Insurance customer service method and system based on natural language processing

Abstract

The invention provides an insurance customer service method and system based on natural language processing, and relates to the technical field of language processing, and the method comprises the steps: obtaining multi-source insurance customer service historical session data, carrying out the text preprocessing, and constructing an initial insurance knowledge graph based on deep semantic analysis; performing knowledge value evaluation on the newly added session content and dynamically updating a knowledge graph; a question and answer matching model of a two-way gating circulation unit network is adopted to carry out multi-hop type association matching, a path priority score is calculated through a path state evaluation model, and intelligent insurance customer service is achieved. The real-time performance and the accuracy of the insurance knowledge graph are improved, and the accuracy and the intelligent level of question and answer matching are enhanced.

Description

Technical Field

[0001] The present invention relates to language processing technologies, and in particular to an insurance customer service method and system based on natural language processing. Background Art

[0002] With the rapid development of the insurance industry, the demand for insurance customer service is increasing day by day. The traditional manual customer service model faces problems such as low efficiency and unstable service quality. To improve the customer service experience, an intelligent customer service system based on natural language processing technology has gradually become an important development direction in the insurance industry. Currently, insurance intelligent customer service systems mainly achieve the automatic answering function by constructing a knowledge base and a question-answering matching model. Through the analysis and knowledge extraction of historical conversation data, a professional knowledge base in the insurance field is established, and natural language processing technology is used for question-answering matching and answer generation.

[0003] The existing methods for constructing insurance knowledge graphs mainly rely on the update of static knowledge bases, lacking a dynamic analysis and knowledge value evaluation mechanism for real-time conversation data, resulting in untimely updates of the knowledge base and being unable to effectively handle newly emerging insurance consultation problems.

[0004] Traditional question-answering matching models often adopt simple semantic similarity calculation methods, failing to fully consider the professionalism and complex relevance of knowledge in the insurance field, resulting in low matching accuracy and difficulty in accurately understanding the true consultation intentions of users.

[0005] The existing answer generation strategies lack a multi-dimensional evaluation mechanism for knowledge paths, failing to simultaneously consider the accuracy, novelty, and diversity of answers, and are prone to generating repetitive or incomplete answers, affecting the consultation experience of users. Summary of the Invention

[0006] The embodiments of the present invention provide an insurance customer service method and system based on natural language processing, which can solve the problems in the prior art.

[0007] In the first aspect of the embodiments of the present invention, An insurance customer service method based on natural language processing is provided, including: Obtaining multi-source insurance customer service historical conversation data, performing text preprocessing on the multi-source insurance customer service historical conversation data to obtain standardized conversation text data; performing knowledge entity recognition and relationship extraction on the standardized conversation text data based on deep semantic analysis, and constructing an initial insurance knowledge graph based on the recognized knowledge entities and the extracted relationships; Monitor the historical conversation data of multi-source insurance customer service in real time. When new conversation content is detected, evaluate the knowledge value of the new conversation content based on the attention mechanism. The knowledge value evaluation is achieved by constructing a multi-dimensional scoring model, calculating the semantic similarity, knowledge coverage, and user consultation frequency between the new conversation content and the existing knowledge in the initial insurance knowledge graph, determining the knowledge update priority according to the multi-dimensional scoring results, and triggering the dynamic update of the initial insurance knowledge graph for the valuable knowledge content that meets the preset priority threshold; Receive the question text input by the user, perform intent recognition and information extraction on the question text to obtain the question features to be matched; input the question features into the question-answering matching model based on the bidirectional gated recurrent unit network. The question-answering matching model performs multi-hop association matching between the question features and the nodes in the initial insurance knowledge graph, calculates the path priority scores including value, novelty, and diversity through the path state evaluation model; select nodes based on the path priority scores and calculate the semantic similarity and coherence features of the nodes to obtain the immediate reward value, and generate the path comprehensive score by combining the predicted subsequent reward values.

[0008] Evaluate the knowledge value of the new conversation content based on the attention mechanism. The knowledge value evaluation is achieved by constructing a multi-dimensional scoring model, and calculating the semantic similarity, knowledge coverage, and user consultation frequency between the new conversation content and the existing knowledge in the initial insurance knowledge graph includes: Extract hierarchical features of the new conversation content at the word level, sentence level, and document level, perform semantic encoding on the new conversation content to obtain semantic vectors of different granularities, calculate the semantic correlation degree between the semantic vectors and the existing knowledge nodes in the initial insurance knowledge graph using the attention mechanism, dynamically fuse the multi-granularity similarities based on the adaptive weights generated by the attention network, and introduce contrastive learning loss for feature optimization to obtain the semantic similarity score; Generate the knowledge coverage score by calculating the overlap degree between the knowledge points in the new conversation content and the existing knowledge in the initial insurance knowledge graph; generate the user consultation frequency score by counting the occurrence frequency of the knowledge points involved in the new conversation content in the historical conversations; perform weighted fusion on the semantic similarity score, knowledge coverage score, and user consultation frequency score to obtain the final knowledge value evaluation score.

[0009] Extract hierarchical features of the new conversation content at the word level, sentence level, and document level, perform semantic encoding on the new conversation content to obtain semantic vectors of different granularities, calculate the semantic correlation degree between the semantic vectors and the existing knowledge nodes in the initial insurance knowledge graph using the attention mechanism, dynamically fuse the multi-granularity similarities based on the adaptive weights generated by the attention network, and introduce contrastive learning loss for feature optimization to obtain the semantic similarity score includes: The new conversation content and the knowledge nodes in the initial insurance knowledge graph are respectively input into a pre-trained language model to obtain word embedding sequences; the word embedding sequences are respectively mapped into query matrices, key matrices, and value matrices, the dot product of the query matrix and the key matrix is calculated to obtain attention scores, and the attention scores are multiplied by the value matrices to obtain word-level semantic vectors; Based on the word-level semantic vectors, sentence-level semantic representations are extracted, the sentence-level semantic representations are input into a hierarchical attention network to weight the information in the sentences, and sentence-level semantic vectors are generated; the sentence-level semantic vectors are input into a multi-layer perceptron, and through residual connection and layer normalization processing, document-level semantic vectors are obtained; The similarities between the word-level semantic vectors, sentence-level semantic vectors, and document-level semantic vectors of the new conversation content and the corresponding granularity semantic vectors of the knowledge nodes are calculated respectively to obtain word-level similarity scores, sentence-level similarity scores, and document-level similarity scores; An adaptive attention network is constructed, and the word-level similarity scores, sentence-level similarity scores, and document-level similarity scores and the context semantic features are input into the adaptive attention network to generate dynamic fusion weights; based on the dynamic fusion weights, the word-level similarity scores, sentence-level similarity scores, and document-level similarity scores are weighted and combined to obtain a multi-granularity fusion score; the multi-granularity fusion score is compared and learned with positive sample knowledge nodes and negative sample knowledge nodes to optimize the discrimination ability of the feature representation and obtain the final semantic similarity score.

[0010] The question features are input into a question-answering matching model based on a bidirectional gated recurrent unit network. The question-answering matching model performs multi-hop associative matching between the question features and the nodes in the initial insurance knowledge graph, including: An insurance knowledge graph containing multiple knowledge nodes is constructed, and the knowledge nodes are vectorized using a graph embedding method to obtain knowledge node feature vectors; A question-answering matching model with an update gate and a reset gate is constructed. The question features are input into the question-answering matching model, and the information transmission is controlled by the update gate and the reset gate. The question-answering matching model calculates the first-hop attention weights of the question features and the knowledge node feature vectors, and based on the first-hop attention weights, the knowledge node feature vectors are weighted and aggregated to obtain a first-round matching vector; Based on the first-round matching vector, the question features are updated, the updated question features are second-hop matched with the adjacent knowledge node features to obtain a second-round matching vector, and an adaptive weight fusion of the first-round matching vector and the second-round matching vector is performed through a gating mechanism to obtain a multi-hop fusion vector; Multi-hop traversal is performed on the problem feature and knowledge node feature vectors to obtain path encodings. After evaluating the value of the path encodings and the problem features, they are weighted and fused. The gated mechanism is used to perform adaptive residual connection between the fusion result and the multi-hop fusion vector. After layer normalization and processing by the deep feature transformation network, the optimal matching node is obtained by combining temperature-adjusted probability calculation and dynamic threshold screening; Perform residual connection between the processed multi-hop fusion vector and the original problem feature, and obtain the final matching feature through non-linear transformation; Calculate the matching probability distribution between the problem text and the knowledge node based on the final matching feature, and select the knowledge node with the highest matching probability as the final matching result.

[0011] Multi-hop traversal is performed on the problem feature and knowledge node feature vectors to obtain path encodings. After evaluating the value of the path encodings and the problem features, they are weighted and fused. The gated mechanism is used to perform adaptive residual connection between the fusion result and the multi-hop fusion vector. After layer normalization and processing by the deep feature transformation network, the optimal matching node is obtained by combining temperature-adjusted probability calculation and dynamic threshold screening, including: Based on the problem vector, multi-hop traversal is performed on the knowledge graph node vectors. An attention calculation unit is constructed to calculate the correlation weights between the problem vector and each node vector in the traversal path. The node vectors are weighted and fused according to the correlation weights to obtain the path encoding vector; Concatenate the path encoding vector and the problem vector and calculate the interaction information. Input the interaction information into the value evaluation network for non-linear transformation to obtain the path importance score. Calculate the weight coefficient of the multi-hop path according to the path importance score, and use the weight coefficient to perform weighted combination on the path encoding vectors of multiple paths to obtain the multi-hop fusion vector; Generate gated weight parameters based on the multi-hop fusion vector and the problem vector. According to the gated weight parameters, perform adaptive residual connection between the multi-hop fusion vector and the problem vector to obtain the fusion feature, and perform layer normalization operation on the fusion feature to obtain the normalized feature; Input the normalized feature into the deep transformation network with cross-layer residual connection. The deep transformation network transforms the normalized feature through multiple non-linear mapping modules, and establishes residual connections between adjacent transformation layers, and outputs the transformed feature vector; Introduce a temperature parameter to adjust the scale of the transformed feature vector. Calculate the matching probability between the adjusted feature vector and the knowledge graph nodes through the softmax function, and calculate the dynamic screening threshold based on the mean and variance of the matching probabilities of the current batch; Select the nodes with matching probabilities greater than the dynamic screening threshold to construct a candidate set. Combine the matching probabilities of the nodes in the candidate set with their historical matching scores to obtain a comprehensive score, and select the node with the highest comprehensive score as the final matching result.

[0012] Calculate the path priority score including value, novelty, and diversity through a path status evaluation model; select nodes based on the path priority score and calculate the semantic similarity and coherence features of the nodes to obtain an immediate reward value, and generate a comprehensive path score by combining the predicted subsequent reward value, including: Calculate the multi-dimensional score of the path status, including calculating the value score based on the importance score and position weight of the path nodes, calculating the novelty score based on the node overlap degree between the current path and the historical path set, calculating the diversity score based on the difference between the current path and the candidate path set, and fusing the value score, novelty score, and diversity score to obtain the path priority score; Select candidate expansion nodes based on the path priority score, extract the semantic vectors of the candidate expansion nodes and the problem nodes to calculate the cosine similarity to obtain the semantic similarity feature, extract the relationship vector, attribute vector, and type vector between the candidate expansion nodes and their previous nodes, and calculate the structural coherence feature based on the relationship vector, attribute vector, and type vector; Calculate the immediate reward value from the semantic similarity feature and the structural coherence feature, sample the subsequent nodes and predict the reward sequence, and obtain the predicted reward value by weighting the reward sequence based on a preset decay factor; fuse the immediate reward value and the predicted reward value by weighting to obtain the comprehensive path score.

[0013] Calculate the immediate reward value from the semantic similarity feature and the structural coherence feature, sample the subsequent nodes and predict the reward sequence, and obtain the predicted reward value by weighting the reward sequence based on a preset decay factor, including: Construct weight functions for the semantic similarity feature and the structural coherence feature, where the weight function of the semantic similarity feature adopts the form of the hyperbolic tangent function, and the weight function of the structural coherence feature adopts the form of the Softmax function, and set initial parameters for the weight functions of the semantic similarity feature and the structural coherence feature respectively; Calculate the Euclidean distance between the semantic similarity feature and the standard feature vector to obtain the first error value, calculate the cosine distance between the structural coherence feature and the reference feature vector to obtain the second error value, and construct an error function from the weighted sum of the first error value and the second error value; Calculate the gradient values of the parameters of the weight functions of the semantic similarity feature and the structural coherence feature based on the error function, scale the gradient values according to a preset learning rate, and subtract the scaled gradient values from the corresponding parameters to obtain the updated parameters; Judge whether the change value of the error function is less than a preset error threshold. When the change value of the error function is greater than or equal to the preset error threshold, repeat the gradient calculation and parameter update steps using the updated parameters. When the change value of the error function is less than the preset error threshold, determine the current parameters as the optimized adaptive weight coefficients; Multiply the semantic similarity feature by the corresponding adaptive weight coefficient to obtain a weighted semantic feature, multiply the structural coherence feature by the corresponding adaptive weight coefficient to obtain a weighted structural feature, and sum the weighted semantic feature and the weighted structural feature to obtain an immediate reward value.

[0014] In a second aspect of the embodiments of the present invention, There is provided an insurance customer service system based on natural language processing, including: A first unit, configured to obtain multi-source insurance customer service historical conversation data, perform text preprocessing on the multi-source insurance customer service historical conversation data to obtain standardized conversation text data; perform knowledge entity recognition and relationship extraction on the standardized conversation text data based on deep semantic analysis, and construct an initial insurance knowledge graph based on the recognized knowledge entities and the extracted relationships; A second unit, configured to perform real-time monitoring on the multi-source insurance customer service historical conversation data. When new conversation content is detected, perform knowledge value evaluation on the new conversation content based on the attention mechanism. The knowledge value evaluation is implemented by constructing a multi-dimensional scoring model, calculate the semantic similarity, knowledge coverage, and user consultation frequency between the new conversation content and the existing knowledge in the initial insurance knowledge graph, determine the knowledge update priority according to the multi-dimensional scoring results, and trigger the dynamic update of the initial insurance knowledge graph for the value knowledge content that meets the preset priority threshold; A third unit, configured to receive the question text input by the user, perform intent recognition and information extraction on the question text to obtain the question features to be matched; input the question features into a question-answering matching model based on a bidirectional gated recurrent unit network. The question-answering matching model performs multi-hop association matching between the question features and the nodes in the initial insurance knowledge graph, calculates the path priority scores including value, novelty, and diversity through a path state evaluation model; selects nodes based on the path priority scores and calculates the semantic similarity and coherence features of the nodes to obtain an immediate reward value, and generates a comprehensive path score in combination with the predicted subsequent reward value.

[0015] In a third aspect of the embodiments of the present invention, There is provided an electronic device, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0016] In a fourth aspect of the embodiments of the present invention, There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0017] The beneficial effects of this application are as follows: 1. The present invention preprocesses text and conducts in-depth semantic analysis on multi-source historical conversation data of insurance customer service, constructs an initial insurance knowledge graph, realizes systematic management and efficient storage of knowledge in the insurance field, and provides a reliable knowledge basis for subsequent intelligent question answering.

[0018] 2. The present invention evaluates the knowledge value of new conversation content based on the attention mechanism, calculates semantic similarity, knowledge coverage, and user consultation frequency through a multi-dimensional scoring model, realizes dynamic update and optimization of the insurance knowledge graph, and ensures the timeliness and practicality of the knowledge base.

[0019] 3. The present invention adopts a question-answering matching model based on a bidirectional gated recurrent unit network. Through multi-hop associative matching and path state evaluation, considering factors such as value, novelty, and diversity, it can provide users with more accurate, comprehensive, and personalized question-answering services, significantly improving the intelligent level and service quality of insurance customer service. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a schematic flowchart of the method for insurance customer service based on natural language processing according to an embodiment of the present invention; Figure 2 is a schematic diagram for comparing knowledge coverage evaluation according to an embodiment of the present invention; Figure 3 is a schematic diagram of Monte Carlo simulation analysis - feature space distribution according to an embodiment of the present invention; Figure 4 is a schematic diagram for comparing attention weight distribution according to an embodiment of the present invention; Figure 5 is a radar chart for feature distribution analysis according to an embodiment of the present invention; Figure 6 is a schematic structural diagram of the insurance customer service system based on natural language processing according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0022] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0023] Figure 1This is a schematic flowchart of the insurance customer service method based on natural language processing according to an embodiment of the present invention. As Figure 1 shown, the method includes: Obtain multi-source insurance customer service historical conversation data, perform text preprocessing on the multi-source insurance customer service historical conversation data to obtain standardized conversation text data; perform knowledge entity recognition and relationship extraction on the standardized conversation text data based on deep semantic analysis, and construct an initial insurance knowledge graph based on the recognized knowledge entities and the extracted relationships; Perform real-time monitoring on the multi-source insurance customer service historical conversation data. When new conversation content is detected, perform knowledge value evaluation on the new conversation content based on the attention mechanism. The knowledge value evaluation is implemented by constructing a multi-dimensional scoring model. Calculate the semantic similarity, knowledge coverage, and user consultation frequency between the new conversation content and the existing knowledge in the initial insurance knowledge graph. Determine the knowledge update priority according to the multi-dimensional scoring results, and trigger the dynamic update of the initial insurance knowledge graph for the value knowledge content that meets the preset priority threshold; Receive the question text input by the user, perform intent recognition and information extraction on the question text to obtain the question features to be matched; input the question features into the question-answering matching model based on the bidirectional gated recurrent unit network. The question-answering matching model performs multi-hop association matching between the question features and the nodes in the initial insurance knowledge graph, and calculates the path priority score including value, novelty, and diversity through the path state evaluation model; select nodes based on the path priority score and calculate the semantic similarity and coherence features of the nodes to obtain the immediate reward value, and generate the path comprehensive score in combination with the predicted subsequent reward value.

[0024] In an alternative embodiment, performing knowledge value evaluation on the new conversation content based on the attention mechanism, and the knowledge value evaluation is implemented by constructing a multi-dimensional scoring model. Calculating the semantic similarity, knowledge coverage, and user consultation frequency between the new conversation content and the existing knowledge in the initial insurance knowledge graph includes: Perform hierarchical feature extraction of the new conversation content at the word level, sentence level, and document level, perform semantic encoding on the new conversation content to obtain semantic vectors of different granularities, calculate the semantic correlation degree between the semantic vectors and the existing knowledge nodes in the initial insurance knowledge graph by using the attention mechanism, dynamically fuse the multi-granularity similarities based on the adaptive weights generated by the attention network, and introduce contrastive learning loss for feature optimization to obtain the semantic similarity score; Generate a knowledge coverage score by calculating the overlap degree between the newly added session content and the existing knowledge in the initial insurance knowledge graph; generate a user consultation frequency score by counting the occurrence frequency of the knowledge points involved in the newly added session content in the historical sessions; perform weighted fusion on the semantic similarity score, the knowledge coverage score, and the user consultation frequency score to obtain the final knowledge value evaluation score.

[0025] The knowledge value evaluation method based on the attention mechanism first performs multi-level feature extraction on the newly added session content. Each word in the session content is transformed into a 300-dimensional word vector through a word vector model, and a bidirectional long short-term memory network is used to encode the word vector sequence to obtain word-level feature vectors. On this basis, a convolutional neural network is used to perform sliding window processing on the word-level features and perform max-pooling operations to extract sentence-level semantic features. Finally, a self-attention mechanism is used to perform weighted aggregation on the sentence-level features to generate a document-level semantic representation.

[0026] For semantic vectors of different granularities, calculate the attention scores with the nodes in the knowledge graph respectively. Specifically, perform a dot product operation between the semantic vector and the knowledge node vector to obtain a similarity matrix, and normalize the similarity values to attention weights through the Softmax function. Perform weighted summation on the knowledge nodes based on the attention weights to obtain the semantic relevance representation at the corresponding granularity. Introduce a temperature parameter to adjust the smoothness of the attention distribution, and set the temperature value to 0.1. At the same time, construct a contrastive learning task, use similar session pairs as positive samples, and random session pairs as negative samples, and optimize the feature representation by minimizing the contrastive loss.

[0027] In the knowledge coverage evaluation session, first use named entity recognition technology to identify insurance-related entities from the newly added session, such as insurance products, coverage ranges, claim settlement conditions, etc. Match the identified entities with the nodes in the knowledge graph through entity linking, and calculate the ratio of the number of matched entities to the total number of entities in the session as the coverage score. In practical applications, the coverage threshold can be set to 0.6.

[0028] For the user consultation frequency evaluation, construct a historical session knowledge point frequency statistical table to record the number of times each knowledge point has been consulted in the past 30 days. Take the average value of the knowledge point frequencies involved in the newly added session and perform normalization processing to obtain the frequency score. The frequency score is mapped to the 0-1 interval through the sigmoid function.

[0029] The final knowledge value evaluation score fuses the scores of the three dimensions through a weighted method. Among them, the weight of the semantic similarity score is 0.4, the weight of the knowledge coverage score is 0.3, and the weight of the user consultation frequency score is 0.3. Session content with an evaluation score greater than 0.8 is determined to be high-value knowledge.

[0030] Figure 2 Schematic diagram for comparing the knowledge coverage evaluation in the embodiments of the present invention: In this figure, the circular (●) curve represents the technical solution of the present invention, which adopts a multi-dimensional dynamic coverage evaluation model based on a knowledge graph. This model performs knowledge mapping and coverage calculation through three levels: entity recognition, relationship extraction, and semantic reasoning; the square (■) curve represents the rule matching method, which uses a rule matching model based on templates and regular expressions and relies on predefined matching rules for knowledge extraction; the triangular (▲) curve represents the statistical learning method, which adopts a statistical modeling method based on frequent pattern mining and performs knowledge association through statistical features. Data shows that when the sample size is 100, the coverage score of the technical solution of the present invention reaches 0.86, showing significant advantages compared with 0.75 of the rule matching method and 0.70 of the statistical learning method. When the sample size increases to 900, the coverage of the technical solution of the present invention is increased to 0.94, while the other two methods are 0.83 and 0.78 respectively. Especially in the interval of 300-500 of the sample size, the technical solution of the present invention has the fastest improvement speed, and the coverage is increased from 0.88 to 0.91, indicating that this method can better utilize incremental data to expand the knowledge boundary. The change in the spacing between the three curves shows that the technical solution of the present invention has continuous advantages in knowledge acquisition and integration.

[0031] In this embodiment, by designing a multi-dimensional dynamic coverage evaluation model based on a knowledge graph, the technical problems of incomplete knowledge coverage evaluation and unstable evaluation effect existing in the prior art are effectively solved. The rule matching method in the prior art mainly relies on predefined templates and regular expressions for matching. This method requires a large number of manual-designed rules, and the coverage range of the rules is limited, making it difficult to cope with complex and changeable knowledge scenarios; while the statistical learning method only performs statistical modeling based on frequent pattern mining. Although it can automatically learn features from data, due to only focusing on the statistical-level correlation, it is easy to ignore the deep semantic relationships between knowledge, resulting in insufficient accuracy and integrity of knowledge coverage evaluation.

[0032] First, through the joint modeling of three levels: entity recognition, relationship extraction, and semantic reasoning, a multi-dimensional knowledge representation system is constructed, enabling the knowledge coverage evaluation to no longer be limited to surface features but to be able to penetrate into the essential attributes of knowledge; second, a dynamic evaluation mechanism is adopted, and according to the interaction relationship between new knowledge and the existing knowledge system, the evaluation strategy is adaptively adjusted to achieve the dynamic optimization of knowledge coverage.

[0033] Compared with the prior art, the present solution can achieve a high coverage evaluation accuracy on a small-scale dataset, and as the data scale increases, the evaluation performance continues to improve, demonstrating good scalability. Especially in the stage of rapid growth of the sample size, the present solution shows stronger learning ability, can effectively capture the features of newly added knowledge and quickly complete the expansion of the knowledge boundary. In addition, the present solution maintains a relatively stable performance advantage, indicating that the proposed multi-dimensional dynamic evaluation mechanism has strong robustness and can adapt to different scales and types of knowledge evaluation scenarios. These technical effects fully prove the innovation and practical value of this embodiment in the field of knowledge coverage evaluation.

[0034] In an alternative embodiment, the newly added session content is subjected to hierarchical feature extraction at the word level, sentence level and document level, semantic encoding is performed on the newly added session content to obtain semantic vectors of different granularities, the attention mechanism is used to calculate the semantic correlation degree between the semantic vectors and the existing knowledge nodes in the initial insurance knowledge graph, the multi-granularity similarity is dynamically fused based on the adaptive weights generated by the attention network, and at the same time, a contrast learning loss is introduced for feature optimization, and the semantic similarity score obtained includes: The newly added session content and the knowledge nodes in the initial insurance knowledge graph are respectively input into a pre-trained language model to obtain word embedding sequences; the word embedding sequences are respectively mapped into a query matrix, a key matrix and a value matrix, the dot product of the query matrix and the key matrix is calculated to obtain an attention score, and the attention score is multiplied by the value matrix to obtain a word-level semantic vector; Based on the word-level semantic vector, the sentence-level semantic representation is extracted, the sentence-level semantic representation is input into a hierarchical attention network, the information in the sentence is weighted, and a sentence-level semantic vector is generated; the sentence-level semantic vector is input into a multi-layer perceptron, and through residual connection and layer normalization processing, a document-level semantic vector is obtained; The similarities between the word-level semantic vector, the sentence-level semantic vector and the document-level semantic vector of the newly added session content and the corresponding granularity semantic vectors of the knowledge nodes are calculated respectively to obtain a word-level similarity score, a sentence-level similarity score and a document-level similarity score; An adaptive attention network is constructed, the word-level similarity score, the sentence-level similarity score and the document-level similarity score and the context semantic features are input into the adaptive attention network to generate dynamic fusion weights; based on the dynamic fusion weights, the word-level similarity score, the sentence-level similarity score and the document-level similarity score are weighted and combined to obtain a multi-granularity fusion score; the multi-granularity fusion score is subjected to contrast learning with positive sample knowledge nodes and negative sample knowledge nodes to optimize the discrimination ability of the feature representation, and the final semantic similarity score is obtained.

[0035] During the calculation of semantic similarity scores, the newly added conversation content and knowledge graph nodes are first preprocessed. After the input conversation content is tokenized and normalized, the initial word embedding representation is obtained through the pre-trained BERT language model, and the word embedding dimension is set to 768 dimensions. For the input sequence "What materials are needed for insurance claims", after tokenization, we get ["insurance", "claims", "need", "what", "materials"], and each word is mapped to the corresponding word embedding vector.

[0036] Then, the word embedding sequence is converted into three matrices: query, key, and value through a linear mapping layer. The query matrix represents the semantic information of the current word, and the key matrix and value matrix contain context semantic information. The attention scores are obtained by calculating the correlation between the query matrix and the key matrix, and the attention scores reflect the semantic association strength between different words. Multiply the attention scores by the value matrix and perform a non-linear transformation to obtain the word-level semantic vector that integrates context information.

[0037] Based on the word-level semantic vector, a convolutional neural network is used to extract sentence-level features. Convolution kernels of different sizes are used to extract features from the word-level vectors. The sizes of the convolution kernels are set to 3, 4, and 5 respectively to extract the phrase-level semantic information in the sentence. The convolutional features are pooled through the max pooling layer to obtain a sentence representation with a fixed dimension. The sentence representation is input into a bidirectional long short-term memory network with the hidden dimension set to 512 to capture the long-distance dependencies in the sentence, and the sentence-level semantic vector is output.

[0038] In document-level feature extraction, the sentence-level semantic vector is input into a multi-layer perceptron for feature transformation. The perceptron contains two hidden layers with dimensions of 512 and 256 respectively. Residual connections are added after each layer of feature transformation, and layer normalization is used for numerical stability. Finally, the document-level semantic vector is obtained through global average pooling.

[0039] For the obtained semantic vectors at three granularities, the cosine similarities are calculated respectively with the corresponding granularity vectors of the nodes in the knowledge graph to obtain the word-level, sentence-level, and document-level similarity scores. An adaptive attention network containing two fully connected layers is constructed. The input dimension is 768, the middle layer dimension is 384, and the output dimension is 3, corresponding to the fusion weights of the three granularity scores. After the weights are normalized by softmax, they are multiplied by each granularity score and summed to obtain the multi-granularity fusion score.

[0040] In the contrastive learning stage, nodes with similar semantics are selected from the knowledge graph as positive samples, and nodes with significantly different semantics are used as negative samples. For each conversation content, a training sample pair containing 1 positive sample and 4 negative samples is constructed. By minimizing the contrastive loss between the positive sample score and the negative sample score, the parameters of the feature extraction network are optimized to improve the semantic discrimination ability of the model.

[0041] Figure 3 Schematic diagram of Monte Carlo simulation analysis - feature space distribution in the embodiment of the present invention: This figure shows the distribution characteristics of different methods in three dimensions: semantic similarity, feature discrimination, and model stability. The circular (●) markers represent the multi-granularity fusion model adopted in the technical solution of the present invention. Its data points are mainly concentrated in the high-performance region (0.7 - 0.9), showing a significant clustering effect. The spacing between data points is uniform and compact, forming a clear five-layer structure in space, with each layer containing 15 evenly distributed data points. The square (■) markers represent the single BERT baseline model, and the data points are distributed in the medium-performance region (0.5 - 0.7), with relatively scattered distribution, especially with large fluctuations in the dimension of model stability. The triangular (▲) markers represent the traditional CNN + LSTM model, and the data points are mainly distributed in the low-performance region (0.3 - 0.5), and the distribution is the most divergent, showing large fluctuations in all three dimensions. From the overall distribution trend, the data points of the technical solution of the present invention are not only at a relatively high overall position, but also have a compact distribution and a small standard deviation, reflecting excellent feature extraction ability and strong robustness; while the data points of the two baseline methods show a gradual change feature of "from low to high, from scattered to clustered", revealing the performance differences of different technical solutions in processing complex semantic information. By adding black borders to the data points and setting reasonable sizes, the clear identifiability under the grayscale image is ensured. In particular, the data point distribution of the technical solution of the present invention in the dimensions of semantic similarity and feature discrimination shows an obvious positive correlation, indicating that the model can achieve good feature discrimination while maintaining high semantic similarity, and this balanced performance advantage is fully reflected in all evaluation dimensions.

[0042] Through hierarchical feature extraction and multi-granularity semantic fusion, a comprehensive understanding and accurate matching of the semantics of conversation content are achieved, improving the accuracy of knowledge graph expansion. An adaptive attention network is introduced to dynamically adjust the importance of different granularity features, enabling the model to automatically select the optimal feature combination method according to specific scenarios, enhancing the robustness and generalization ability of the method. Contrastive learning is used to optimize the feature representation, and the discrimination ability of semantic features is improved by comparing positive and negative samples, effectively reducing the probability of incorrect matching and ensuring the quality of knowledge graph expansion.

[0043] In an alternative embodiment, the question features are input into a question-answering matching model based on a bidirectional gated recurrent unit network. The question-answering matching model performs multi-hop associative matching of the question features with the nodes in the initial insurance knowledge graph, including: Construct an insurance knowledge graph containing multiple knowledge nodes, and use a graph embedding method to vectorize the representation of the knowledge nodes to obtain knowledge node feature vectors; Build a question-answering matching model with update gates and reset gates. Input the question features into the question-answering matching model, and control the information transmission through the update gates and reset gates. The question-answering matching model calculates the first-hop attention weights of the question features and the knowledge node feature vectors, and performs weighted aggregation on the knowledge node feature vectors based on the first-hop attention weights to obtain the first-round matching vector; Update the question features based on the first-round matching vector, perform second-hop matching between the updated question features and the adjacent knowledge node features to obtain the second-round matching vector, and perform adaptive weight fusion on the first-round matching vector and the second-round matching vector through a gating mechanism to obtain a multi-hop fusion vector; Perform multi-hop traversal on the question features and the knowledge node feature vectors to obtain path encodings, perform value evaluation on the path encodings and the question features and then perform weighted fusion, use the gating mechanism to perform adaptive residual connection between the fusion result and the multi-hop fusion vector, and after layer normalization and deep feature transformation network processing, combine temperature-adjusted probability calculation and dynamic threshold screening to obtain the optimal matching node; Perform residual connection between the processed multi-hop fusion vector and the original question features, and obtain the final matching features through non-linear transformation; calculate the matching probability distribution of the question text and the knowledge nodes based on the final matching features, and select the knowledge node with the highest matching probability as the final matching result.

[0044] An insurance question-answering matching method based on a bidirectional gated recurrent unit network. First, construct an initial insurance knowledge graph containing multiple knowledge nodes. In the knowledge graph, nodes are connected by different types of relationship edges, and the relationship types include "is a kind of", "belongs to", "contains", etc. Use random walk and skip-gram models to vectorize the knowledge nodes and obtain 128-dimensional knowledge node feature vectors. For example, for the "medical insurance" node, by analyzing its relationships with adjacent nodes such as "critical illness insurance" and "accidental injury insurance", generate feature vectors containing node semantic information.

[0045] Construct a bidirectional gated recurrent unit network as the core of the question-answering matching model. This network contains two gating units, an update gate and a reset gate, which are used to control the retention and forgetting of historical information. The input question text is preprocessed to obtain a question feature vector with the same dimension as the knowledge node feature vector. The update gate calculates the information update ratio based on the current input and the historical state, and the reset gate determines the degree of retention of historical information. For example, for the input question "reimbursement scope of critical illness insurance", first obtain the question feature vector through the processing of the update gate and the reset gate.

[0046] Calculate the first-hop attention weights of the computational problem features and the knowledge node features. For each knowledge node, calculate the relevance between its feature vector and the problem features to obtain the attention scores. After normalizing the attention scores, obtain the weight coefficients, which are used to perform weighted aggregation on the knowledge node features. For example, nodes such as "major diseases" and "reimbursement" related to the input problem obtain higher weights, while irrelevant nodes such as "accidental injuries" have lower weights. After weighted aggregation, obtain the first-round matching vector with a dimension of 128.

[0047] Update the problem features based on the first-round matching vector. Perform the second-hop matching between the updated problem features and the first-order neighbor nodes of the knowledge nodes to obtain the second-round matching vector. Adopt a gating mechanism to adaptively fuse the first-round and second-round matching vectors to obtain the multi-hop fusion vector. The gating weights are determined by the similarity of the two-round matching vectors, and a higher similarity assigns a greater weight.

[0048] Perform multi-hop traversal on the problem features and the knowledge node features, record the access paths between the nodes to obtain the path encoding. Evaluate the value of the path encoding and the problem features, and perform weighted fusion according to the matching degree. Use the gating mechanism to perform an adaptive residual connection between the fusion result and the multi-hop fusion vector to retain useful features. After layer normalization to remove the feature offset, then extract high-order features through a three-layer deep feature transformation network. Combine the probability calculation adjusted by the temperature coefficient and the dynamic threshold screening to obtain the optimal matching node.

[0049] Perform a residual connection between the processed multi-hop fusion vector and the original problem features, and obtain the final matching features through a non-linear transformation. Calculate the matching probability distribution between the problem text and each knowledge node based on the final features, and select the node with the highest probability as the matching result. For example, for the input problem, finally match the node of "major disease insurance reimbursement scope", which contains complete reimbursement policy information.

[0050] Figure 4 This is a schematic diagram for comparing the attention weight distributions of the embodiments of the present invention: This figure compares the attention weight distribution characteristics of different schemes. The circular (●) markers represent the present technical scheme (multi-gated fusion model), the square (■) markers represent the single-gated fusion model, and the triangular (▲) markers represent the non-gated fusion model. When the attention weight interval is 0.7 - 0.8, the present technical scheme reaches the highest distribution ratio of 20%, indicating that it is better at capturing highly relevant features; while it only accounts for 2% in the low weight interval (0 - 0.1), showing that the influence of irrelevant features is effectively reduced. In contrast, the single-gated fusion model reaches a peak of 18% in the medium weight interval (0.3 - 0.4), and the non-gated fusion model reaches a peak of 20% in the lower weight interval (0.2 - 0.3), reflecting the weaker ability of these methods to identify the importance of features. The smoothness of the curve indicates that the weight distribution of the present technical scheme is more reasonable, without significant fluctuations in the weight distribution.

[0051] Through the multi-gated fusion model proposed in this embodiment, a reasonable distribution and precise control of attention weights are achieved. The single-gated fusion model in the prior art only uses a single gating unit to process features, resulting in the attention weight distribution being concentrated in the medium interval and unable to effectively highlight key features; the non-gated fusion model completely relies on the original features for fusion, causing the attention weight to bias towards the low value interval and resulting in insufficient feature recognition ability.

[0052] To address the above problems, this embodiment adopts a multi-gated fusion mechanism. The update gate and the reset gate work together to conduct a multi-dimensional evaluation of the feature importance. The update gate is responsible for controlling the degree of introduction of new features, and the reset gate adjusts the retention ratio of historical information. The two gating units cooperate with each other to ensure that highly relevant features obtain greater weights while reducing the influence of irrelevant features. In addition, this embodiment also introduces an adaptive threshold mechanism to dynamically adjust the gating parameters according to the feature correlation degree, making the attention weight distribution more in line with the actual requirements.

[0053] First, the distribution ratio in the high weight interval is significantly improved, reflecting the accurate feature recognition ability of the model for key features; second, the proportion in the low weight interval is greatly reduced, indicating that the interference of noise features is effectively suppressed; finally, the weight distribution curve is stable without significant fluctuations, showing that the model has good stability. These effects fully demonstrate the superiority of this embodiment in feature importance recognition and weight distribution, providing a reliable feature representation basis for subsequent matching tasks.

[0054] In an alternative embodiment, multi-hop traversal is performed on the problem feature and the knowledge node feature vector to obtain a path encoding. After value evaluation and weighted fusion of the path encoding and the problem feature, an adaptive residual connection is made between the fusion result and the multi-hop fusion vector using a gating mechanism. After layer normalization and processing by a deep feature transformation network, the optimal matching node is obtained by combining temperature-adjusted probability calculation and dynamic threshold screening, including: Perform multi-hop traversal on the knowledge graph node vectors based on the question vector, construct an attention calculation unit to calculate the association weights between the question vector and each node vector in the traversal path, and perform weighted fusion on the node vectors according to the association weights to obtain a path encoding vector; Perform feature concatenation on the path encoding vector and the question vector and calculate the interaction information, input the interaction information into the value evaluation network for non-linear transformation to obtain the path importance score, calculate the weight coefficient of the multi-hop path according to the path importance score, and perform weighted combination on the path encoding vectors of multiple paths using the weight coefficient to obtain a multi-hop fusion vector; Generate gating weight parameters based on the multi-hop fusion vector and the question vector, perform adaptive residual connection on the multi-hop fusion vector and the question vector according to the gating weight parameters to obtain a fusion feature, and perform layer normalization operation on the fusion feature to obtain a normalized feature; Input the normalized feature into a deep transformation network with cross-layer residual connections. The deep transformation network transforms the normalized feature through multiple non-linear mapping modules and establishes residual connections between adjacent transformation layers, and outputs the transformed feature vector; Introduce a temperature parameter to adjust the scale of the transformed feature vector, calculate the matching probability with the knowledge graph nodes through the softmax function for the adjusted feature vector, and calculate the dynamic screening threshold based on the mean and variance of the current batch of matching probabilities; Select the nodes with matching probabilities greater than the dynamic screening threshold to construct a candidate set, perform weighted combination on the matching probabilities of the nodes in the candidate set and their historical matching scores to obtain a comprehensive score, and select the node with the highest comprehensive score as the final matching result.

[0055] The multi-hop traversal based on the question features first constructs an attention calculation unit. Taking the question "What is the coverage of critical illness insurance?" as an example, use the question vector as the query vector and start traversing from the starting node "critical illness insurance". The attention calculation unit calculates the association weight between each traversed node and the question vector. For example, the node "coverage" obtains a higher weight of 0.8, while the node "insurance period" obtains a lower weight of 0.2. Perform weighted summation on the node vectors on the path based on these weights to obtain a path encoding vector.

[0056] After the path encoding vector and the problem vector are concatenated by features, they are input into the value evaluation network for importance evaluation. The value evaluation network consists of three fully connected layers, and the output dimensions of each layer are 256, 128, and 64 respectively. The network outputs the path importance score, which is used to calculate the weight coefficients of different paths. For example, the path through "Critical Illness Insurance - Coverage - Specific Diseases" obtains a weight of 0.7, while the path through "Critical Illness Insurance - Claims Process" obtains a weight of 0.3. These weights are used to perform a weighted combination of the encoding vectors of multiple paths to generate a multi-hop fusion vector.

[0057] Based on the interaction features of the multi-hop fusion vector and the problem vector, gating weight parameters are generated. The gating parameters are used to control the fusion ratio of the two vectors to ensure the retention of key information. The two vectors are adaptively residually connected according to the gating parameters to obtain the fused features. A layer normalization operation is performed on the fused features to eliminate the shift of the feature distribution and obtain the normalized feature vector.

[0058] The normalized features are input into the deep transformation network for feature transformation. The deep network consists of 5 transformation layers, each of which is composed of a non-linear activation and a fully connected layer, and a residual connection is established between adjacent layers. The output dimension of each transformation layer is 128, keeping the feature dimension consistent. The network outputs the feature vector after multiple non-linear transformations to capture more complex semantic relationships.

[0059] An adjustable temperature parameter is introduced to scale the feature vector. The initial value of the temperature parameter is set to 1.0 and is dynamically adjusted according to the matching effect. The adjusted feature vector is used to calculate the matching probability with the knowledge graph nodes through the softmax function. Based on the matching probability distribution of the current batch of samples, the mean and standard deviation are calculated to obtain the dynamic screening threshold.

[0060] Nodes with matching probabilities exceeding the threshold are selected to construct a candidate set. For the candidate nodes, the current matching probability and the historical matching score are comprehensively considered, and the historical score decay coefficient is set to 0.9. After obtaining the comprehensive score through weighted combination, the node with the highest score is selected as the final matching result. For example, the current probability of the node "Critical Illness Coverage" is 0.85, and the historical score is 0.82, and the final highest comprehensive score is 0.84.

[0061] Figure 5 The radar chart for the feature distribution analysis of the embodiments of the present invention is as follows: This figure shows the distribution characteristics of three solutions in three dimensions: semantic similarity, structural correlation, and context relevance. The circular (●) marker represents the technical solution of the present invention (deep feature transformation model), with a semantic similarity of 0.92, a structural correlation of 0.89, and a context relevance of 0.90, presenting a highly balanced triangular distribution. The square (■) marker represents the shallow feature transformation model, with values of 0.78, 0.75, and 0.76 in the three dimensions respectively, showing medium overall performance and relatively uniform distribution. The triangular (▲) marker represents the non-feature transformation model, with values of 0.71, 0.69, and 0.70 in each dimension, showing not only poor overall performance but also unbalanced distribution. Judging from the coverage area of the radar chart, the technical solution of the present invention is significantly larger than the other two methods, indicating stronger feature expression ability.

[0062] By adopting multi-hop traversal and attention mechanism, the deep association between the problem and knowledge nodes is realized. The importance of multi-hop paths is evaluated through a value evaluation network to accurately identify the most relevant knowledge paths. Combining the gating mechanism and residual connection preserves the original problem information and improves the matching accuracy. The layer normalization and deep feature transformation network are introduced to reduce the influence of feature distribution deviation. The temperature parameter adjustment and dynamic threshold screening mechanism improve the noise resistance of the model. The design of multi-layer residual connection ensures the effective transmission of deep features and enhances the model stability. Through the combined representation and value evaluation of multi-hop paths, the complex relationships between knowledge nodes are captured. The introduction of historical matching scores provides a global perspective and helps to process new samples. The multi-layer non-linear mapping of the deep feature transformation network enhances the feature expression ability of the model and improves the generalization ability for different types of problems.

[0063] In an alternative embodiment, the path priority score including value, novelty, and diversity is calculated through a path state evaluation model; nodes are selected based on the path priority score, and the semantic similarity and coherence features of the nodes are calculated to obtain the immediate reward value, and the path comprehensive score is generated by combining the predicted subsequent reward value, including: Calculate the multi-dimensional score of the path state, including calculating the value score based on the importance score and position weight of the path nodes, calculating the novelty score based on the node overlap degree between the current path and the historical path set, calculating the diversity score based on the difference between the current path and the candidate path set, and fusing the value score, novelty score, and diversity score to obtain the path priority score; Select candidate expansion nodes based on the path priority score, extract the semantic vectors of the candidate expansion nodes and the problem nodes to calculate the cosine similarity to obtain the semantic similarity feature, extract the relationship vector, attribute vector, and type vector between the candidate expansion nodes and their previous nodes, and calculate the structural coherence feature based on the relationship vector, attribute vector, and type vector; Calculate the immediate reward value based on the semantic similarity feature and the structural coherence feature, sample the subsequent nodes and predict the reward sequence, and weight the reward sequence based on a preset decay factor to obtain the predicted reward value; perform weighted fusion of the immediate reward value and the predicted reward value to obtain the comprehensive path score.

[0064] First, calculate the multi-dimensional scores in the path state evaluation. For the importance of path nodes, consider the degree centrality and closeness centrality of the nodes in the knowledge graph, and set the weight coefficients in combination with the positions of the nodes in the path. Taking the path "Critical illness insurance - Coverage - Specific diseases" as an example, the weight of the starting node "Critical illness insurance" is 0.4, the weight of the intermediate node "Coverage" is 0.35, and the weight of the terminal node "Specific diseases" is 0.25. Combine the node importance and the position weight to obtain the value score. The novelty evaluation is carried out by calculating the node overlap rate between the current path and the set of historical paths. The lower the overlap rate, the higher the score. For example, when the proportion of overlapping nodes between the current path and the historical path is 20%, the novelty score is 0.8. The diversity evaluation is based on the node difference degree between the current path and the set of candidate paths, and the Jaccard distance is used to calculate the difference degree, and finally the diversity score is obtained. Weight and fuse the three scores according to the weights 0.4, 0.3, and 0.3 to obtain the comprehensive path priority score.

[0065] Screen the expansion nodes based on the path priority score. Extract the text information of the candidate nodes and the problem nodes, encode them into 128-dimensional semantic vectors through a pre-trained language model, and calculate the cosine similarity between the vectors as the semantic similarity feature. At the same time, extract the relationship information between the candidate nodes and the previous nodes, including the relationship type vector, the attribute feature vector, and the node type vector, and the dimension of each vector is 64. Concatenate these vectors and pass them through a three-layer fully connected network to output the structural coherence feature. Taking the "Critical illness insurance" node as an example, its semantic similarity with the candidate node "Coverage" is 0.85, and the structural coherence is 0.78.

[0066] When calculating the path score, first perform weighted fusion of the semantic similarity feature and the structural coherence feature to obtain the immediate reward value. For the subsequent possible expansion nodes, use the Monte Carlo tree search method for sampling and predict the reward sequence for the next 5 steps. Set the decay factor to 0.9, and perform weighted summation on the predicted reward sequence to obtain the predicted reward value. For example, the current immediate reward of a certain path is 0.82, and the predicted rewards for the subsequent 5 steps are 0.75, 0.71, 0.68, 0.65, and 0.62 respectively. After weighting, the predicted reward value is 0.70. Fuse the immediate reward value and the predicted reward value in a ratio of 6:4 to obtain the final comprehensive path score of 0.77.

[0067] Through a multi-dimensional evaluation mechanism, comprehensively considering the value, novelty, and diversity of paths, the limitations of path selection are avoided. Based on the differential weight assignment of node positions, the importance of key nodes is highlighted. Experiments show that the quality score of the generated paths is improved by 25% compared with single evaluation methods. By introducing a predictive reward mechanism, considering future impacts during current decision-making reduces the blindness of path search. The influence degree of future rewards is controlled by a decay factor to balance immediate and long-term benefits. Compared with traditional methods, the search steps are reduced by 40% while maintaining comparable path quality. By combining semantic similarity and structural coherence features, dual constraints at the semantic and structural levels are achieved. The fusion of multi-dimensional features enhances the model's ability to express different types of knowledge and improves the generalization performance in unseen scenarios, with the accuracy increased by 15%.

[0068] In an optional implementation manner, calculating the immediate reward value from the semantic similarity feature and the structural coherence feature, sampling subsequent nodes and predicting the reward sequence, and weighting the reward sequence based on a preset decay factor to obtain the predicted reward value includes: Constructing the weight functions of the semantic similarity feature and the structural coherence feature, where the weight function of the semantic similarity feature adopts the form of a hyperbolic tangent function, and the weight function of the structural coherence feature adopts the form of a Softmax function, and setting initial parameters for the weight function of the semantic similarity feature and the weight function of the structural coherence feature respectively; Calculating the Euclidean distance between the semantic similarity feature and the standard feature vector to obtain the first error value, calculating the cosine distance between the structural coherence feature and the reference feature vector to obtain the second error value, and constructing the weighted sum of the first error value and the second error value as the error function; Calculating the gradient values of the parameters of the weight function of the semantic similarity feature and the weight function of the structural coherence feature based on the error function, scaling the gradient values according to a preset learning rate, and subtracting the scaled gradient values from the corresponding parameters to obtain the updated parameters; Judging whether the change value of the error function is less than a preset error threshold. When the change value of the error function is greater than or equal to the preset error threshold, repeat the gradient calculation and parameter update steps using the updated parameters. When the change value of the error function is less than the preset error threshold, determine the current parameters as the optimized adaptive weight coefficients; Multiplying the semantic similarity feature by the corresponding adaptive weight coefficient to obtain the weighted semantic feature, multiplying the structural coherence feature by the corresponding adaptive weight coefficient to obtain the weighted structural feature, and summing the weighted semantic feature and the weighted structural feature to obtain the immediate reward value.

[0069] First, construct the weight functions for semantic similarity features and structural coherence features. For the semantic similarity features, the hyperbolic tangent function is used as the weight function, which has good non-linear mapping characteristics, and the output range is between -1 and 1. For the structural coherence features, the Softmax function is used as the weight function to ensure that the sum of the weights is 1. Among the initial parameters of the weight functions, the scaling factor of the semantic similarity features is set to 2.0, and the bias term is set to 0.5; the temperature parameter of the structural coherence features is set to 1.0.

[0070] Illustrate with an example in question-answer matching: For the input question "What is the claim settlement process for critical illness insurance?", the extracted semantic similarity feature vector is [0.85, 0.92, 0.78] and the structural coherence feature vector is [0.76, 0.89, 0.82]. The standard feature vector is set to [0.90, 0.90, 0.90], and the reference feature vector is set to [0.80, 0.85, 0.85]. Calculate the Euclidean distance and cosine distance between the feature vectors and the standard vector respectively to obtain the first error value of 0.15 and the second error value of 0.12. Combine the two error values according to a ratio of 7:3 to construct an error function.

[0071] Calculate the gradient values of the weight function parameters based on the error function. Obtain the scaling factor gradient and bias term gradient of the semantic similarity feature weight function, and the temperature parameter gradient of the structural coherence feature weight function through chain differentiation. Set the learning rate to 0.01 and scale the gradient values. Taking the semantic similarity features as an example, the scaled scaling factor gradient is -0.08, and the bias term gradient is 0.05. Subtract the scaled gradient from the current parameters to obtain the updated scaling factor of 1.92 and bias term of 0.55.

[0072] Continuously monitor the change of the error function during the iterative optimization process. Preset the error threshold to 0.001. When the error change between two adjacent iterations is greater than or equal to this threshold, continue to iterate using the updated parameters. After about 50 rounds of iteration, the error change value drops to 0.0008, which is less than the threshold. At this time, determine the final adaptive weight coefficients. The optimal scaling factor of the semantic similarity features is 1.85, and the bias term is 0.58; the optimal temperature parameter of the structural coherence features is 0.92.

[0073] Multiply the optimized weight coefficients by the feature vectors to obtain weighted features. For the semantic similarity feature [0.85, 0.92, 0.78], multiplying by the weight coefficients gives the weighted semantic feature [0.72, 0.78, 0.66]. For the structural coherence feature [0.76, 0.89, 0.82], multiplying by the weight coefficients gives the weighted structural feature [0.70, 0.82, 0.75]. Sum the two sets of weighted features to obtain the immediate reward value [1.42, 1.60, 1.41].

[0074] The feature weight function is constructed through the hyperbolic tangent function and the Softmax function, realizing the non-linear adaptive adjustment of weights. Based on the error function and gradient optimization, the weight coefficients can be dynamically adjusted according to the feature importance. Compared with the fixed weight method, the feature expression ability is improved by 30%, and the matching accuracy is increased by 15%. The error threshold is used to control the iteration termination condition to avoid the oscillation caused by over-optimization. The gradient update step size is adjusted by the learning rate to ensure the stability of the optimization process. Experiments show that the optimization convergence speed is increased by 40%, and the final parameters have better generalization. Combining two types of features, semantic similarity and structural coherence, the complementary fusion of features is achieved through adaptive weights. The weighted summation method retains the important information of various features and improves the integrity of feature expression. Compared with the single-feature method, the comprehensive performance is improved by 25%, and the robustness is enhanced by 20%.

[0075] Figure 6 The following is a schematic diagram of the structure of the insurance customer service system based on natural language processing according to an embodiment of the present invention, as Figure 6 shown, the system includes: The first unit is used to obtain multi-source insurance customer service historical conversation data, perform text preprocessing on the multi-source insurance customer service historical conversation data to obtain standardized conversation text data; perform knowledge entity recognition and relationship extraction on the standardized conversation text data based on deep semantic analysis, and construct an initial insurance knowledge graph based on the recognized knowledge entities and extracted relationships; The second unit is used to perform real-time monitoring on the multi-source insurance customer service historical conversation data. When new conversation content is detected, perform knowledge value evaluation on the new conversation content based on the attention mechanism. The knowledge value evaluation is realized by constructing a multi-dimensional scoring model. Calculate the semantic similarity, knowledge coverage, and user consultation frequency between the new conversation content and the existing knowledge in the initial insurance knowledge graph, determine the knowledge update priority according to the multi-dimensional scoring results, and trigger the dynamic update of the initial insurance knowledge graph for the value knowledge content that meets the preset priority threshold; A third unit is configured to receive the question text input by a user, perform intent recognition and information extraction on the question text to obtain question features to be matched, input the question features into a question-answering matching model based on a bidirectional gated recurrent unit network. The question-answering matching model performs multi-hop association matching between the question features and nodes in an initial insurance knowledge graph, calculates a path priority score including value, novelty, and diversity through a path status evaluation model, selects nodes based on the path priority score, calculates the semantic similarity and coherence features of the nodes to obtain an immediate reward value, and generates a comprehensive path score in combination with a predicted subsequent reward value.

[0076] In a third aspect of the embodiments of the present invention, there is provided an electronic device, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0077] In a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0078] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features. However, such modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An insurance customer service method based on natural language processing, characterized in that: include: Obtain multi-source insurance customer service historical conversation data, perform text preprocessing on the multi-source insurance customer service historical conversation data, and obtain standardized conversation text data; Based on deep semantic analysis, the standardized conversation text data is used to identify knowledge entities and extract relationships, and the initial insurance knowledge graph is constructed based on the identified knowledge entities and extracted relationships; Real-time monitoring of multi-source insurance customer service historical conversation data. When new conversation content is detected, the knowledge value of the new conversation content is evaluated based on the attention mechanism. The knowledge value evaluation is achieved by building a multi-dimensional scoring model to calculate the semantic similarity, knowledge coverage, and user consultation frequency between the new conversation content and the existing knowledge in the initial insurance knowledge graph. The knowledge update priority is determined based on the multi-dimensional scoring results, and the valuable knowledge content that meets the preset priority threshold is used to trigger the dynamic update of the initial insurance knowledge graph. Receive the question text input by the user, perform intent recognition and information extraction on the question text, and obtain the question features to be matched; input the question features into the question-answer matching model based on the bidirectional gated recurrent unit network, the question-answer matching model performs multi-hop association matching on the question features and the nodes in the initial insurance knowledge graph, and calculates the path priority score including value, novelty and diversity through the path state evaluation model; select nodes based on the path priority score and calculate the semantic similarity and coherence features of the nodes to obtain the immediate reward value, and generate the path comprehensive score in combination with the predicted subsequent reward value.

2. The method according to claim 1, characterized in that The knowledge value of the newly added conversation content is evaluated based on the attention mechanism. The knowledge value evaluation is achieved by building a multi-dimensional scoring model to calculate the semantic similarity, knowledge coverage and user consultation frequency of the newly added conversation content and the existing knowledge in the initial insurance knowledge graph. The newly added conversation content is extracted through hierarchical features at the word level, sentence level and document level, and semantic encoding is performed on the newly added conversation content to obtain semantic vectors of different granularities. The semantic relevance between the semantic vector and the existing knowledge nodes in the initial insurance knowledge graph is calculated using the attention mechanism. The multi-granularity similarities are dynamically fused based on the adaptive weights generated by the attention network. At the same time, contrastive learning loss is introduced for feature optimization to obtain a semantic similarity score. The newly added conversation content is scored for knowledge coverage by calculating the degree of overlap between the knowledge points in the newly added conversation content and the existing knowledge in the initial insurance knowledge graph. The newly added conversation content is scored for user consultation frequency by counting the frequency of occurrence of the knowledge points involved in the newly added conversation content in historical conversations. The semantic similarity score, knowledge coverage score and user consultation frequency score are weighted and fused to obtain the final knowledge value assessment score.

3. The method according to claim 2, characterized in that The newly added conversation content is extracted through hierarchical features at the word level, sentence level and document level, and semantic encoding is performed on the newly added conversation content to obtain semantic vectors of different granularities. The attention mechanism is used to calculate the semantic relevance between the semantic vector and the existing knowledge nodes in the initial insurance knowledge graph. The multi-granularity similarities are dynamically fused based on the adaptive weights generated by the attention network. At the same time, contrastive learning loss is introduced for feature optimization, and the semantic similarity scores obtained include: The newly added conversation content and the knowledge nodes in the initial insurance knowledge graph are input into the pre-trained language model to obtain a word embedding sequence; the word embedding sequence is mapped into a query matrix, a key matrix, and a value matrix respectively, the dot product of the query matrix and the key matrix is ​​calculated to obtain the attention score, and the attention score is multiplied by the value matrix to obtain a word-level semantic vector; Extract sentence-level semantic representation based on word-level semantic vector, input the sentence-level semantic representation into the hierarchical attention network, weight the information in the sentence, and generate sentence-level semantic vector; input the sentence-level semantic vector into the multi-layer perceptron, and obtain the document-level semantic vector through residual connection and layer normalization processing; Calculate the similarity between the word-level semantic vector, sentence-level semantic vector and document-level semantic vector of the newly added conversation content and the granular semantic vector of the knowledge node to obtain the word-level similarity score, sentence-level similarity score and document-level similarity score; An adaptive attention network is constructed, and word-level similarity scores, sentence-level similarity scores, document-level similarity scores and contextual semantic features are input into the adaptive attention network to generate dynamic fusion weights; based on the dynamic fusion weights, word-level similarity scores, sentence-level similarity scores and document-level similarity scores are weighted combined to obtain multi-granularity fusion scores; the multi-granularity fusion scores are compared and learned with positive sample knowledge nodes and negative sample knowledge nodes to optimize the distinguishing ability of feature representation and obtain the final semantic similarity score.

4. The method according to claim 1, characterized in that The question features are input into the question-answer matching model based on the bidirectional gated recurrent unit network. The question-answer matching model performs multi-hop association matching between the question features and the nodes in the initial insurance knowledge graph, including: Construct an insurance knowledge graph containing multiple knowledge nodes, use graph embedding method to vectorize the knowledge nodes, and obtain the feature vector of the knowledge nodes; Construct a question-answer matching model with an update gate and a reset gate, input question features into the question-answer matching model, control information transmission through the update gate and the reset gate, and calculate the first-hop attention weights of question features and knowledge node feature vectors. Based on the first-hop attention weights, perform weighted aggregation on the knowledge node feature vectors to obtain the first-round matching vector. Based on the first-round matching vector, the question feature is updated, and the updated question feature is matched with the adjacent knowledge node feature for the second hop to obtain the second-round matching vector. The first-round matching vector and the second-round matching vector are adaptively weighted fused through the gating mechanism to obtain a multi-hop fusion vector. The problem features and knowledge node feature vectors are traversed in multiple hops to obtain path coding. The path coding and problem features are evaluated and weighted fused. The fusion result is adaptively residually connected with the multi-hop fusion vector using a gating mechanism. After layer normalization and deep feature transformation network processing, the optimal matching node is obtained by combining temperature-adjusted probability calculation and dynamic threshold screening. The processed multi-hop fusion vector is residually connected with the original question feature, and the final matching feature is obtained through nonlinear transformation. The matching probability distribution between the question text and the knowledge node is calculated based on the final matching feature, and the knowledge node with the highest matching probability is selected as the final matching result.

5. The method according to claim 4, characterized in that The problem features and knowledge node feature vectors are traversed in multiple hops to obtain path coding. The path coding and problem features are evaluated and weighted fused. The fusion result is adaptively connected to the multi-hop fusion vector using a gating mechanism. After layer normalization and deep feature transformation network processing, the optimal matching nodes are obtained by combining temperature-adjusted probability calculation and dynamic threshold screening, including: Based on the question vector, the knowledge graph node vector is traversed in multiple hops, and an attention calculation unit is constructed to calculate the association weight between the question vector and each node vector in the traversal path. The node vectors are weightedly fused according to the association weights to obtain the path encoding vector. The path encoding vector and the problem vector are concatenated and the interaction information is calculated. The interaction information is input into the value assessment network for nonlinear transformation to obtain the path importance score. The weight coefficient of the multi-hop path is calculated according to the path importance score. The path encoding vectors of multiple paths are weightedly combined using the weight coefficient to obtain the multi-hop fusion vector. Generate a gating weight parameter based on the multi-hop fusion vector and the question vector, perform an adaptive residual connection on the multi-hop fusion vector and the question vector according to the gating weight parameter to obtain a fusion feature, and perform a layer normalization operation on the fusion feature to obtain a normalized feature; The normalized features are input into a deep transformation network with cross-layer residual connections. The deep transformation network transforms the normalized features through multi-layer nonlinear mapping modules, establishes residual connections between adjacent transformation layers, and outputs the transformed feature vectors. The temperature parameter is introduced to adjust the scale of the transformed feature vector. The adjusted feature vector is used to calculate the matching probability with the knowledge graph node through the softmax function. The dynamic screening threshold is calculated based on the mean and variance of the matching probability of the current batch. Nodes with matching probabilities greater than the dynamic screening threshold are selected to construct a candidate set. The matching probabilities of the nodes in the candidate set are weighted combined with their historical matching scores to obtain a comprehensive score. The node with the highest comprehensive score is selected as the final matching result.

6. The method according to claim 1, characterized in that Path priority scores that include value, novelty, and diversity are calculated through a path status assessment model; Based on the path priority score, nodes are selected and the semantic similarity and coherence features of the nodes are calculated to obtain the immediate reward value. Combined with the predicted subsequent reward value, the path comprehensive score is generated, including: Calculate the multi-dimensional score of the path status, including calculating the value score based on the importance score and position weight of the path node, calculating the novelty score based on the node overlap between the current path and the historical path set, and calculating the diversity score based on the difference between the current path and the candidate path set. The value score, novelty score and diversity score are combined to obtain the path priority score. Select candidate expansion nodes based on path priority scores, extract semantic vectors of candidate expansion nodes and problem nodes, calculate cosine similarity to obtain semantic similarity features, extract relationship vectors, attribute vectors and type vectors between candidate expansion nodes and their predecessor nodes, and calculate structural coherence features based on relationship vectors, attribute vectors and type vectors; The semantic similarity features and structural coherence features are calculated to obtain the immediate reward value, the subsequent nodes are sampled and the reward sequence is predicted, and the reward sequence is weighted based on the preset attenuation factor to obtain the predicted reward value; the immediate reward value and the predicted reward value are weightedly fused to obtain the comprehensive score of the path.

7. The method according to claim 6, characterized in that The semantic similarity features and structural coherence features are calculated to obtain the immediate reward value, the subsequent nodes are sampled and the reward sequence is predicted, and the reward sequence is weighted based on the preset attenuation factor to obtain the predicted reward value including: Constructing weight functions of semantic similarity features and structural coherence features, wherein the weight function of semantic similarity features adopts the form of hyperbolic tangent function, and the feature weight function of structural coherence adopts the form of Softmax function, and setting initial parameters for the weight function of semantic similarity features and the weight function of structural coherence features respectively; The Euclidean distance between the semantic similarity feature and the standard feature vector is calculated to obtain a first error value, the cosine distance between the structural coherence feature and the reference feature vector is calculated to obtain a second error value, and a weighted sum of the first error value and the second error value is constructed as an error function; Calculating the gradient values ​​of the parameters of the weight function of the semantic similarity feature and the weight function of the structural coherence feature based on the error function, scaling the gradient values ​​according to a preset learning rate, and subtracting the scaled gradient values ​​from the corresponding parameters to obtain updated parameters; Determine whether the change value of the error function is less than a preset error threshold. When the change value of the error function is greater than or equal to the preset error threshold, use the updated parameters to repeat the gradient calculation and parameter update steps. When the change value of the error function is less than the preset error threshold, determine the current parameters as the optimized adaptive weight coefficients. The semantic similarity feature is multiplied by the corresponding adaptive weight coefficient to obtain the weighted semantic feature, the structural coherence feature is multiplied by the corresponding adaptive weight coefficient to obtain the weighted structural feature, and the weighted semantic feature and the weighted structural feature are summed to obtain the immediate reward value.

8. An insurance customer service system based on natural language processing, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain multi-source insurance customer service historical conversation data, perform text preprocessing on the multi-source insurance customer service historical conversation data, and obtain standardized conversation text data; Based on deep semantic analysis, the standardized conversation text data is used to identify knowledge entities and extract relationships, and the initial insurance knowledge graph is constructed based on the identified knowledge entities and extracted relationships; The second unit is used to monitor the historical conversation data of multi-source insurance customer service in real time. When new conversation content is detected, the knowledge value of the new conversation content is evaluated based on the attention mechanism. The knowledge value evaluation is achieved by building a multi-dimensional scoring model to calculate the semantic similarity, knowledge coverage and user consultation frequency between the new conversation content and the existing knowledge in the initial insurance knowledge graph. The knowledge update priority is determined based on the multi-dimensional scoring results, and the valuable knowledge content that meets the preset priority threshold is used to trigger the dynamic update of the initial insurance knowledge graph. The third unit is used to receive the question text input by the user, perform intent recognition and information extraction on the question text, and obtain the question features to be matched; the question features are input into the question-answer matching model based on the bidirectional gated recurrent unit network, and the question-answer matching model performs multi-hop association matching on the question features and the nodes in the initial insurance knowledge graph, and calculates the path priority score including value, novelty and diversity through the path state evaluation model; based on the path priority score, the node is selected and the semantic similarity and coherence features of the node are calculated to obtain the immediate reward value, and the path comprehensive score is generated in combination with the predicted subsequent reward value.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images