization knowledge is determined based on its source, and its weight is determined based on the authority, timeliness, and consistency of the optimization knowledge. The expression is as follows:

[0032] ,

[0033] ,

[0034] in To optimize knowledge i Knowledge weight, To optimize the authority of knowledge, For decay rate, For the current time, This refers to the last time the information was updated. For the number of conflict relations, The total number of relations. To optimize knowledge i Source category weight, For the sake of the source's authority, To optimize knowledge i The number of citations; the above source categories include public datasets, expert-annotated data, user interaction data, and other data;

[0035] To optimize knowledge, timestamps and category labels are set. Knowledge vectors are generated from category labels, timestamps, and knowledge weights. Timestamps are set for the optimized knowledge graph. An industry knowledge base is constructed based on the optimized knowledge and the optimized knowledge graph. The industry knowledge base includes a graph database, a source database, and a vector database. The graph database stores optimized knowledge graphs from different industries. The source database stores optimized knowledge from different industries. The vector database stores different knowledge vectors.

[0036] Furthermore, the method for obtaining industry description information and user profiles includes:

[0037] Multimodal question content is input into a question fusion model to obtain fused question content. The question fusion model includes a feature processing layer, a spatiotemporal alignment strategy, a hierarchical fusion strategy, and an adversarial alignment enhancement strategy. The feature processing layer processes the multimodal question content through text units, audio units, and image units to obtain multimodal question features. The multimodal question features include text features, speech-text features, and image-text features. The spatiotemporal alignment strategy synchronizes text features with speech-text features in time and spatially registers text features with image-text features. The hierarchical fusion strategy fuses the multimodal question features to obtain fused question features. The adversarial alignment enhancement strategy improves the accuracy of modality fusion through a modality discriminator and generative adversarial training.

[0038] User information is divided into first user features and user interaction logs. The user interaction logs are hierarchically clustered to obtain first question features and first interaction needs. The merged question content is hierarchically clustered to obtain merged user features and merged industry features. The merged user features include second question features and second user features. The merged industry features include first user industry features and first question industry features.

[0039] The third user feature is obtained by cross-referencing the second user feature with the first user industry feature. The fourth user feature is obtained by fusing the third user feature with the first user feature. The second question feature is corrected using the first question feature to obtain the third question feature. The user profile is obtained by splicing the first interaction requirement, the third question feature and the fourth user feature.

[0040] The first question industry feature is modified by using the first question feature to obtain the second question industry feature. The intersection of the first user feature and the first user industry feature is obtained to obtain the second user industry feature. The second question industry feature and the second user industry feature are concatenated to obtain the industry description information.

[0041] Furthermore, the method for obtaining the answer content includes:

[0042] A question-answering reinforcement learning model is constructed, and the question content is input into the model to obtain the answer content; the question-answering reinforcement learning model includes a semantic layer, a syntactic layer, and a policy layer;

[0043] The semantic layer performs intent detection and negative sample enhancement on the fused question content to obtain question intent features. The semantic layer includes an intent detection module and a negative sample generation module. The intent detection module uses an MLP classifier to obtain the question intent in the fused question content. The negative sample generation module obtains question intent features by semantic substitution and adversarial negative sample processing.

[0044] The grammar layer determines candidate answers based on question intent features, industry description information, and industry knowledge base. The specific steps are as follows: calculate the cosine similarity between the industry description information vector and the knowledge vector in the industry knowledge base; extract the timestamp, category label, and knowledge weight of the knowledge vector with the highest similarity; determine the optimized knowledge graph and optimized knowledge based on the timestamp; and select the subgraphs of the optimized knowledge graph based on question intent features to obtain candidate answers.

[0045] The strategy layer determines candidate answers and generates response content based on user profiles and knowledge weights. The specific steps are as follows: select a response template based on category labeling, sort candidate answers according to knowledge weights and user profiles, take the candidate answers with the highest scores as response features, input the response features into the response template to obtain template response content, and input the template response content and user profiles into the language module for language polishing to obtain response content; the language module is a pre-trained LLAMA-2.

[0046] Furthermore, the method for obtaining a first reward signal by performing repeatability detection on the answer content includes:

[0047] Set up a sliding dialog window and define the question-and-answer rounds. Divide the answers within the sliding dialog window into multiple answer items, each containing at least one response feature. Calculate the Euclidean distance between each answer item and the corresponding industry optimization knowledge. Select the optimization knowledge with the closest Euclidean distance as the similarity feature. Count the historical usage frequency of the corresponding similar feature. Determine the first reward signal based on the historical usage frequency and knowledge weight of the similar feature. The expression is:

[0048] ,

[0049] in This is the first reward signal. m This represents the number of categories of response features corresponding to similar features within the sliding dialog window. Similarity features i The corresponding knowledge weight, Similar features within a sliding dialog window i The frequency of responses corresponding to response characteristics. Similarity features i Historical extraction count Similarity features i The standard number of extractions.

[0050] Furthermore, the method for determining the second reward signal includes:

[0051] The system collects user feedback information within a sliding dialog window; this feedback information includes evaluation attitude and question frequency; the evaluation attitude includes positive and negative evaluations; and the question frequency includes progressive retrieval frequency and repetitive retrieval frequency.

[0052] The time weight is determined based on the question-and-answer time, the positive and negative evaluation levels are determined based on the evaluation attitude, and the second reward signal is determined based on the time weight, positive evaluation level, negative evaluation level, progressive retrieval frequency, and repeated retrieval frequency; the time weight is negatively correlated with the question time; the second reward signal is positively correlated with the positive evaluation level and progressive retrieval frequency, and negatively correlated with the negative evaluation level and repeated retrieval frequency.

[0053] Furthermore, the method for obtaining dynamic reward weights includes:

[0054] A question-and-answer memory adjustment model is constructed. The question intent features, answer features, and user question response information within the entire dialogue window are input into the question-and-answer memory adjustment model to obtain dynamic reward weights. The dynamic reward weights include a first reward signal weight and a second reward signal weight.

[0055] The question-and-answer memory adjustment model includes a recurrent spatiotemporal layer and a self-attention layer. The recurrent spatiotemporal layer uses a parallel layer of RNN and LSTM to capture long-term dialogue dependencies based on question intent features, answer features, and user question response information, and outputs a first historical dialogue feature and a second historical dialogue feature respectively. The first historical dialogue feature and the second historical dialogue feature are weighted and fused to obtain the historical dialogue feature. The self-attention layer divides the historical dialogue feature according to the degree of answer repetition and the intensity of user response, performs numerical encoding to obtain answer repetition encoding and user response encoding, and uses a self-attention mechanism to obtain the first reward signal weight and the second reward signal weight based on the answer repetition encoding and the user response encoding.

[0056] Furthermore, the method for obtaining optimized answer content includes:

[0057] A reward objective function is constructed based on the first reward signal, the second reward signal, and the dynamic reward weights. The reward objective function includes a repetition reward item, a user response reward item, and a stability constraint item. The repetition reward item is positively correlated with the first reward signal and its weight. The user response reward item is positively correlated with the second reward signal and its weight. The stability constraint item limits the update amplitude of the strategy using KL divergence to prevent parameter mutations.

[0058] A group search strategy is used to optimize the parameters of the question-answering reinforcement learning model. The model is iterated multiple times until the reward objective function is maximized or the maximum number of iterations is reached. The optimal parameters of the question-answering reinforcement learning model are then output. Multimodal question content is input into the optimized question-answering reinforcement learning model to obtain optimized answer content.

[0059] The beneficial effects of this invention are:

[0060] This invention is a training method for an intelligent question-answering system based on machine learning. Compared with existing technologies, this invention has the following technical advantages:

[0061] This invention, through steps of acoustic data optimization, multimodal fusion, model construction, reward signal feedback, and model optimization, can enhance data preprocessing capabilities in the training of intelligent question-answering systems. It can reduce the cost and time of model training, thereby improving the efficiency and accuracy of intelligent question-answering system training. Optimizing intelligent question-answering system training technology improves the performance and user experience of the intelligent question-answering system, significantly saving resources and increasing work efficiency. It provides strong technical support for intelligent question-answering system training and has significant practical implications for promoting the deep application of intelligent question-answering technology in vertical fields. Attached Figure Description

[0062] Picture 1 This is a flowchart illustrating the steps of a machine learning-based intelligent question-answering system training method according to the present invention. Detailed Implementation

[0063] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0064] The present invention provides a training method for an intelligent question-answering system based on machine learning, comprising the following steps:

[0065] like Picture 1 As shown, this embodiment includes the following steps:

[0066] Obtain knowledge learning samples, optimize and classify the knowledge learning samples to obtain optimized knowledge, determine knowledge weights based on the sources of the optimized knowledge, and construct an industry knowledge base based on the optimized knowledge and the knowledge weights;

[0067] Obtain user information and multimodal question content, perform multimodal fusion on the multimodal question content to obtain fused question content, and obtain industry description information and user profile based on the fused question content and the user information;

[0068] The fused question content, the industry description information, and the user profile are input into the question-answering reinforcement learning model to obtain the answer content. The answer content is subjected to repetition detection to obtain a first reward signal. User question response information is obtained, and a second reward signal is determined based on the user question response information.

[0069] The answer content and user question response information are input into the question-and-answer memory adjustment model to obtain dynamic reward weights;

[0070] The question-answering reinforcement learning model is adjusted based on the first reward signal, the second reward signal, and the dynamic reward weight. The multimodal question content is input into the adjusted question-answering reinforcement learning model to obtain optimized answer content.

[0071] A user interaction log is generated based on the fused question content and the optimized answer content, and the industry knowledge base is updated using the user interaction log and the knowledge weight.

[0072] In this embodiment, the method for obtaining optimization knowledge includes:

[0073] A dual knowledge channel is constructed to acquire knowledge learning samples; the dual knowledge channel includes a static knowledge channel and a dynamic knowledge channel.

[0074] The static knowledge channel injects and preprocesses industry ontology information; the industry ontology information includes an industry ontology library, industry standards, an industry structured knowledge base, and an industry language model.

[0075] The dynamic knowledge channel injects and preprocesses real-time industry data; the real-time industry data includes industry news streams, user interaction logs, and sensor data streams; the update condition for the dynamic knowledge channel is triggered when both the new knowledge confidence score and the graph connectivity are greater than the corresponding update threshold; the expression for calculating the new knowledge confidence score is:

[0076] ,

[0077] in for t Time information e confidence level Weighting based on the authority of the source, SA To ensure the authority of the information source, For decay rate, For the current time, This refers to the last time the information was updated.

[0078] A support vector machine classifier is used to classify knowledge learning samples according to industry, constructing basic knowledge graphs for different industries. Dynamic pruning is then applied to these basic knowledge graphs, with the following pruning conditions:

[0079] ,

[0080] ,

[0081] ,

[0082] in For nodes A function to determine whether the pruning conditions are met. For the current time, For nodes timestamp, T As a time threshold, For nodes Similarity to the corresponding cluster centers The similarity threshold for the current time. For nodes Importance rating The threshold for node importance. Based on the basic similarity threshold, The update frequency for the current time. K The update frequency for the current time. For nodes Degree centrality, For degree centrality weights, For nodes betweenness centrality, For betweenness centrality weights, For nodes PageRank value, As the PageRank value weight;

[0083] Based on the basic knowledge graph, knowledge distillation is performed on the corresponding knowledge learning samples to obtain knowledge distillation samples. Based on the knowledge distillation samples, the corresponding basic knowledge graph is incrementally updated. The knowledge distillation and incremental update steps are repeated until the graph stability index reaches the stability threshold. Optimized knowledge and optimized knowledge graph are then output.

[0084] The knowledge distillation is achieved through a teacher-student model, a target distillation loss function, and dynamic curriculum learning. The teacher-student model includes a teacher model and a student model. The teacher model outputs the entity relationship probability distribution based on a relational graph convolutional network of the basic knowledge graph. The student model outputs high-confidence prediction results based on a lightweight GraphSAGE model. The target distillation loss function is expressed as follows:

[0085] ,

[0086] ,

[0087] ,

[0088] in Let the target distillation loss function be... For the knowledge distillation loss function, , , These are the weighting coefficients. For Kullback-Leibler divergence, , Output probabilities for the teacher and student models. , The adjacency matrix for the teacher model and the student model. For Jensen-Shannon divergence, , Embed matrices for teacher and student model nodes. To balance the intensity of the adversarial training, Let the loss function be that of the discriminator. To distinguish the loss term output by the teacher model, To differentiate the loss term output by the student model, D For discriminator, As input knowledge learning samples, Output for the teacher model, Output for the student model. To maximize the student model loss;

[0089] In practical evaluation, taking intelligent question-answering training in the field of oilfield safety and environmental protection as an example, industry ontology information (oilfield environmental protection standard documents, such as "People's Republic of China Petroleum and Natural Gas Industry Standard SY / T 6788-2020", etc.; industry structured knowledge base, such as open datasets and paper databases in the oilfield field; oilfield ontology library, such as oilfield thesaurus, pollutant classification, detection methods, etc.) is injected through static knowledge channels, while real-time industry data (industry news stream, such as oilfield news and announcements; user interaction logs, such as discussion posts and user-uploaded question logs; sensor data stream, such as soil trace element content detection records of a certain oilfield) is injected through dynamic knowledge channels.

[0090] Taking the updating of sensor data streams in a dynamic knowledge channel as an example, the monitoring records of trace element content in oilfield soil (weighting based on the authority of the source) Take 0.7, the authority of the information source. SA Take 0.85, decay rate If the value is set to 0.02 (the difference between the current time and the last update time is 6 hours), the confidence level of the new entity is 0.862 (confidence threshold 0.8) and the graph connectivity is 0.78 (connectivity threshold 0.7), and the update condition is triggered at this time.

[0091] A basic knowledge graph for the oilfield industry is constructed based on knowledge learning samples. This basic knowledge graph is then dynamically pruned. Taking a subgraph of the knowledge graph constructed according to the "Petroleum and Natural Gas Industry Standard SY / T 6788-2010 of the People's Republic of China" as an example, a time threshold is used. T Similarity threshold for 5 years and the current time The value is 0.7, which corresponds to the current time of the knowledge graph subgraph of this standard. With nodes timestamp The difference is 15 years, and the current cluster center similarity is 0.3, so a pruning operation is performed on this subgraph;

[0092] The teacher model outputs entity relationships based on the pruned basic knowledge graph (e.g., node 1 corresponds to the safety threshold of lead 300mg / kg, and node 2 corresponds to the detection method of ICP-MS). The student model outputs prediction results with a confidence level greater than 0.9 based on the lightweight GraphSAGE model. The optimized knowledge and optimized knowledge graph are output through repeated distillation using the target distillation loss function.

[0093] In this embodiment, the method for constructing an industry knowledge base includes the following steps:

[0094] The authority of optimization knowledge is determined based on its source, and its weight is determined based on the authority, timeliness, and consistency of the optimization knowledge. The expression is as follows:

[0095] ,

[0096] ,

[0097] in To optimize knowledge i Knowledge weight, To optimize the authority of knowledge, For decay rate, For the current time, This refers to the last time the information was updated. For the number of conflict relations, The total number of relations. To optimize knowledge i Source category weight, For the sake of the source's authority, To optimize knowledge i The number of citations; the above source categories include public datasets, expert-annotated data, user interaction data, and other data;

[0098] To optimize knowledge, timestamps and category labels are set. Knowledge vectors are generated from category labels, timestamps, and knowledge weights. Timestamps are set for the optimized knowledge graph. An industry knowledge base is constructed based on the optimized knowledge and the optimized knowledge graph. The industry knowledge base includes a graph database, a source database, and a vector database. The graph database stores optimized knowledge graphs from different industries. The source database stores optimized knowledge from different industries. The vector database stores different knowledge vectors.

[0099] In actual assessments, taking the knowledge weight calculation of user-submitted oilfield trace element monitoring data as an example, the user-submitted data "lead content 320mg / kg" (source authority) The value of 0.5 (which conflicts with the national standard of 300 mg / kg) indicates a conflict. The total number of relations is 1. The number of times the national standard is cited is 5. The value is 10, the time difference between the last update and the current time is 4 years, and the decay rate is 10. Optimize knowledge to 0.1 i Source category weight A value of 0.5 is used to maximize the knowledge authority. The knowledge weight of the user-submitted oilfield trace element monitoring data is calculated to be 1.08, with a value of 0.9.

[0100] In this embodiment, the method for obtaining industry description information and user profiles includes:

[0101] Multimodal question content is input into a question fusion model to obtain fused question content. The question fusion model includes a feature processing layer, a spatiotemporal alignment strategy, a hierarchical fusion strategy, and an adversarial alignment enhancement strategy. The feature processing layer processes the multimodal question content through text units, audio units, and image units to obtain multimodal question features. The multimodal question features include text features, speech-text features, and image-text features. The spatiotemporal alignment strategy synchronizes text features with speech-text features in time and spatially registers text features with image-text features. The hierarchical fusion strategy fuses the multimodal question features to obtain fused question features. The adversarial alignment enhancement strategy improves the accuracy of modality fusion through a modality discriminator and generative adversarial training.

[0102] User information is divided into first user features and user interaction logs. The user interaction logs are hierarchically clustered to obtain first question features and first interaction needs. The merged question content is hierarchically clustered to obtain merged user features and merged industry features. The merged user features include second question features and second user features. The merged industry features include first user industry features and first question industry features.

[0103] The third user feature is obtained by cross-referencing the second user feature with the first user industry feature. The fourth user feature is obtained by fusing the third user feature with the first user feature. The second question feature is corrected using the first question feature to obtain the third question feature. The user profile is obtained by splicing the first interaction requirement, the third question feature and the fourth user feature.

[0104] The first question industry feature is modified by using the first question feature to obtain the second question industry feature. The intersection of the first user feature and the first user industry feature is obtained to obtain the second user industry feature. The second question industry feature and the second user industry feature are concatenated to obtain the industry description information.

[0105] In practical evaluation, the text unit uses the BERT-wwm model to process the text query content to obtain text features, the audio unit uses the Conformer model to convert speech to text and perform paralinguistic extraction to obtain speech-text features, and the image unit uses VisionTransformer, YOLOv7 algorithm and graph convolutional network to process the image query content to obtain image-text features. The spatiotemporal alignment strategy synchronizes the text features with the speech-text features in time by using the speech timestamp and the text word segmentation position, and establishes region-word alignment rules through an attention mechanism to spatially register the text features with the image-text features. The hierarchical fusion strategy processes multimodal query features through feature concatenation, cross-modal Transformer and weighted voting of multimodal prediction results to obtain fused query features.

[0106] The question fusion model fuses multimodal question data (text question: I am an oilfield engineer. When the soil lead content is measured at 307.1 mg / kg, does it meet the 2023 environmental protection standards?; image question: upload a screenshot of the test report; voice question: the test point is oilfield area A, and the sampling depth is 1.5 meters). First, the voice text features are aligned with the text features using the voice timestamp (00:05-00:10). Then, the image text features are aligned with the text features based on "307.1" using an attention mechanism. Finally, the fused question content is obtained through a hierarchical fusion strategy [oilfield engineer, Pb, 307.1 mg / kg, environmental protection standards, pollution assessment, monitoring point, A, depth 1.5 m, urgency 0.8].

[0107] User information is divided into first user characteristics [30 years old, male, oilfield engineer, scientific research] and user interaction logs. The user interaction logs are hierarchically clustered to obtain first question characteristics [urgent, detailed, numerical, comparison] and first interaction needs [timely, rigorous, conclusion, analysis, standard]. The merged question content is hierarchically clustered to obtain merged user characteristics and merged industry characteristics. The merged user characteristics include second question characteristics [rigorous, numerical, comparison] and second user characteristics [oilfield engineer]. The merged industry characteristics include first user industry characteristics [scientific research, numerical, standard] and first question industry characteristics [oil and gas industry, environmental protection, safety].

[0108] The third user feature [oilfield engineer, scientific research] is obtained by cross-referencing the second user feature with the first user industry feature. The fourth user feature [30 years old, male, oilfield engineer, scientific research] is obtained by fusing the third user feature with the first user feature. The second question feature is corrected using the first question feature to obtain the third question feature [urgent, rigorous, numerical, comparison]. The user profile [30 years old, male, oilfield engineer, scientific research, urgent, rigorous, numerical, comparison, timeliness, strict, conclusion, analysis, standard] is obtained by splicing the first interaction requirement, the third question feature and the fourth user feature.

[0109] The first question industry feature is modified using the first question feature to obtain the second question industry feature [oil and gas industry, environmental protection, safety, detailed, numerical]. The intersection of the first user feature and the first user industry feature is taken to obtain the second user industry feature [oil field, scientific research]. The second question industry feature and the second user industry feature are concatenated to obtain the industry description information [oil and gas industry, scientific research, environmental protection, safety, detailed, numerical].

[0110] In this embodiment, the method for obtaining the answer content includes:

[0111] A question-answering reinforcement learning model is constructed, and the question content is input into the model to obtain the answer content; the question-answering reinforcement learning model includes a semantic layer, a syntactic layer, and a policy layer;

[0112] The semantic layer performs intent detection and negative sample enhancement on the fused question content to obtain question intent features. The semantic layer includes an intent detection module and a negative sample generation module. The intent detection module uses an MLP classifier to obtain the question intent in the fused question content. The negative sample generation module obtains question intent features by semantic substitution and adversarial negative sample processing.

[0113] The grammar layer determines candidate answers based on question intent features, industry description information, and industry knowledge base. The specific steps are as follows: calculate the cosine similarity between the industry description information vector and the knowledge vector in the industry knowledge base; extract the timestamp, category label, and knowledge weight of the knowledge vector with the highest similarity; determine the optimized knowledge graph and optimized knowledge based on the timestamp; and select the subgraphs of the optimized knowledge graph based on question intent features to obtain candidate answers.

[0114] The strategy layer determines candidate answers and generates response content based on user profiles and knowledge weights. The specific steps are as follows: select a response template based on category labeling, sort candidate answers according to knowledge weights and user profiles, take the group of candidate answers with the highest score as response features, input the response features into the response template to obtain template response content, and input the template response content and user profile into the language module for language polishing to obtain response content; the language module is a pre-trained LLAMA-2.

[0115] In practical assessments, taking intelligent question-answering training in the field of oilfield safety and environmental protection as an example, the integrated question content [oilfield engineer, Pb, 307.1 mg / kg, environmental protection standards, pollution assessment, monitoring point, A, depth 1.5m, urgency 0.8] is input into the question-answering reinforcement learning model. In the semantic layer, the intent detection module uses an MLP classifier to obtain the question intent "pollution risk assessment - numerical precision comparison" and labels the question-answer pairs with the intent. Focal Loss is used to solve the class imbalance problem. The negative sample generation module first predicts replaceable positions using BiLSTM, then samples from the thesaurus of knowledge graphs for semantic replacement. At the same time, it locates the subject, verb, and object based on syntactic analysis, inverts them according to the rule base, and finally uses GAN to generate a semantically reasonable but logically contradictory question: "What is the latest environmental protection standard for soil lead content? If the lead content is '307.1 mg / kg,' is it exceeding the standard?".

[0116] At the grammatical level, the industry description information vector is matched to the oilfield industry knowledge base. The knowledge vector corresponds to the timestamp of 2023, the category label is safety and environmental protection, and the knowledge weight is 1.08. Based on the timestamp, the optimized knowledge graph and optimized knowledge are determined. Based on the question intent "pollution risk assessment - numerical precision comparison" and the question vector [Pb, 307.1, environmental protection standard], the candidate answers are obtained as "《People's Republic of China Petroleum and Natural Gas Industry Standard SY / T 6788-2020》", "2020", "safety threshold 300mg / kg", "sampling depth", "heavy metal migration analysis", "pollution source", and "minor pollution risk".

[0117] In the strategy layer, the response template "According to {standard}, the safe threshold for {pollutant} is {threshold}, the current detection value is {detection value}, {conclusion}, {recommendation}" is selected based on the category "Safety and Environmental Protection". The corresponding template response content is "According to the SY / T 6788-2020 standard, the safe threshold for lead content in oilfield soil is 300 mg / kg. The current detection value is 307.1 mg / kg. The lead content in the A1 oilfield area exceeds the standard. It is recommended to investigate the pollution source." Based on the user profile, the response content is refined to "Since there are no updated oilfield safety and environmental protection standards in 2023, according to the SY / T 6788-2020 standard issued in 2020, the safe threshold for lead content in oilfield soil is 300 mg / kg. The lead content detection value you obtained from the A1 oilfield area is 307.1 mg / kg, which exceeds the safe threshold of 300 mg / kg. It is recommended that you verify the following: 1. Re-test to confirm the accuracy of the data; 2. Investigate possible pollution sources; 3. Conduct heavy metal migration analysis."

[0118] In this embodiment, the method for obtaining a first reward signal by performing repeatability detection on the answer content includes:

[0119] Set up a sliding dialog window and define the question-and-answer rounds. Divide the answers within the sliding dialog window into multiple answer items, each containing at least one response feature. Calculate the Euclidean distance between each answer item and the corresponding industry optimization knowledge. Select the optimization knowledge with the closest Euclidean distance as the similarity feature. Count the historical usage frequency of the corresponding similar feature. Determine the first reward signal based on the historical usage frequency and knowledge weight of the similar feature. The expression is:

[0120] ,

[0121] in This is the first reward signal. m This represents the number of categories of response features corresponding to similar features within the sliding dialog window. Similarity features i The corresponding knowledge weight, Similar features within a sliding dialog window i The frequency of responses corresponding to response characteristics. Similarity features i Historical extraction count Similarity features i The standard number of extractions;

[0122] In the actual evaluation, the dialogue continued based on the previous questions. The number of question-and-answer rounds in the sliding dialogue window was set to 3. Three similar features were obtained from the 3 rounds of dialogue: “SY / T 6788-2020 standard”, “300mg / kg”, and “ICP-MS”. The corresponding knowledge weights, answer frequencies, historical extraction times, and standard extraction times were 0.75 / 3 / 10 / 5, 0.8 / 2 / 8 / 5, and 0.6 / 1 / 6 / 5, respectively. The first reward signal was calculated to be 2.59.

[0123] In this embodiment, the method for determining the second reward signal includes:

[0124] The system collects user feedback information within a sliding dialog window; this feedback information includes evaluation attitude and question frequency; the evaluation attitude includes positive and negative evaluations; and the question frequency includes progressive retrieval frequency and repetitive retrieval frequency.

[0125] The time weight is determined based on the question-and-answer time; the positive and negative evaluation levels are determined based on the evaluation attitude; and the second reward signal is determined based on the time weight, positive evaluation level, negative evaluation level, progressive retrieval frequency, and repeated retrieval frequency. The time weight is negatively correlated with the question-and-answer time; the second reward signal is positively correlated with the positive evaluation level and progressive retrieval frequency, and negatively correlated with the negative evaluation level and repeated retrieval frequency.

[0126] In the actual evaluation, the customer feedback during the three rounds of question-and-answer dialogue was "the solution is effective" (positive evaluation level 4), and the customer made three further searches. There were no negative evaluations or duplicate searches. The interval between the most recent questions was... Time weighting according to The second reward signal was calculated as 4.416, based on the product of the evaluation level, the number of searches, and the time weight.

[0127] In this embodiment, the method for obtaining dynamic reward weights includes:

[0128] A question-and-answer memory adjustment model is constructed. The question intent features, answer features, and user question response information within the entire dialogue window are input into the question-and-answer memory adjustment model to obtain dynamic reward weights. The dynamic reward weights include a first reward signal weight and a second reward signal weight.

[0129] The question-and-answer memory adjustment model includes a recurrent spatiotemporal layer and a self-attention layer. The recurrent spatiotemporal layer uses a parallel layer of RNN and LSTM to capture long-term dialogue dependencies based on question intent features, answer features, and user question response information, and outputs a first historical dialogue feature and a second historical dialogue feature respectively. The first historical dialogue feature and the second historical dialogue feature are weighted and fused to obtain the historical dialogue feature. The self-attention layer divides the historical dialogue feature according to the degree of answer repetition and the intensity of user response, performs numerical encoding to obtain answer repetition encoding and user response encoding, and uses a self-attention mechanism to obtain the first reward signal weight and the second reward signal weight based on the answer repetition encoding and the user response encoding.

[0130] In this embodiment, the method for obtaining optimized answer content includes:

[0131] A reward objective function is constructed based on the first reward signal, the second reward signal, and the dynamic reward weights. The reward objective function includes a repetition reward item, a user response reward item, and a stability constraint item. The repetition reward item is positively correlated with the first reward signal and its weight. The user response reward item is positively correlated with the second reward signal and its weight. The stability constraint item limits the update amplitude of the strategy using KL divergence to prevent parameter mutations.

[0132] The parameters of the question-answering reinforcement learning model are optimized by a group search strategy. The model is iterated multiple times until the reward objective function is maximized or the maximum number of iterations is reached. The optimal parameters of the question-answering reinforcement learning model are output. The multimodal question content is input into the optimized question-answering reinforcement learning model to obtain the optimized answer content.

[0133] In the actual evaluation, the question intent features, response features, and user question response information within the entire dialogue window were input into the question-answering memory adjustment model to obtain a first reward signal weight of 0.38 and a second reward signal weight of 0.62.

[0134] The repetition reward is calculated using the product of the first reward signal weight and the first reward signal, which is 0.984. The user feedback reward is calculated using the product of the second reward signal weight and the second reward signal, which is 2.738. The stability constraint is set to 0.05. The reward objective function before iteration is 3.772. A group search strategy is used to optimize the parameters of the question-answering reinforcement learning model until the reward objective function reaches its maximum of 5 in the 15th iteration. At this point, the optimal parameters of the question-answering reinforcement learning model are output (in the semantic layer: the word vector dimension of the word vector model is adjusted from 300 to 512, and the number of hidden layer nodes of the fully connected neural network of the intent classifier is adjusted from 256 to 128; in the syntactic layer: the text knowledge weight and case knowledge weight of the knowledge graph are adjusted from fixed weights to dynamic weights).

[0135] The multimodal question content is input into the optimized question-answering reinforcement learning model to obtain the optimized answer content: "Since there are no updated oilfield safety and environmental protection standards in 2023, according to the SY / T 6788-2020 standard issued in 2020, the safe threshold for lead content in oilfield soil is 300 mg / kg. The lead content test value you obtained from the A1 oilfield area is 307.1 mg / kg, which exceeds the safe threshold of 300 mg / kg by 1.7 mg / kg, with an exceedance rate of 0.57%. We suggest you perform the following verification: 1. Re-test to confirm the accuracy of the data; 2. Investigate possible sources of pollution, such as drilling operations or pipeline leaks; 3. Combine the ICP-MS detection method to conduct heavy metal migration analysis."

[0136] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A training method for an intelligent question-answering system based on machine learning, characterized in that, Includes the following steps: S1. Obtain knowledge learning samples, optimize and classify the knowledge learning samples to obtain optimized knowledge, determine knowledge weights based on the sources of the optimized knowledge, and construct an industry knowledge base based on the optimized knowledge and the knowledge weights. S2. Obtain user information and multimodal question content, perform multimodal fusion on the multimodal question content to obtain fused question content, and obtain industry description information and user profile based on the fused question content and the user information; S3. Input the fused question content, the industry description information and the user profile into the question-answering reinforcement learning model to obtain the answer content, perform repeatability detection on the answer content to obtain the first reward signal, obtain user question response information, and determine the second reward signal based on the user question response information; S4. Input the answer content and user question response information into the question-and-answer memory adjustment model to obtain dynamic reward weights; S5. Adjust the question-answering reinforcement learning model according to the first reward signal, the second reward signal and the dynamic reward weight, and input the multimodal question content into the adjusted question-answering reinforcement learning model to obtain optimized answer content; S6. Generate a user interaction log based on the fused question content and the optimized answer content, and update the industry knowledge base using the user interaction log and the knowledge weight; The method for constructing an industry knowledge base includes the following steps: The authority of optimization knowledge is determined based on its source, and its weight is determined based on the authority, timeliness, and consistency of the optimization knowledge. The expression is as follows: ； ； in To optimize knowledge Knowledge weight, To optimize the authority of knowledge, For decay rate, For the current time, This refers to the last time the information was updated. For the number of conflict relations, The total number of relations. To optimize knowledge Source category weight, For the sake of the source's authority, To optimize knowledge The number of citations; the above source categories include public datasets, expert-annotated data, user interaction data, and other data; To optimize knowledge, timestamps and category labels are set. Knowledge vectors are generated from category labels, timestamps, and knowledge weights. Timestamps are set for the optimized knowledge graph. An industry knowledge base is constructed based on the optimized knowledge and the optimized knowledge graph. The industry knowledge base includes a graph database, a source database, and a vector database. The graph database stores optimized knowledge graphs from different industries. The source database stores optimized knowledge from different industries. The vector database stores different knowledge vectors. The method for obtaining a first reward signal by performing repeatability detection on the answer content includes: Set up a sliding dialog window and define the question-and-answer rounds. Divide the answers within the sliding dialog window into multiple answer items, each containing at least one response feature. Calculate the Euclidean distance between each answer item and the corresponding industry optimization knowledge. Select the optimization knowledge with the closest Euclidean distance as the similarity feature. Count the historical usage frequency of the corresponding similar feature. Determine the first reward signal based on the historical usage frequency and knowledge weight of the similar feature. The expression is: ； in This is the first reward signal. This represents the number of categories of response features corresponding to similar features within the sliding dialog window. Similarity features The corresponding knowledge weight, Similar features within a sliding dialog window The frequency of responses corresponding to response characteristics. Similarity features Historical extraction count Similarity features The standard number of extractions.

2. The training method for an intelligent question-answering system based on machine learning according to claim 1, characterized in that, The method for obtaining optimization knowledge includes: A dual knowledge channel is constructed to acquire knowledge learning samples; the dual knowledge channel includes a static knowledge channel and a dynamic knowledge channel. The static knowledge channel injects and preprocesses industry ontology information; the industry ontology information includes an industry ontology library, industry standards, an industry structured knowledge base, and an industry language model. The dynamic knowledge channel injects and preprocesses real-time industry data; the real-time industry data includes industry news streams, user interaction logs, and sensor data streams; the update condition for the dynamic knowledge channel is triggered when both the new knowledge confidence score and the graph connectivity are greater than the corresponding update threshold; the expression for calculating the new knowledge confidence score is: ； in for Time information confidence level Weighting based on the authority of the source, To ensure the authority of the information source, For decay rate, For the current time, This refers to the last time the information was updated. A support vector machine classifier is used to classify knowledge learning samples according to industry, constructing basic knowledge graphs for different industries. Dynamic pruning is then applied to these basic knowledge graphs, with the following pruning conditions: ； ； ； in For nodes A function to determine whether the pruning conditions are met. For the current time, For nodes timestamp, As a time threshold, For nodes Similarity to the corresponding cluster centers The similarity threshold for the current time. For nodes Importance rating The threshold for node importance. Based on the basic similarity threshold, The update frequency for the current time. For the domain constant, For nodes Degree centrality, For degree centrality weights, For nodes betweenness centrality, For betweenness centrality weights, For nodes PageRank value, As the PageRank value weight; Based on the basic knowledge graph, knowledge distillation is performed on the corresponding knowledge learning samples to obtain knowledge distillation samples. Based on the knowledge distillation samples, the corresponding basic knowledge graph is incrementally updated. The knowledge distillation and incremental update steps are repeated until the graph stability index reaches the stability threshold. Optimized knowledge and optimized knowledge graph are then output. The knowledge distillation is achieved through a teacher-student model, a target distillation loss function, and dynamic curriculum learning. The teacher-student model includes a teacher model and a student model. The teacher model outputs the entity relationship probability distribution based on a relational graph convolutional network of the basic knowledge graph. The student model outputs high-confidence prediction results based on a lightweight GraphSAGE model. The target distillation loss function is expressed as follows: ； ； ； in Let the target distillation loss function be... For the knowledge distillation loss function, , , These are the weighting coefficients. For Kullback-Leibler divergence, , Output probabilities for the teacher and student models. , The adjacency matrix for the teacher model and the student model. For Jensen-Shannon divergence, , Embed matrices for teacher and student model nodes. To balance the intensity of the adversarial training, Let the loss function be that of the discriminator. To distinguish the loss term output by the teacher model, To differentiate the loss term output by the student model, For discriminator, As input knowledge learning samples, Output for the teacher model, Output for the student model. To maximize the student model loss.

3. The training method for an intelligent question-answering system based on machine learning according to claim 1, characterized in that, The methods for obtaining industry description information and user profiles include: Multimodal question content is input into a question fusion model to obtain fused question content. The question fusion model includes a feature processing layer, a spatiotemporal alignment strategy, a hierarchical fusion strategy, and an adversarial alignment enhancement strategy. The feature processing layer processes the multimodal question content through text units, audio units, and image units to obtain multimodal question features. The multimodal question features include text features, speech-text features, and image-text features. The spatiotemporal alignment strategy synchronizes text features with speech-text features in time and spatially registers text features with image-text features. The hierarchical fusion strategy fuses the multimodal question features to obtain fused question features. The adversarial alignment enhancement strategy improves the accuracy of modality fusion through a modality discriminator and generative adversarial training. User information is divided into first user features and user interaction logs. The user interaction logs are hierarchically clustered to obtain first question features and first interaction needs. The merged question content is hierarchically clustered to obtain merged user features and merged industry features. The merged user features include second question features and second user features. The merged industry features include first user industry features and first question industry features. The third user feature is obtained by cross-referencing the second user feature with the first user industry feature. The fourth user feature is obtained by fusing the third user feature with the first user feature. The second question feature is corrected using the first question feature to obtain the third question feature. The user profile is obtained by splicing the first interaction requirement, the third question feature and the fourth user feature. The first question industry feature is modified by using the first question feature to obtain the second question industry feature. The intersection of the first user feature and the first user industry feature is obtained to obtain the second user industry feature. The second question industry feature and the second user industry feature are concatenated to obtain the industry description information.

4. The training method for an intelligent question-answering system based on machine learning according to claim 1, characterized in that, The method for obtaining the answer content includes: A question-answering reinforcement learning model is constructed, and the question content is input into the model to obtain the answer content; the question-answering reinforcement learning model includes a semantic layer, a syntactic layer, and a policy layer; The semantic layer performs intent detection and negative sample enhancement on the fused question content to obtain question intent features. The semantic layer includes an intent detection module and a negative sample generation module. The intent detection module uses an MLP classifier to obtain the question intent in the fused question content. The negative sample generation module obtains question intent features by semantic substitution and adversarial negative sample processing. The grammar layer determines candidate answers based on question intent features, industry description information, and industry knowledge base. The specific steps are as follows: calculate the cosine similarity between the industry description information vector and the knowledge vector in the industry knowledge base; extract the timestamp, category label, and knowledge weight of the knowledge vector with the highest similarity; determine the optimized knowledge graph and optimized knowledge based on the timestamp; and select the subgraphs of the optimized knowledge graph based on question intent features to obtain candidate answers. The strategy layer determines candidate answers and generates response content based on user profiles and knowledge weights. The specific steps are as follows: select a response template based on category labeling, sort candidate answers according to knowledge weights and user profiles, take the candidate answers with the highest scores as response features, input the response features into the response template to obtain template response content, and input the template response content and user profiles into the language module for language polishing to obtain response content; the language module is a pre-trained LLAMA-2.

5. The training method for an intelligent question-answering system based on machine learning according to claim 1, characterized in that, The method for determining the second reward signal includes: The system collects user feedback information within a sliding dialog window; this feedback information includes evaluation attitude and question frequency; the evaluation attitude includes positive and negative evaluations; and the question frequency includes progressive retrieval frequency and repetitive retrieval frequency. The time weight is determined based on the question-and-answer time, the positive and negative evaluation levels are determined based on the evaluation attitude, and the second reward signal is determined based on the time weight, positive evaluation level, negative evaluation level, progressive retrieval frequency, and repeated retrieval frequency; the time weight is negatively correlated with the question time; the second reward signal is positively correlated with the positive evaluation level and progressive retrieval frequency, and negatively correlated with the negative evaluation level and repeated retrieval frequency.

6. The training method for an intelligent question-answering system based on machine learning according to claim 1, characterized in that, The method for obtaining dynamic reward weights includes: A question-and-answer memory adjustment model is constructed. The question intent features, answer features, and user question response information within the entire dialogue window are input into the question-and-answer memory adjustment model to obtain dynamic reward weights. The dynamic reward weights include a first reward signal weight and a second reward signal weight. The question-and-answer memory adjustment model includes a recurrent spatiotemporal layer and a self-attention layer. The recurrent spatiotemporal layer uses a parallel layer of RNN and LSTM to capture long-term dialogue dependencies based on question intent features, answer features, and user question response information, and outputs a first historical dialogue feature and a second historical dialogue feature respectively. The first historical dialogue feature and the second historical dialogue feature are weighted and fused to obtain the historical dialogue feature. The self-attention layer divides the historical dialogue feature according to the degree of answer repetition and the intensity of user response, performs numerical encoding to obtain answer repetition encoding and user response encoding, and uses a self-attention mechanism to obtain the first reward signal weight and the second reward signal weight based on the answer repetition encoding and the user response encoding.

7. The training method for an intelligent question-answering system based on machine learning according to claim 1, characterized in that, The method for obtaining optimized answer content includes: A reward objective function is constructed based on the first reward signal, the second reward signal, and the dynamic reward weights. The reward objective function includes a repetition reward item, a user response reward item, and a stability constraint item. The repetition reward item is positively correlated with the first reward signal and its weight. The user response reward item is positively correlated with the second reward signal and its weight. The stability constraint item limits the update amplitude of the strategy using KL divergence to prevent parameter mutations. A group search strategy is used to optimize the parameters of the question-answering reinforcement learning model. The model is iterated multiple times until the reward objective function is maximized or the maximum number of iterations is reached. The optimal parameters of the question-answering reinforcement learning model are then output. Multimodal question content is input into the optimized question-answering reinforcement learning model to obtain optimized answer content.

Images