Intelligent question-answering system training method based on machine learning

Abstract

The invention discloses an intelligent question-answering system training method based on machine learning, and the method comprises the steps: optimizing a knowledge learning sample to obtain optimized knowledge, determining a knowledge weight, constructing an industry knowledge base, fusing multi-modal question content to obtain fused question content, and carrying out the training of the intelligent question-answering system. Obtaining industry description information and a user portrait according to the fused question content and the user information, and inputting the fused question content, the industry description information and the user portrait into a question and answer reinforcement learning model to obtain answer content, and obtaining a first reward signal, a second reward signal and a dynamic reward weight according to the answer content and the user question response information, adjusting the question and answer reinforcement learning model, inputting the multi-modal question content into the adjusted question and answer reinforcement learning model to obtain optimized answer content, and updating the industry knowledge base by adopting a user interaction log and the knowledge weight. According to the method, intelligent question and answer system training can be accurately and efficiently carried out, and meanwhile powerful support is provided for intelligent question and answer services of all industries.

Description

Technical Field

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

[0002] With the booming development of artificial intelligence technology, as the core carrier of human-computer interaction, intelligent question answering systems play a key role in fields such as customer service, intelligent education, and medical health. They can automatically understand user questions and output accurate answers. Therefore, building a highly accurate and strongly 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, there are obvious bottlenecks in traditional intelligent question answering system training technologies: First, most solutions have insufficient capabilities in fusing multi-source heterogeneous knowledge and lack a state determination mechanism for knowledge weights, resulting in limited accuracy of industry knowledge bases. Second, in multi-modal interaction scenarios, the ability to fuse and process multi-modal question content is weak, making it difficult to deeply explore the true needs of users. Finally, the model optimization mechanism is single and lacks a dynamic adjustment strategy for reward signals, restricting the refined training of the model. In response to the above dilemmas, the present invention proposes a training method for an intelligent question answering system based on machine learning. By obtaining knowledge learning samples through a dual knowledge channel, combining multi-modal fusion to generate industry description information and user portraits, through the collaboration of a question answering reinforcement learning model and a question answering memory adjustment model, realizing model optimization based on multiple reward signals and dynamic reward weights, and using user interaction logs to reversely update the industry knowledge base, effectively overcoming the shortcomings of traditional technologies. It can not only effectively solve problems such as low efficiency of knowledge integration and rough model training in traditional technologies, but also significantly improve the industry adaptability, answer accuracy, and self-evolution ability of intelligent question answering systems, and has important practical significance for promoting the in-depth application of intelligent question answering technology in vertical fields. Summary of the Invention

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

[0005] To achieve the above object, the present invention is implemented according to the following technical solutions: The present invention includes the following steps: Obtain knowledge learning samples, optimize and classify the knowledge learning samples to obtain optimized knowledge, determine knowledge weights according to the sources of the optimized knowledge, and construct an industry knowledge base based on the optimized knowledge and the knowledge weights; Obtain user information and multi-modal question content, perform multi-modal fusion on the multi-modal question content to obtain fused question content, and obtain industry description information and user portraits according to the fused question content and the user information; Input the fused question content, the industry description information, and the user profile into the question-answering reinforcement learning model to obtain an answer content, perform a repeatability detection on the answer content to obtain a first reward signal, obtain user question response information, and determine a second reward signal according to the user question response information; Input the answer content and the user question response information into the question-answering memory adjustment model to obtain a dynamic reward weight; 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 multi-modal question content into the adjusted question-answering reinforcement learning model to obtain an optimized answer content; Form a user interaction log according to the fused question content and the optimized answer content, and update the industry knowledge base by using the user interaction log and the knowledge weight.

[0006] Further, the method for obtaining optimized knowledge includes: Construct a dual knowledge channel, and obtain knowledge learning samples through the dual knowledge channel; 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 industry real-time data; the industry real-time data includes an industry news stream, a user interaction log, and a sensor data stream; the update condition of the dynamic knowledge channel is that the new knowledge confidence and the graph connectivity are both greater than the corresponding update thresholds to trigger the update condition; the calculation expression of the new knowledge confidence is: , where is t the confidence of the information at time e , is the weight of the source authority, SA is the authority of the information source, is the decay rate, is the current time, is the last update time of the information; Use a support vector machine classifier to classify the knowledge learning samples by industry, construct a basic knowledge graph for different industries, and perform dynamic pruning on the basic knowledge graph. The pruning conditions are: , , , where is the node Judgment function for whether pruning conditions are met is the current time is the node 's timestamp T is the time threshold is the node 's similarity with the corresponding cluster center is the similarity threshold at the current time is the node 's importance score is the node importance threshold is the basic similarity threshold is the update frequency at the current time K is the update frequency at the current time is the node 's degree centrality is the degree centrality weight is the node 's betweenness centrality is the betweenness centrality weight is the node 's PageRank value is the PageRank value weight; According to the basic knowledge graph, the corresponding knowledge learning samples are used for knowledge distillation to obtain knowledge distillation samples. According to 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, and the optimized knowledge and optimized knowledge graph are output; The knowledge distillation is realized 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 the relational graph convolutional network of the basic knowledge graph; the student model outputs high-confidence prediction results based on the lightweight GraphSAGE model; the expression of the target distillation loss function is: , , , where is the target distillation loss function is the knowledge distillation loss function , , are weight coefficients is the Kullback-Leibler divergence , are the output probabilities of the teacher model and the student model , is the adjacency matrix of the teacher model and the student model, is the Jensen-Shannon divergence, and are the node embedding matrices of the teacher model and the student model, is to balance the adversarial training intensity, is the loss function of the discriminator, is the loss term for distinguishing the output of the teacher model, is the loss term for distinguishing the output of the student model, D is the discriminator, is the input knowledge learning sample, is the output of the teacher model, is the output of the student model, is to maximize the loss of the student model.

[0007] Further, the method for constructing the industry knowledge base includes the following steps: Determine the authority of the optimized knowledge according to the source of the optimized knowledge, and determine the knowledge weight according to the authority, timeliness and consistency of the optimized knowledge. The expression is: , , where is the optimized knowledge i is the knowledge weight of is the authority of the optimized knowledge, is the decay rate, is the current time, is the last update time of the information, is the number of conflict relationships, is the total number of relationships, is the source category weight of the optimized knowledge i is the source authority of, is the source authority of is the optimized knowledge i is the number of citations of the optimized knowledge; the above source categories include public data sets, expert-annotated data, user interaction data, and other data; Set a timestamp and category annotation for the optimized knowledge, generate a knowledge vector from the category annotation, timestamp and knowledge weight, set a timestamp for the optimized knowledge graph, and construct an industry knowledge base according to 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 the optimized knowledge graphs of different industries; the source database stores the optimized knowledge of different industries; the vector database stores different knowledge vectors.

[0008] Further, the method for obtaining the industry description information and the user portrait includes: Input the multi-modal question content into the question fusion model to obtain the fused question content; the question fusion model includes a feature processing layer, a spatio-temporal alignment strategy, a hierarchical fusion strategy, and an adversarial alignment enhancement strategy; the feature processing layer processes the multi-modal question content through a text unit, an audio unit, and an image unit respectively to obtain multi-modal question features; the multi-modal question features include text features, speech text features, and image text features; the spatio-temporal alignment strategy synchronizes the text features with the speech text features in time and registers the text features with the image text features in space; the hierarchical fusion strategy fuses the multi-modal question features to obtain the fused question features; the adversarial alignment enhancement strategy improves the modal fusion accuracy through a modal discriminator and generative adversarial training; Divide the user information into first user features and user interaction logs, hierarchically cluster the user interaction logs to obtain first question features and first interaction requirements, and hierarchically cluster the fused question content to obtain fused user features and fused industry features; the fused user features include second question features and second user features; the fused industry features include first user industry features and first question industry features; Perform feature crossing on the second user features and the first user industry features to obtain third user features, perform feature fusion on the third user features and the first user features to obtain fourth user features, use the first question features to correct the second question features to obtain third question features, and splice the first interaction requirements, the third question features, and the fourth user features to obtain a user portrait; Use the first question features to correct the first question industry features to obtain second question industry features, take the intersection of the first user features and the first user industry features to obtain second user industry features, and splice the second question industry features and the second user industry features to obtain industry description information.

[0009] Further, the method for obtaining the answer content includes: Construct a question-answering reinforcement learning model, and input the fused question content into the question-answering reinforcement learning model to obtain the answer content; the question-answering reinforcement learning model includes a semantic layer, a grammar 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, and 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 processes the question intent through semantic replacement and adversarial negative samples to obtain question intent features; The syntax layer determines candidate answers through question intention features, industry description information, and an industry knowledge base. The specific steps are as follows: Calculate the cosine similarity between the industry description information vector and the knowledge vectors in the industry knowledge base, extract the timestamp, category annotation, and knowledge weight of the knowledge vector with the highest similarity, determine the optimized knowledge graph and optimized knowledge according to the timestamp, and filter the subgraph of the optimized knowledge graph based on the question intention features to obtain candidate answers; The strategy layer determines the candidate answers and generates response content based on the user profile and knowledge weight. The specific steps are as follows: Select a response template according to the category annotation, rank the candidate answers according to the knowledge weight and user profile, take the group of candidate answers with the highest score as the response feature, input the response feature into the response template to obtain the template response content, and input the template response content and the user profile into the language module for language polishing to obtain the response content; The language module is the pre-trained LLAMA-2.

[0010] Furthermore, the method for obtaining the first reward signal by performing repetitive detection on the response content includes: Set a sliding conversation window, set the number of question-and-answer rounds, divide the response content within the sliding conversation window into multiple response items, each response item contains at least one response feature, calculate the Euclidean distance between the response item and the corresponding industry-optimized knowledge, take the optimized knowledge with the closest Euclidean distance as the similar feature, count the historical usage times of the corresponding similar feature, and determine the first reward signal according to the historical usage times of the similar feature and the knowledge weight. The expression is: , where is the first reward signal, m is the number of categories of the response features corresponding to the similar feature within the sliding conversation window, is the similar feature i corresponding knowledge weight, is the similar feature within the sliding conversation window i corresponding response feature answer frequency, is the similar feature i historical extraction times, is the similar feature i standard extraction times.

[0011] Furthermore, the method for determining the second reward signal includes: Statistical user question response information within the sliding conversation window; The user question response information includes evaluation attitude and question frequency; The evaluation attitude includes positive evaluation and negative evaluation; The question frequency includes progressive retrieval frequency and repeated retrieval frequency; Determine the time weight according to the Q&A time, determine the positive evaluation level and negative evaluation level according to the evaluation attitude, and determine the second reward signal according to 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.

[0012] Further, the method for obtaining the dynamic reward weight includes: Construct a Q&A memory adjustment model, and input the question intention features, reply features, and user question response information within the entire dialogue window into the Q&A memory adjustment model to obtain the dynamic reward weight; the dynamic reward weight includes the first reward signal weight and the second reward signal weight; The Q&A memory adjustment model includes a recurrent spatio-temporal layer and a self-attention layer; the recurrent spatio-temporal layer uses a parallel layer of RNN and LSTM to capture long-term dialogue dependencies based on the question intention features, reply features, and user question response information, and respectively outputs the first historical dialogue feature and the second historical dialogue feature, and the historical dialogue feature is obtained by weighted fusion of the first historical dialogue feature and the second historical dialogue feature; the self-attention layer divides the historical dialogue feature according to the answer repetition degree and user response intensity, performs numerical encoding to obtain the answer repetition encoding and user response encoding, and uses the self-attention mechanism to obtain the first reward signal weight and the second reward signal weight according to the answer repetition encoding and user response encoding.

[0013] Further, the method for obtaining the optimized answer content includes: Construct a reward objective function according to the first reward signal, the second reward signal, and the dynamic reward weight; the reward objective function includes a repetition reward term, a user response reward term, and a stability constraint term; the repetition reward term is positively correlated with the first reward signal and the first reward signal weight; the user response reward term is positively correlated with the second reward signal and the second reward signal weight; the stability constraint term restricts the update amplitude through the KL divergence strategy to prevent parameter mutation; Use the group search strategy to optimize the parameters of the Q&A reinforcement learning model, iterate multiple times until the reward objective function is maximized or the maximum number of iterations is reached, output the optimal parameters of the Q&A reinforcement learning model, and input the multi-modal question content into the optimized Q&A reinforcement learning model to obtain the optimized answer content.

[0014] The beneficial effects of the present invention are: The present invention is a method for training an intelligent Q&A system based on machine learning. Compared with the prior art, the present invention has the following technical effects: Through steps of acoustic data optimization, multimodal fusion, model construction, reward signal feedback, and model optimization, the present invention can improve the data preprocessing ability in the training of intelligent question-answering systems, reduce the cost and time of model training, thereby improving the efficiency and accuracy of intelligent question-answering system training, optimizing the intelligent question-answering system training technology, enhancing the performance and user experience of intelligent question-answering systems, greatly saving resources, improving work efficiency, providing strong technical support for intelligent question-answering system training, and having important practical significance for promoting the in-depth application of intelligent question-answering technology in vertical fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of the steps of a method for training an intelligent question-answering system based on machine learning according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The present invention will be further described below through specific embodiments. The illustrative embodiments and explanations of the present invention are used to explain the present invention, but do not limit the present invention.

[0017] A method for training an intelligent question-answering system based on machine learning according to the present invention includes the following steps: As Figure 1 shown, in this embodiment, it includes the following steps: Obtain knowledge learning samples, optimize and classify the knowledge learning samples to obtain optimized knowledge, determine knowledge weights according to the source of the optimized knowledge, and construct an industry knowledge base based on the optimized knowledge and the knowledge weights; 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 portraits according to the fused question content and the user information; Input the fused question content, the industry description information, and the user portrait into a question-answering reinforcement learning model to obtain an answer content, perform a repeatability detection on the answer content to obtain a first reward signal, obtain user question response information, and determine a second reward signal according to the user question response information; Input the answer content and the user question response information into a question-answering memory adjustment model to obtain a dynamic reward weight; 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 an optimized answer content; Form a user interaction log according to the fused question content and the optimized answer content, and update the industry knowledge base using the user interaction log and the knowledge weights.

[0018] In this embodiment, the method for obtaining the optimized knowledge includes: Construct a dual knowledge channel and obtain knowledge learning samples through the dual knowledge channel; 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 industry real-time data; the industry real-time data includes an industry news stream, user interaction logs, and sensor data streams; the update condition of the dynamic knowledge channel is that the new knowledge confidence and graph connectivity are both greater than the corresponding update thresholds to trigger the update condition; the calculation expression of the new knowledge confidence is: , where is t the confidence of the time information e , is the weight of the source authority, SA is the authority of the information source, is the attenuation rate, is the current time, is the last update time of the information; Use a support vector machine classifier to classify the knowledge learning samples by industry, construct basic knowledge graphs for different industries, and perform dynamic pruning on the basic knowledge graphs. The pruning conditions are: , , , where is the judgment function for whether the node meets the pruning condition, is the current time, is the timestamp of the node , T is the time threshold, is the similarity between the node and the corresponding clustering center, is the similarity threshold at the current time, is the importance score of the node , is the node importance threshold, is the basic similarity threshold, is the update frequency at the current time, K is the update frequency at the current time, is the degree centrality of the node , is the degree centrality weight, For a node is the betweenness centrality, is the betweenness centrality weight, For a node is the PageRank value, is the PageRank value weight; According to the basic knowledge graph, the corresponding knowledge learning samples are subjected to knowledge distillation to obtain knowledge distillation samples. According to 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, and the optimized knowledge and optimized knowledge graph are output; The knowledge distillation is implemented 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 the relational graph convolutional network of the basic knowledge graph; the student model outputs high-confidence prediction results based on the lightweight GraphSAGE model; the expression of the target distillation loss function is: , , , where is the target distillation loss function, is the knowledge distillation loss function, , , are weight coefficients, is the Kullback-Leibler divergence, , are the output probabilities of the teacher model and the student model, , are the adjacency matrices of the teacher model and the student model, is the Jensen-Shannon divergence, , are the node embedding matrices of the teacher model and the student model, is to balance the adversarial training intensity, is the loss function of the discriminator, is the loss term for distinguishing the output of the teacher model, is the loss term for distinguishing the output of the student model, D is the discriminator, is the input knowledge learning sample, is the output of the teacher model, is the output of the student model, is to maximize the loss of the student model; In actual evaluation, taking the intelligent Q&A training in the field of oilfield safety and environmental protection as an example, industry ontology information (oilfield environmental protection standard documents, such as "Petroleum and Natural Gas Industry Standard SY / T 6788-2020 of the People's Republic of China", etc.; industry structured knowledge bases, such as publicly available datasets and paper libraries in the oilfield field; oilfield field ontology libraries, such as oilfield field thesaurus, pollutant classification, detection methods, etc.) is injected through static knowledge channels, and industry real-time data (industry news streams, such as oilfield field news, announcements, etc.; user interaction logs, such as discussion posts, question logs uploaded by users, etc.; sensor data streams, such as detection records of trace element content in the soil of a certain oilfield) is injected through dynamic knowledge channels; Taking the update of the sensor data stream in the dynamic knowledge channel as an example, the detection record of trace element content in the oilfield soil (the weight of the source authority is taken as 0.7, the authority of the information source SA is taken as 0.85, the decay rate is taken as 0.02, and the time difference between the current time and the last update time is 6h) corresponds to the new entity confidence of 0.862 (confidence threshold 0.8) and the graph connectivity of 0.78 (connectivity threshold 0.7). At this time, the update condition is triggered; Construct a basic knowledge graph for the oilfield industry according to the knowledge learning samples, and perform dynamic pruning on the basic knowledge graph. Taking the knowledge graph subgraph constructed based on "Petroleum and Natural Gas Industry Standard SY / T 6788-2010 of the People's Republic of China" as an example, the time threshold T is taken as 5 years, and the similarity threshold at the current time is taken as 0.7. The time at the current time of the knowledge graph subgraph corresponding to this standard and the node time stamp has a difference of 15 years, and the current clustering center similarity is 0.3. Therefore, pruning operation is performed on this subgraph; The teacher model is based on the probability distribution of the output entity relationships (such as node 1 lead corresponding to the attribute safety threshold of 300mg / kg, node 2 ICP-MS corresponding to the attribute detection method) of the pruned basic knowledge graph. The student model outputs a prediction result with a confidence greater than 0.9 based on the lightweight GraphSAGE model, and repeats distillation output to optimize knowledge and optimize the knowledge graph through the target distillation loss function.

[0019] In this embodiment, the method for constructing the industry knowledge base includes the following steps: Determine the authority of the optimized knowledge according to the source of the optimized knowledge, and determine the knowledge weight according to the authority, timeliness and consistency of the optimized knowledge. The expression is: , , where To optimize knowledge i of the knowledge weight, To optimize the authority of knowledge, is the attenuation rate, is the current time, is the last update time of the information, is the number of conflict relationships, is the total number of relationships, To optimize knowledge i of the source category weight, is the source authority, To optimize knowledge i of the citation times; the above source categories include public datasets, expert-annotated data, user interaction data, and other data; Set timestamps and category annotations for the optimized knowledge, generate knowledge vectors from category annotations, timestamps, and knowledge weights, set timestamps for the optimized knowledge graph, and construct an industry knowledge base 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 the optimized knowledge graphs of different industries; the source database stores the optimized knowledge of different industries; the vector database stores different knowledge vectors; In actual evaluation, taking the calculation of the knowledge weight of the oilfield trace element monitoring data submitted by the user as an example, the user submits the data "lead content 320 mg / kg" (source authority is 0.5) conflicts with the national standard "300 mg / kg", the number of conflict relationships is 1, the total number of relationships is 5, the national standard citation times is 10, the difference between the last update time of the information and the current time is 4 years, the attenuation rate is 0.1, the optimized knowledge i of the source category weight is 0.5, the maximum optimized knowledge authority is 0.9, and the calculated knowledge weight of the oilfield trace element monitoring data submitted by the user is 1.08.

[0020] In this embodiment, the method for obtaining the industry description information and the user portrait includes: Input the multi-modal question content into the question fusion model to obtain the fused question content; the question fusion model includes a feature processing layer, a spatio-temporal alignment strategy, a hierarchical fusion strategy, and an adversarial alignment enhancement strategy; the feature processing layer processes the multi-modal question content through a text unit, an audio unit, and an image unit respectively to obtain multi-modal question features; the multi-modal question features include text features, speech text features, and image text features; the spatio-temporal alignment strategy synchronizes the text features with the speech text features in time and registers the text features with the image text features in space; the hierarchical fusion strategy fuses the multi-modal question features to obtain the fused question features; the adversarial alignment enhancement strategy improves the modality fusion accuracy through a modality discriminator and generative adversarial training; Divide the user information into first user features and user interaction logs, hierarchically cluster the user interaction logs to obtain first question features and first interaction requirements, and hierarchically cluster the fused question content to obtain fused user features and fused industry features; the fused user features include second question features and second user features; the fused industry features include first user industry features and first question industry features; Perform feature cross between the second user features and the first user industry features to obtain third user features, perform feature fusion between the third user features and the first user features to obtain fourth user features, use the first question features to correct the second question features to obtain third question features, and splice the first interaction requirements, the third question features, and the fourth user features to obtain a user portrait; Use the first question features to correct the first question industry features to obtain second question industry features, take the intersection of the first user features and the first user industry features to obtain second user industry features, and splice the second question industry features and the second user industry features to obtain industry description information; In the actual evaluation, the text unit uses the BERT-wwm model to process the text question 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 the VisionTransformer, YOLOv7 algorithm, and graph convolutional network to process the image question content to obtain image text features; the spatio-temporal alignment strategy synchronizes and aligns the text features with the speech text features in time through the speech timestamp and the text tokenization position, and registers the text features with the image text features in space through the attention mechanism to establish a region-word alignment rule; the hierarchical fusion strategy processes the multi-modal question features through feature splicing, cross-modal Transformer, and weighted voting of multi-modal prediction results to obtain the fused question features; The question fusion model fuses multi-modal question data (text question: I am an oilfield engineer. When the detected value of soil lead content is 307.1 mg / kg, does it meet the environmental protection standards in 2023? Image question: Upload a screenshot of the test report; Voice question: The detection point is Area A of the oilfield, and the sampling depth is 1.5 meters). First, it aligns the voice text features with the text features through the voice timestamp (00:05 - 00:10), binds the image text features with the text features according to "307.1" through the attention mechanism, and obtains the fused question content [oilfield engineer, Pb, 307.1 mg / kg, environmental protection standards, pollution assessment, monitoring point, A, depth 1.5m, urgency 0.8] through the hierarchical fusion strategy; The user information is divided into the first user features [30 years old, male, oilfield engineer, scientific research] and the user interaction log. The user interaction log is hierarchically clustered to obtain the first question features [urgent, detailed, numerical, comparison] and the first interaction requirements [timeliness, strictness, conclusion, analysis, standard]. The fused question content is hierarchically clustered to obtain the fused user features and the fused industry features; The fused user features include the second question features [rigorous, numerical, comparison] and the second user features [oilfield engineer]; The fused industry features include the first user industry features [scientific research, numerical, standard], and the first question industry features [oil and gas industry, environmental protection, safety]; The second user features and the first user industry features are cross-characterized to obtain the third user features [oilfield engineer, scientific research]. The third user features and the first user features are feature-fused to obtain the fourth user features [30 years old, male, oilfield engineer, scientific research]. The first question features are used to correct the second question features to obtain the third question features [urgent, rigorous, numerical, comparison]. The first interaction requirements, the third question features, and the fourth user features are feature-stitched to obtain the user portrait [30 years old, male, oilfield engineer, scientific research, urgent, rigorous, numerical, comparison, timeliness, strictness, conclusion, analysis, standard]; The first question features are used to correct the first question industry features to obtain the second question industry features [oil and gas industry, environmental protection, safety, detailed, numerical]. The intersection of the first user features and the first user industry features is taken to obtain the second user industry features [oilfield, scientific research]. The second question industry features and the second user industry features are feature-stitched to obtain the industry description information [oil and gas industry, scientific research, environmental protection, safety, detailed, numerical].

[0021] In this embodiment, the method for obtaining the answer content includes: Construct a question-and-answer reinforcement learning model, and input the fused question content into the question-and-answer reinforcement learning model to obtain the answer content; The question-and-answer reinforcement learning model includes a semantic layer, a grammar 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 processes the question intent through semantic replacement and adversarial negative samples to obtain question intent features. The syntax layer determines candidate answers based on the 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 vectors in the industry knowledge base, extract the timestamp, category annotation, and knowledge weight of the knowledge vector with the highest similarity. Determine the optimized knowledge graph and optimized knowledge according to the timestamp. Screen the sub-graph of the optimized knowledge graph based on the question intent features to obtain candidate answers. The strategy layer determines the candidate answers and generates the response content based on the user profile and knowledge weight. The specific steps are as follows: Select a response template according to the category annotation, rank the candidate answers according to the knowledge weight and user profile, obtain the group of candidate answers with the highest score as the response feature, input the response feature into the response template to obtain the template response content, and input the template response content and user profile into the language module for language polishing to obtain the response content. The language module is the pre-trained LLAMA-2. In the actual evaluation, taking the intelligent question-answering training in the oilfield safety and environmental protection field as an example, the fused question content [Oilfield engineer, Pb, 307.1 mg / kg, environmental protection standard, pollution assessment, monitoring point, A, depth 1.5 m, 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 precise comparison" and mark the question-answer pairs of the intent. The Focal Loss is used to solve the class imbalance problem. The negative sample generation module first predicts the replaceable positions through BiLSTM, then samples and replaces them from the knowledge graph synonym library for semantic replacement. At the same time, it locates the subject, verb, and object according to the syntactic analysis, inverts them according to the rule library, and finally generates a question with reasonable semantics but logical contradictions through GAN: "What is the soil lead content standard in the latest environmental protection standard? If the lead content is '307.1 mg / kg', is it exceeding the standard?" In the syntax layer, the oilfield industry knowledge base is matched through the industry description information vector. The timestamp corresponding to the knowledge vector is 2023, the category annotation is safety and environmental protection, and the knowledge weight is 1.08. Determine the optimized knowledge graph and optimized knowledge according to the timestamp. According to the question intent "Pollution risk assessment - numerical precise comparison" and the question vector [Pb, 307.1, environmental protection standard], the candidate answers "Petroleum and Natural Gas Industry Standard of the People's Republic of China SY / T 6788-2020", "2020", "Safety threshold 300 mg / kg", "Sampling depth", "Heavy metal migration analysis", "Pollution source", "Slight pollution risk" are obtained. In the policy layer, according to the category label "safety and environmental protection", the reply template "According to the {standard}, the safety threshold of {pollutant} is {threshold}, the current detection value is {detection value}, {conclusion}, {recommendation}" is selected. The corresponding template reply content is "According to the SY / T 6788-2020 standard, the safety content threshold of lead in the soil of the oilfield area is 300 mg / kg, the current detection value is 307.1 mg / kg, the lead content in the A1 oilfield area exceeds the standard, and it is recommended to investigate the pollution source". According to the user profile, the language is polished to obtain the answer content "Since there is no updated oilfield safety and environmental protection standard in 2023, according to the SY / T 6788-2020 standard issued in 2020, the safety content threshold of lead in the soil of the oilfield area is 300 mg / kg. The lead content detection value you obtained from the A1 oilfield area is 307.1 mg / kg, and this detection value exceeds the safety content threshold of 300 mg / kg. It is recommended that you conduct the following verifications: 1. Retest to confirm the data accuracy; 2. Investigate possible pollution sources; 3. Conduct an analysis of the mobility of heavy metals".

[0022] In this embodiment, the method for obtaining the first reward signal by performing a repeatability detection on the answer content includes: Set a sliding dialogue window, set the number of question-and-answer rounds, divide the answer content within the sliding dialogue window into multiple answer items, each answer item contains at least one reply feature, calculate the Euclidean distance between the answer item and the optimized knowledge of the corresponding industry, and take the optimized knowledge with the closest Euclidean distance as the similar feature. Count the historical usage times of the corresponding similar feature, and determine the first reward signal according to the historical usage times of the similar feature and the knowledge weight. The expression is: , where is the first reward signal, m is the number of categories of the reply features corresponding to the similar feature within the sliding dialogue window, is the similar feature i corresponding knowledge weight, is the similar feature within the sliding dialogue window i corresponding reply feature answer frequency, is the similar feature i historical extraction times, is the similar feature i standard extraction times; In the actual evaluation, continue the conversation based on the previous questions. Set the number of question-and-answer rounds in the sliding conversation window to 3. Obtain three similar features, namely "SY / T 6788-2020 standard", "300 mg / kg", and "ICP-MS", from the 3 rounds of conversation. The corresponding knowledge weights, response frequencies, historical extraction times, and standard extraction times are 0.75 / 3 / 10 / 5, 0.8 / 2 / 8 / 5, 0.6 / 1 / 6 / 5 respectively. Calculate the first reward signal to be 2.59.

[0023] In this embodiment, the method for determining the second reward signal includes: Statistically analyze the user question response information within the sliding conversation window; the user question response information includes evaluation attitude and question frequency; the evaluation attitude includes positive evaluation and negative evaluation; the question frequency includes progressive retrieval frequency and repeated retrieval frequency; Determine the time weight according to the question-and-answer time, determine the positive evaluation level and negative evaluation level according to the evaluation attitude, and determine the second reward signal according to 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; In the actual evaluation, in the 3 rounds of question-and-answer conversations that continue, the customer feedback is "the solution is effective" (positive evaluation level 4), and there are 3 progressive retrievals, no negative evaluation and repeated retrieval, and the time interval of the most recent question , the time weight is calculated according to Calculate, and calculate the second reward signal to be 4.416 according to the product of the evaluation level, retrieval times, and time weight.

[0024] In this embodiment, the method for obtaining the dynamic reward weight includes: Construct a question-and-answer memory adjustment model, and input the question intention features, reply features, and user question response information within the entire conversation window into the question-and-answer memory adjustment model to obtain the dynamic reward weight; the dynamic reward weight includes the first reward signal weight and the second reward signal weight; The question-and-answer memory adjustment model includes a recurrent spatio-temporal layer and a self-attention layer; the recurrent spatio-temporal layer uses a parallel layer of RNN and LSTM to capture long-term dialogue dependencies based on the question intent feature, the reply feature, and the user question response information, and respectively outputs a first historical dialogue feature and a second historical dialogue feature. The historical dialogue feature is obtained by weighted fusion of the first historical dialogue feature and the second historical dialogue feature; the self-attention layer divides the historical dialogue feature according to the answer repetition degree and the user response intensity, performs numerical encoding to obtain an answer repetition encoding and a user response encoding, and uses the self-attention mechanism to obtain a first reward signal weight and a second reward signal weight according to the answer repetition encoding and the user response encoding.

[0025] In this embodiment, the method for obtaining the optimized answer content includes: Construct a reward objective function according to the first reward signal, the second reward signal, and the dynamic reward weight; the reward objective function includes a repetition reward term, a user response reward term, and a stability constraint term; the repetition reward term is positively correlated with the first reward signal and the first reward signal weight; the user response reward term is positively correlated with the second reward signal and the second reward signal weight; the stability constraint term restricts the update amplitude through the KL divergence strategy to prevent parameter mutation; Adopt a population search strategy to optimize the parameters of the question-and-answer reinforcement learning model, iterate multiple times until the reward objective function is maximized or the maximum number of iterations is reached, output the optimal parameters of the question-and-answer reinforcement learning model, and input the multi-modal question content into the optimized question-and-answer reinforcement learning model to obtain the optimized answer content; In the actual evaluation, the question intent feature, the reply feature, and the user question response information within the entire dialogue window are input into the question-and-answer memory adjustment model to obtain a first reward signal weight of 0.38 and a second reward signal weight of 0.62; The repetition reward term takes the product of the first reward signal weight and the first reward signal, which is 0.984. The user response reward term takes the product of the second reward signal weight and the second reward signal, which is 2.738. The stability constraint term takes 0.05. Calculate the reward objective function before iteration to be 3.772. Adopt a population search strategy to optimize the parameters of the question-and-answer reinforcement learning model. Until the 15th iteration, the reward objective function reaches the maximum of 5. At this time, output the optimal parameters of the question-and-answer reinforcement learning model (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 intent classifier fully connected neural network is adjusted from 256 to 128; in the syntax layer: the text knowledge weight and the case knowledge weight of the knowledge graph are adjusted from fixed weights to dynamic weights); Input the multi-modal question content into the optimized question-answering reinforcement learning model to obtain the optimized answer content: "Since there is no updated oilfield safety and environmental protection standard in 2023, according to the SY / T 6788-2020 standard issued in 2020, the safety content threshold of lead in the soil of the oilfield area is 300 mg / kg. The lead content detection value you obtained from the A1 oilfield area is 307.1 mg / kg, which exceeds the safety content threshold of 300 mg / kg, with an excess amount of 1.7 mg / kg and an over-standard rate of 0.57%. It is recommended that you conduct the following verifications: 1. Recheck to confirm the data accuracy; 2. Investigate possible pollution sources, such as drilling operations or pipeline leaks; 3. Combine the ICP-MS detection method to carry out the analysis of heavy metal mobility."

[0026] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in 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, Including the following steps: S1. Obtain knowledge learning samples, optimize and classify the knowledge learning samples to obtain optimized knowledge, determine knowledge weights according to 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 portraits according to the fused question content and the user information; S3. Input the fused question content, the industry description information, and the user portraits into a question-answering reinforcement learning model to obtain an answer content, perform a repetition detection on the answer content to obtain a first reward signal, obtain user question response information, and determine a second reward signal according to the user question response information; S4. Input the answer content and the user question response information into a question-answering memory adjustment model to obtain a dynamic reward weight; 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 an optimized answer content; S6. Form a user interaction log according to the fused question content and the optimized answer content, and update the industry knowledge base by using the user interaction log and the knowledge weights.

2. The training method of an intelligent question answering system based on machine learning according to claim 1, characterized in that, The method for obtaining the optimized knowledge includes: Construct a dual knowledge channel and obtain knowledge learning samples through the dual knowledge channel; 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 industry real-time data; the industry real-time data includes an industry news stream, a user interaction log, and a sensor data stream; the update condition of the dynamic knowledge channel is that the new knowledge confidence and the graph connectivity are both greater than the corresponding update thresholds to trigger the update condition; the calculation expression of the new knowledge confidence is: ， wherein is t the confidence of the e time information, is the weight of the source authority, SA is the authority of the information source, is the attenuation rate, is the current time, is the last update time of the information; Use a support vector machine classifier to classify the knowledge learning samples by industry, construct basic knowledge graphs for different industries, and perform dynamic pruning on the basic knowledge graphs. The pruning conditions are: ， ， ， where is a judgment function for whether the node satisfies the pruning condition, is the current time, is the timestamp of the node , T is the time threshold, is the similarity between the node and the corresponding cluster center, is the similarity threshold at the current time, is the importance score of the node , is the node importance threshold, is the basic similarity threshold, is the update frequency at the current time, K is the update frequency at the current time, is the degree centrality of the node , is the degree centrality weight, is the betweenness centrality of the node , is the betweenness centrality weight, is the PageRank value of the node , is the PageRank value weight; Perform knowledge distillation on the corresponding knowledge learning samples according to the basic knowledge graphs to obtain knowledge distillation samples, perform incremental updates on the corresponding basic knowledge graphs according to the knowledge distillation samples, and repeat the knowledge distillation and incremental update steps until the graph stability index reaches the stability threshold, and output the optimized knowledge and the optimized knowledge graph; The knowledge distillation is realized 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 an entity relationship probability distribution based on the relational graph convolutional network of the basic knowledge graph; the student model outputs high-confidence prediction results based on a lightweight GraphSAGE model; the expression of the target distillation loss function is: ， ， ， where is the target distillation loss function, is the knowledge distillation loss function, , , are weight coefficients, is the Kullback-Leibler divergence, , are the output probabilities of the teacher model and the student model, , are the adjacency matrices of the teacher model and the student model, is the Jensen-Shannon divergence, , are the node embedding matrices of the teacher model and the student model, is to balance the intensity of adversarial training, is the loss function of the discriminator, is the loss term for distinguishing the output of the teacher model, is the loss term for distinguishing the output of the student model, D is the discriminator, is the input knowledge learning sample, is the output of the teacher model, is the output of the student model, is to maximize the loss of the student model.

3. The training method of an intelligent question-answering system based on machine learning according to claim 1, characterized in that The method for constructing the industry knowledge base includes the following steps: Determine the authority of the optimized knowledge according to the source of the optimized knowledge, and determine the knowledge weight according to the authority, timeliness and consistency of the optimized knowledge. The expression is as follows: ， ， Among them To optimize knowledge i The knowledge weight of To optimize the authority of knowledge Is the attenuation rate Is the current time Is the last update time of the information Is the number of conflict relationships Is the total number of relationships To optimize knowledge i The source category weight of Is the source authority of To optimize knowledge i The number of citations of; The above source categories include public datasets, expert-annotated data, user interaction data, and other data; Set time stamps and category labels for the optimized knowledge, generate knowledge vectors from the category labels, time stamps and knowledge weights, set time stamps for the optimized knowledge graph, and construct an industry knowledge base according to 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 the optimized knowledge graphs of different industries; the source database stores the optimized knowledge of different industries; the vector database stores different knowledge vectors.

4. A training method for an intelligent question-answering system based on machine learning according to claim 1, characterized in that, The method for obtaining industry description information and user portraits includes: Input the multi-modal question content into a question fusion model to obtain fused question content; the question fusion model includes a feature processing layer, a spatio-temporal alignment strategy, a hierarchical fusion strategy and an adversarial alignment enhancement strategy; the feature processing layer processes the multi-modal question content through a text unit, an audio unit and an image unit respectively to obtain multi-modal question features; the multi-modal question features include text features, speech text features and image text features; the spatio-temporal alignment strategy synchronizes the text features with the speech text features in time and registers the text features with the image text features in space; the hierarchical fusion strategy fuses the multi-modal question features to obtain fused question features; the adversarial alignment enhancement strategy improves the accuracy of modal fusion through a modal discriminator and generative adversarial training. Divide the user information into first user features and user interaction logs, hierarchically cluster the user interaction logs to obtain first question features and first interaction requirements, and hierarchically cluster the fused question content to obtain fused user features and fused industry features; the fused user features include second question features and second user features; the fused industry features include first user industry features and first question industry features. Perform feature crossing on the second user features and the first user industry features to obtain third user features, perform feature fusion on the third user features and the first user features to obtain fourth user features, use the first question features to correct the second question features to obtain third question features, and splice the first interaction requirements, the third question features and the fourth user features to obtain a user portrait. Use the first question features to correct the first question industry features to obtain second question industry features, take the intersection of the first user features and the first user industry features to obtain second user industry features, and splice the second question industry features and the second user industry features to obtain industry description information.

5. The training method of an intelligent question answering system based on machine learning according to claim 1, wherein The method for obtaining answer content includes: Construct a question-answering reinforcement learning model, and input the fused question content into the question-answering reinforcement learning model to obtain answer content; the question-answering reinforcement learning model includes a semantic layer, a grammar 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 processes the question intent through semantic replacement and adversarial negative samples to obtain question intent features. The syntax layer determines candidate answers based on the question intention 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 vectors in the industry knowledge base, extract the timestamp, category annotation, and knowledge weight of the knowledge vector with the highest similarity, determine the optimized knowledge graph and optimized knowledge according to the timestamp, and filter the sub-graph of the optimized knowledge graph based on the question intention features to obtain candidate answers; The strategy layer determines the candidate answers based on the user profile and knowledge weight to generate the answer content. The specific steps are as follows: Select the reply template according to the category annotation, rank the candidate answers according to the knowledge weight and user profile, obtain the group of candidate answers with the highest score as the reply feature, input the reply feature into the reply template to obtain the template reply content, and input the template reply content and user profile into the language module for language polishing to obtain the answer content; The language module is the pre-trained LLAMA-2.

6. The training method of an intelligent question answering system based on machine learning according to claim 1, characterized in that, The method for obtaining the first reward signal by detecting the repeatability of the answer content includes: Set a sliding dialogue window, set the number of question-and-answer rounds, divide the answer content in the sliding dialogue window into multiple answer items, each answer item contains at least one reply feature, calculate the Euclidean distance between the answer item and the corresponding industry-optimized knowledge, take the optimized knowledge with the closest Euclidean distance as the similar feature, count the historical usage times of the corresponding similar feature, and determine the first reward signal according to the historical usage times of the similar feature and the knowledge weight. The expression is: ， Among them is the first reward signal, m is the number of categories of response features corresponding to similar features within the sliding dialogue window, is the similar feature i corresponding knowledge weight, is the similar feature within the sliding dialogue window i corresponding response frequency of the response feature, is the similar feature i historical extraction times, is the similar feature i standard extraction times.

7. A 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: Count the user question response information in the sliding dialogue window; The user question response information includes evaluation attitude and question frequency; The evaluation attitude includes positive evaluation and negative evaluation; The question frequency includes progressive retrieval frequency and repeated retrieval frequency; Determine the time weight according to the question-and-answer time, determine the positive evaluation level and negative evaluation level according to the evaluation attitude, and determine the second reward signal according to 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.

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

9. The training method of an intelligent question answering system based on machine learning according to claim 1, wherein The method for obtaining the optimized answer content includes: Constructing a reward objective function according to the first reward signal, the second reward signal, and the dynamic reward weight; the reward objective function includes a repetition reward term, a user response reward term, and a stability constraint term; the repetition reward term is positively correlated with the first reward signal and the first reward signal weight; the user response reward term is positively correlated with the second reward signal and the second reward signal weight; the stability constraint term restricts the update amplitude through the KL divergence strategy to prevent parameter mutation; Using a population search strategy to optimize the parameters of the question-and-answer reinforcement learning model, stopping after multiple iterations until the reward objective function is maximized or the maximum number of iterations is reached, outputting the optimal parameters of the question-and-answer reinforcement learning model, and inputting the multi-modal question content into the optimized question-and-answer reinforcement learning model to obtain the optimized answer content.

Images