Semantic analysis-based car insurance case retrieval method and system

Abstract

The invention relates to the technical field of data retrieval, in particular to a semantic analysis-based car insurance case retrieval method and system. Obtaining unstructured text data and visual modal data of a to-be-retrieved case; performing structured analysis on the unstructured text data, extracting entity objects and attribute parameters, converting the attribute parameters into space-time semantic sequence data, and constructing an event evolution graph structure; inputting the event evolution graph structure into a graph neural network model for coding, generating a graph embedding vector, and executing vector similarity search to obtain a preliminarily matched candidate case set; performing vectorization processing on the visual modal data, extracting a visual semantic feature vector, and calculating a semantic logic consistency metric value between the visual semantic feature vector and the space-time semantic sequence data; and reordering the candidate case set based on a weighted fusion result of the similarity score of the graph embedding vector and the semantic logic consistency metric value, and outputting an ordered target case list.

Description

Technical Field

[0001] This invention relates to the field of data retrieval technology, specifically to a method and system for retrieving car insurance cases based on semantic analysis. Background Technology

[0002] In intelligent data processing systems, massive amounts of multimodal data records are stored. Each record typically contains unstructured natural language descriptions and unstructured visual image data. The retrieval task of this type of data aims to find target records from the database that are highly matched with the input query in terms of semantic logic and physical characteristics.

[0003] Existing text retrieval technologies primarily rely on mapping statistical features or static semantic features. However, when processing text data describing physical dynamic processes, existing vectorization methods suffer from significant technical shortcomings: they typically treat text as a sequential set of word vectors, losing information about the temporal evolution of the data object and its spatial topological structure. In cross-modal retrieval, existing heterogeneous data alignment techniques often employ a strategy of mapping image feature vectors to text feature vectors in a common subspace. However, this mapping is usually based on statistical correlations of overall semantics, lacking geometric consistency verification of physical-level features. Specifically, the text description implicitly contains logical quantities of motion physics parameters (such as trajectory direction and velocity vectors), while the features extracted from images are typically statistical quantities of pixel distribution. This makes it impossible for the system to determine whether the logical motion trajectory described in the text is geometrically consistent with the physical form presented in the image.

[0004] In summary, existing technologies lack a retrieval mechanism that can parse unstructured text into a computer-computable dynamic graph structure and achieve precise alignment of text physical parameters with visual geometric features in vector space when processing such data.

[0005] To address this, a method and system for retrieving auto insurance cases based on semantic analysis is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for retrieving car insurance cases based on semantic analysis. By weighted fusion of similarity scores of graph embedding vectors and cross-modal consistency metrics, the candidate case set is reordered and a sorted list of target cases is output.

[0007] To achieve the above objectives, this invention provides a method for retrieving auto insurance cases based on semantic analysis, comprising: Acquire multimodal input data for the cases to be retrieved, including unstructured text data and visual modal data; Unstructured text data is parsed in a structured manner to extract entity objects and attribute parameters. Based on a preset action semantic mapping table, the attribute parameters are transformed into spatiotemporal semantic sequence data, and an event evolution graph structure representing the evolution logic of an accident is constructed. The event evolution graph structure includes physical entity nodes and virtual state nodes generated from spatiotemporal semantic sequence data. The event evolution graph structure is input into the graph neural network model for encoding, generating graph embedding vectors for the cases to be retrieved, and performing vector similarity search in a pre-set graph vector database to obtain a preliminary set of candidate cases based on the similarity score; Visual modal data is vectorized to extract visual semantic feature vectors that represent the morphology and location of damage, and the semantic logical consistency metric between the visual semantic feature vectors and the spatiotemporal semantic sequence data is calculated. Based on the weighted fusion result of the similarity score of graph embedding vectors and the semantic logical consistency metric, the candidate case set is reordered and the sorted target case list is output.

[0008] The process of constructing an event evolution graph structure specifically includes: The natural language processing model is used to identify dynamic behavior descriptions in the unstructured text data and extract action predicates and location modifiers. Establish an action semantic mapping table to transform action predicates and location modifiers into behavior state vectors and relative location identifiers; Set a logical evolution step size, perform discretization state deduction on the entity object based on the behavior state vector and relative orientation identifier, and generate spatiotemporal semantic sequence data containing continuous logical temporal sequence; instantiate the state features in the spatiotemporal semantic sequence data into virtual state nodes in the graph structure, and establish temporal connection edges between physical entity nodes and virtual state nodes to form the event evolution graph structure.

[0009] The data definition of the event evolution graph structure includes: Node layer data definition: includes a first type of node that identifies physical entities and stores entity ID and status attributes; includes a second type of node that identifies virtual spatiotemporal states and stores the logical timing and relative orientation status bits; Edge layer data definition: includes temporal edges, used to link the state changes of the same entity node under continuous logical temporal sequence, storing the state evolution weight; includes spatial relationship edges, used to link physical entity nodes and virtual spatiotemporal state nodes, storing the spatial topology relationship type; By defining data at the node and edge layers, graph index entries reflecting dynamic data relationships are constructed in the database.

[0010] The process of performing vector similarity search in a pre-defined graph vector database specifically includes: An attention-based graph isomorphic network is used as an encoder to perform message passing and aggregation operations on the event evolution graph structure, generating a fixed-dimensional graph embedding vector. In the graph vector database, an approximate nearest neighbor search algorithm is used to retrieve the K historical case vectors that are closest to the Euclidean distance, using the graph embedding vector as the query key. The retrieval process is based on the topological isomorphism of the graph structure to filter out a subset of cases with similar node connection patterns and data evolution paths.

[0011] The process of generating visual semantic feature vectors includes: Pre-trained deep neural network models are used to extract features from visual modal data to generate feature maps containing texture distribution information. The feature map is used to identify regions of interest and locate the deformation texture features of the damaged parts of the vehicle. Morphological encoding is performed on the deformation texture features to extract the distribution pattern and stress semantic features of the damaged area, and then encoded into a visual semantic feature vector. The visual semantic feature vector is used to characterize the logical attributes of physical impact and damage in a high-dimensional semantic space, serving as a semantic verification parameter in cross-modal retrieval.

[0012] The calculation of semantic logical consistency metrics specifically includes: Extract text semantic vectors representing dynamic behavior from spatiotemporal semantic sequence data; construct a unified hidden layer feature mapping space and map the visual semantic feature vectors and the text semantic vectors into the unified hidden layer feature mapping space; calculate the vector similarity between the mapped visual semantic feature vectors and the text semantic vectors. By normalizing the vector similarity, a semantic logical consistency metric is obtained, which represents the degree of physical and logical consistency between the text description and the visual presentation.

[0013] Reordering the candidate case set specifically includes dynamic weighted query processing; Specifically, this involves parsing the intent of the retrieval request and determining the query configuration parameters; If the query configuration parameters point to logical relevance retrieval, the weight coefficient of the graph embedding vector similarity in the ranking function is increased; if the query configuration parameters point to physical semantic consistency retrieval, the weight coefficient of the semantic logical consistency metric in the ranking function is increased; a comprehensive score is calculated based on the adjusted weight coefficients, and the candidate case set is sorted in descending order according to the comprehensive score.

[0014] A semantic analysis-based auto insurance case retrieval system includes: The data acquisition module acquires multimodal input data of the cases to be retrieved, including unstructured text data and visual modal data; The text parsing module performs structured parsing on unstructured text data, extracts entity objects and attribute parameters, transforms the attribute parameters into spatiotemporal semantic sequence data based on a preset action semantic mapping table, and constructs an event evolution graph structure that represents the evolution logic of the accident; the event evolution graph structure includes physical entity nodes and virtual state nodes generated from spatiotemporal semantic sequence data. The preliminary matching module encodes the event evolution graph structure into the graph neural network model, generates graph embedding vectors for the cases to be retrieved, performs vector similarity search in a pre-set graph vector database, and obtains a preliminary matching candidate case set based on the similarity score. The consistency measurement module performs vectorization processing on visual modal data, extracts visual semantic feature vectors that represent the characteristics of damaged morphology and location, and calculates the semantic logical consistency measurement value between the visual semantic feature vectors and the spatiotemporal semantic sequence data. The ranking optimization module re-ranks the candidate case set based on the weighted fusion result of the similarity score of the graph embedding vector and the semantic logical consistency metric, and outputs the ranked target case list.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention effectively solves the retrieval difficulties caused by the ambiguity of natural language descriptions and the lack of quantitative data by constructing an event evolution graph containing virtual state nodes; it transforms unstructured text streams into structured logical streams, and by introducing discretization deduction, it can deduce the dynamic process of an accident, enabling machines to understand not only the accident text but also the accident's occurrence process; this graph structure definition has extremely strong generalization ability, even if the wording of the reporting person is different, as long as the accident logic is consistent, the generated graph topology will be similar, thereby significantly improving the precision rate.

[0016] 2. The semantic logical consistency metric proposed in this invention provides double insurance for the accuracy and authenticity of search results; in fact, it performs an automated logical reasoning to determine whether the case description and evidence photos form a complete chain of evidence; this not only filters out those interference items that are merely keyword matches but are completely different from the actual case, but also helps to identify potential fraud risks, significantly improving the reference value and confidence of search results.

[0017] 3. This invention endows the retrieval system with strong human-computer interaction adaptability through a dynamic weight re-ranking mechanism, changing the traditional static ranking method of search engines. It can intelligently adjust the recommendation strategy according to the work priorities of personnel in different positions such as surveying, damage assessment, and claims settlement. This not only significantly reduces the time users spend manually filtering irrelevant results, but also makes the recommended results more practical and targeted in specific business scenarios, thereby improving the processing efficiency of the entire auto insurance claims process. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a semantic analysis-based method for retrieving car insurance cases according to the present invention. Figure 2 This is a data logic diagram of a car insurance case retrieval method based on semantic analysis according to the present invention; Figure 3 This is a schematic diagram of the structure of a car insurance case retrieval system based on semantic analysis according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: This invention proposes a method for retrieving auto insurance cases based on semantic analysis. The process of the method is as follows: Figure 1 As shown, the data logic of the method is as follows: Figure 2 As shown, it includes: Acquire multimodal input data for the cases to be retrieved, including unstructured text data and visual modal data; Unstructured text data is parsed in a structured manner to extract entity objects and attribute parameters. Based on a preset action semantic mapping table, the attribute parameters are transformed into spatiotemporal semantic sequence data, and an event evolution graph structure representing the evolution logic of an accident is constructed. The event evolution graph structure includes physical entity nodes and virtual state nodes generated from spatiotemporal semantic sequence data. The event evolution graph structure is input into the graph neural network model for encoding, generating graph embedding vectors for the cases to be retrieved, and performing vector similarity search in a pre-set graph vector database to obtain a preliminary set of candidate cases based on the similarity score; Visual modal data is vectorized to extract visual semantic feature vectors that represent the morphology and location of damage, and the semantic logical consistency metric between the visual semantic feature vectors and the spatiotemporal semantic sequence data is calculated. Based on the weighted fusion result of the similarity score of graph embedding vectors and the semantic logical consistency metric, the candidate case set is reordered and the sorted target case list is output.

[0021] The process of constructing an event evolution graph structure specifically includes: The natural language processing model is used to identify dynamic behavior descriptions in the unstructured text data and extract action predicates and location modifiers. Establish an action semantic mapping table to transform action predicates and location modifiers into behavior state vectors and relative location identifiers; Set a logical evolution step size, perform discretization state deduction on the entity object based on the behavior state vector and relative orientation identifier, and generate spatiotemporal semantic sequence data containing continuous logical temporal sequence; instantiate the state features in the spatiotemporal semantic sequence data into virtual state nodes in the graph structure, and establish temporal connection edges between physical entity nodes and virtual state nodes to form the event evolution graph structure.

[0022] The data definition of the event evolution graph structure includes: Node layer data definition: includes a first type of node that identifies physical entities and stores entity ID and status attributes; includes a second type of node that identifies virtual spatiotemporal states and stores the logical timing and relative orientation status bits; Edge layer data definition: includes temporal edges, used to link the state changes of the same entity node under continuous logical temporal sequence, storing the state evolution weight; includes spatial relationship edges, used to link physical entity nodes and virtual spatiotemporal state nodes, storing the spatial topology relationship type; By defining data at the node and edge layers, graph index entries reflecting dynamic data relationships are constructed in the database.

[0023] In order to extract key information from complex police reports, this embodiment uses a pre-trained language model (such as the Chinese BERT model) based on a transformer architecture as a feature extractor, with a bidirectional long short-term memory network and a conditional random field layer cascaded in the back end.

[0024] The workflow of a natural language processing model includes: the input text sequence first passes through a transformer layer to generate context-sensitive word vectors; then, a bidirectional long short-term memory network captures long-distance semantic dependencies; finally, a conditional random field layer constrains the output label sequence; for example, ensuring that locative words immediately follow verbs.

[0025] The model can accurately identify physical entities, including our vehicles, third-party vehicles, and roadside guardrails; action predicates, including lane change, emergency braking, rear-end collision, and scrape; and directional modifiers, including left, same direction, and rear.

[0026] Since the extracted vocabulary is static, it needs to be transformed into a dynamic logical process; Action semantic mapping table: A high-dimensional mapping table is pre-defined; for example, the word "change lanes" is not just a word, but is mapped to a state vector of "non-zero lateral displacement and maintained longitudinal velocity" in the system; "emergency braking" is mapped to a state vector of "negative longitudinal acceleration"; Discretization deduction logic: Set the "logic evolution step size", which is usually divided into three logical time steps: "initial state", "intermediate action state" and "final contact state". The discretization deduction process includes two stages: the first stage is the automatic determination of the number of steps, and the second stage is the deduction of the state within each step size.

[0027] The first stage involves the automatic determination of the step size. Example of deduction: If the text is "When driving in the same direction, the other party forcibly changed lanes and hit my side"; Time step one (start): Based on "same direction", two virtual nodes are generated, representing the two vehicles respectively, and the logical coordinates show that the two are parallel; Time step two (action): Based on "lane change", according to the mapping table, deduce that the virtual node of the other vehicle will generate lateral displacement and move closer to the node of our vehicle.

[0028] Step 3 (Contact): Based on the collision, determine that the virtual nodes of the two vehicles overlap in the logical space and establish an interactive relationship.

[0029] The second stage involves the deduction of the state within each step. Based on the order and position of action predicates in the paragraph, the corresponding action is mapped to that step length; if there is no explicit action description within that step length, the state of the previous step length is inherited.

[0030] For the i-th step, a virtual state node is created. The attributes of this node include: node ID, logical sequence label, entity reference set, behavior state vector, and relative orientation identifier. Among them, the node ID and logical sequence label are obtained by sorting the accumulated virtual nodes; the entity reference set contains all physical entities involved in this step, such as our vehicles and the enemy vehicles; the behavior state vector is obtained by querying the action semantic mapping table based on the action predicate of this step; and the relative orientation identifier is obtained by querying the mapping table based on the orientation modifiers involved in this step. For the same physical vehicle, connect its virtual state nodes at step i and step i+1 to create a temporal edge; for two different vehicles within the same logical temporal i, if there is a spatial relationship description in the text, connect their corresponding virtual state nodes to create a spatial relationship edge.

[0031] Result generated: This process does not rely on real GPS data, but generates a set of spatiotemporal semantic sequence data, which is essentially a string of logical state codes that represent the behavior pattern of objects on the time axis.

[0032] Node layer design: The event evolution graph structure contains two types of nodes; The first type is physical entity nodes, which serve as anchor points in the graph and store static attributes such as vehicle ID and vehicle model. The second type is virtual state nodes, which are instantaneous state snapshots generated at each time step in the above deduction process, storing the logical coordinates and behavior vectors at that time.

[0033] Edge design: The event evolution graph structure contains two types of edges; Temporal edges: virtual nodes of the same entity at different time steps. The weight of the edge represents the order of time passage, forming a behavioral chain of a single vehicle.

[0034] Spatial relationship edge: A virtual node connecting different entities at the same time step, used to represent the relative positional relationship between two vehicles, such as "close" or "far apart".

[0035] The resulting graph structure not only has a physical skeleton but also includes behavioral interactions that evolve over time.

[0036] Furthermore, when constructing temporal edges, it is not simply a matter of connecting nodes at two consecutive time points. A numerical state evolution weight is also calculated and stored to measure the drasticness and logical importance of the entity's behavior changes between two consecutive logical moments.

[0037] The state evolution weights are dynamically calculated based on two dimensions: semantic vector distance dimension and time decay dimension. Semantic vector distance dimension: Calculate the semantic distance between the behavior state vector of the previous moment and the behavior state vector of the current moment; if the vehicle changes from "driving straight at a constant speed" to "driving straight at a constant speed", the state remains unchanged and the vector distance between the two is minimal, so this time step is given a low weight, indicating that this stage is redundant information; if the vehicle suddenly changes from "driving straight" to "sudden braking" or "sudden left turn", the state changes abruptly and the vector distance between the two is huge, so this time step is given a high weight, indicating that a critical physical transition has occurred in this stage; Time decay dimension: Introducing a time attention mechanism; in terms of logical timing, the time step closer to the "collision occurrence time" will be given an additional gain coefficient for its corresponding time edge weight; while the weight of the initial stage that is farther away from the collision will be reduced accordingly.

[0038] These weights are stored directly as edge attributes in the edge table of the graph database, acting as flow valves during subsequent message passing in the graph neural network. For high-weight temporal edges, more gradient and feature information are allowed to pass through, making the model highly focused on key actions that lead to accidents, such as "sudden braking" and "lane changes." For low-weight temporal edges, information transmission is suppressed, thereby automatically ignoring background noise that is meaningless for determining liability, such as normal vehicle driving.

[0039] This invention endows graph structures with the ability to distinguish between priorities by introducing state evolution weights into temporal edges.

[0040] Key Feature Focus: Traditional methods often treat the entire accident process equally, resulting in the crucial "one second before collision" information being overwhelmed by a large amount of "normal driving" information; this solution focuses on the turning point of the accident by calculating the degree of state change, enabling the retrieval model to accurately capture the root cause of the accident.

[0041] Improved noise resistance: The lengthy background information in the report description is automatically filtered out during the graph encoding process because its state evolution weight is extremely low, thus ensuring that the generated graph embedding vector is pure and has high discriminative power.

[0042] This invention effectively solves the retrieval challenges caused by vague natural language descriptions and a lack of quantitative data by constructing an event evolution graph containing virtual state nodes. It transforms unstructured text streams into structured logical streams, and by introducing discretization deduction, it can deduce the dynamic process of an accident, enabling machines to understand not only the accident text but also the accident's occurrence process. This graph structure definition has extremely strong generalization ability; even if the wording of the reporting person is different, as long as the accident logic is consistent, the generated graph topology will be similar, thereby significantly improving the precision rate.

[0043] The process of performing vector similarity search in a pre-defined graph vector database specifically includes: An attention-based graph isomorphic network is used as an encoder to perform message passing and aggregation operations on the event evolution graph structure, generating a fixed-dimensional graph embedding vector. In the graph vector database, an approximate nearest neighbor search algorithm is used to retrieve the K historical case vectors that are closest to the Euclidean distance, using the graph embedding vector as the query key. The retrieval process is based on the topological isomorphism of the graph structure to filter out a subset of cases with similar node connection patterns and data evolution paths.

[0044] In this embodiment, a graph isomorphic network based on an attention mechanism is selected as the core encoder.

[0045] Message passing and aggregation mechanism: In a graph neural network, each node sends information to its neighboring nodes; for each virtual state node, it aggregates information from the previous time step (past state) and neighboring vehicles (spatial interaction) at the same time step.

[0046] Application of the attention mechanism: Not all states are equally important; the network introduces a self-attention module to automatically calculate the weight coefficient of each node. For example, the weight of a node at the moment of a collision will be automatically amplified, while the weight of a node at the moment of normal driving will be reduced.

[0047] Graph embedding vector generation: After multiple layers of message passing, the features of all nodes in the graph are aggregated into a fixed-length global feature vector, i.e., the graph embedding vector; in high-dimensional space, it represents the complete logical topology of the incident.

[0048] Graph Vector Database Construction: Graph embedding vectors of historical cases are pre-stored in a high-performance vector database.

[0049] Retrieval Algorithm: An approximate nearest neighbor search algorithm, such as the hierarchical navigation small-world algorithm, is employed. The Euclidean distance between the vector of the case to be retrieved and the vectors in the database is calculated. Due to the characteristics of isomorphic graph networks, the closer the distance, the more similar the topological structures of the two graphs, meaning the more similar the occurrence processes of the two incidents, such as in liability determination logic.

[0050] This invention achieves deep semantic alignment of complex traffic accident dynamic processes by employing graph isomorphic networks for encoding. Compared to traditional keyword matching, this scheme can capture the structured dependencies between nodes; through an attention mechanism, it can automatically focus on the key turning points of the accident, ignoring irrelevant background noise; and the application of an approximate nearest neighbor search algorithm enables the system to filter logically isomorphic candidate sets from a massive historical case database within milliseconds, greatly satisfying the stringent timeliness requirements of auto insurance claims.

[0051] Furthermore, the steps for automatically calculating the weight coefficient of each node include: Step 1: Feature gating aggregation based on edge weights; In the message passing phase of the graph neural network, the state evolution weights stored in the temporal edges of the event evolution graph are directly called; the state evolution weights are used as flow control coefficients when neighboring node features are transmitted to the current node; for temporal edges with high weights, the feature signals of their connected predecessor nodes are amplified, and for temporal edges with low weights, their feature signals are suppressed, thereby completing the initial screening of key action information at the feature aggregation level.

[0052] Step 2: Generate motion-sensitive augmented node vectors; The initial state attributes of the current node are updated using the aggregated feature signals to generate an enhanced node vector. The numerical intensity of the enhanced node vector implicitly includes the intensity of the preceding action, so that the virtual state node that has undergone high-weight actions such as sudden braking and lane changing has higher activation significance in the feature space.

[0053] Step 3: Adaptive scoring based on feature saliency; A globally trainable query vector is introduced to map and calculate the matching degree of the enhanced node vectors of all nodes. Since the key nodes have been integrated with high-intensity edge weight information in step two, their matching score with the query vector is automatically improved, thereby calculating the node weight coefficient.

[0054] Step 4: Weighted generation of global graph embedding; Using the node weight coefficients, the vectors of all virtual state nodes in the entire graph are weighted and summed to generate the final graph embedding vector. This process ensures that the generated graph embedding vector is mainly dominated by high-weight key incident evolution nodes, realizing the logical transmission from the intensity of the action to the importance of semantic representation.

[0055] The process of generating visual semantic feature vectors includes: Pre-trained deep neural network models are used to extract features from visual modal data to generate feature maps containing texture distribution information. The feature map is used to identify regions of interest and locate the deformation texture features of the damaged parts of the vehicle. Morphological encoding is performed on the deformation texture features to extract the distribution pattern and stress semantic features of the damaged area, and then encoded into a visual semantic feature vector. The visual semantic feature vector is used to characterize the logical attributes of physical impact and damage in a high-dimensional semantic space, serving as a semantic verification parameter in cross-modal retrieval.

[0056] The deep neural network model performs deep feature extraction and region of interest identification as follows: the backbone network uses a deep residual network (such as ResNet-101) combined with a feature pyramid structure to extract multi-scale feature maps of the image; the mask region convolutional neural network is used to segment the damaged parts of the vehicle, accurately cropping out the regions of interest containing dents, scratches, and tears, and removing background interference.

[0057] Furthermore, the morphological encoding and force semantic extraction process includes texture flow analysis and force semantic encoding: The texture flow analysis includes: for the cropped damaged area, calculating the distribution histogram of its pixel gradient; if the gradient direction is highly concentrated and singular, it is encoded as a linear scraping feature, corresponding to physical tangential force, such as lane change scraping; if the gradient direction is radial or disordered, it is encoded as an aggregated deformation feature, corresponding to physical normal impact, such as rear-end collision or perpendicular side collision.

[0058] The force semantic encoding includes mapping the aforementioned texture distribution features into a visual semantic feature vector. The dimensions of this vector no longer represent color or shape, but rather the direction of force and contact pattern; for example, the first dimension of the vector represents the probability of lateral force, and the second dimension represents the probability of longitudinal force.

[0059] This invention, by introducing morphological coding, overcomes the limitations of traditional image retrieval, which only focuses on appearance similarity, and achieves retrieval based on physical mechanism similarity. It can infer the stress process from static photographs, distinguishing between seemingly similar stains and genuine physical dents. This feature extraction based on stress semantics provides a solid physical basis for subsequent verification of whether the accident process described in the text matches the damage result shown in the image.

[0060] The calculation of semantic logical consistency metrics specifically includes: Extract text semantic vectors representing dynamic behavior from spatiotemporal semantic sequence data; construct a unified hidden layer feature mapping space and map the visual semantic feature vectors and the text semantic vectors into the unified hidden layer feature mapping space; calculate the vector similarity between the mapped visual semantic feature vectors and the text semantic vectors. By normalizing the vector similarity, a semantic logical consistency metric is obtained, which represents the degree of physical and logical consistency between the text description and the visual presentation.

[0061] Since the spatiotemporal semantic sequence generated from text and the visual semantic vector generated from images belong to different data modalities, they cannot be directly compared. Therefore, a shared unified hidden layer feature mapping space is constructed. The mapping process is as follows: two independent fully connected layer networks are used to project text features and visual features into this shared space, respectively.

[0062] Furthermore, the construction of the unified hidden layer feature mapping space and model parameters include: Choice of mapping space dimension: In this embodiment, the dimension of the unified hidden layer feature mapping space is set to M=256. 256 dimensions are sufficient to represent the core semantic information of both text and visual high-dimensional input data, while avoiding the computational bottleneck caused by excessively high dimensions; at the same time, the moderate dimension of 256 can effectively constrain the number of parameters of the mapping network and prevent overfitting when training data is limited.

[0063] The structural design of the mapping network is the same for the two mapping networks that project text semantic vectors and visual semantic feature vectors into a unified space. The structure of each mapping network is as follows: input layer, fully connected layer 1, batch normalization layer, dropout layer, fully connected layer 2, layer normalization, and output layer; The optimizer used is the Adam optimizer, with an initial learning rate of 1e-4 and a cosine annealing method for learning rate decay. The norm of the gradient is limited to within 1.0 to prevent gradient explosion. The mini-batch size is 256, and the number of training epochs is 100.

[0064] Training strategy: During the model training phase, a contrastive learning strategy is adopted; matching text-image pairs are used as positive samples to narrow their distance in the shared space; and mismatched text-image pairs are used as negative samples to widen their distance.

[0065] The consistency metric calculation process includes similarity score calculation and normalization processing; Similarity score: In the shared space, the cosine similarity between the mapped text vector and the visual vector is calculated, which is the cosine value of the angle between the two vectors; Normalization: Normalize the cosine value to the range of 0 to 1; the closer the score is to 1, the more closely the lane change process described in the text matches the linear scratches on the side shown in the picture in terms of physical logic; a low score suggests that there may be a contradiction, such as the text saying a rear-end collision, but the picture shows side damage.

[0066] The semantic-logical consistency metric proposed in this invention provides double insurance for the accuracy and authenticity of search results; in fact, it performs an automated logical reasoning to determine whether the case description and evidence photos form a complete chain of evidence; this not only filters out those interference items that are merely keyword matches but are completely different from the actual case, but also helps to identify potential fraud risks, significantly improving the reference value and confidence of search results.

[0067] Reordering the candidate case set specifically includes dynamic weighted query processing; Specifically, this involves parsing the intent of the retrieval request and determining the query configuration parameters; If the query configuration parameters point to logical relevance retrieval, the weight coefficient of the graph embedding vector similarity in the ranking function is increased; if the query configuration parameters point to physical semantic consistency retrieval, the weight coefficient of the semantic logical consistency metric in the ranking function is increased; a comprehensive score is calculated based on the adjusted weight coefficients, and the candidate case set is sorted in descending order according to the comprehensive score.

[0068] Logical correlation retrieval mode: When users (such as accident liability determiners) are mainly concerned with accident liability determination, adjust the configuration parameters to focus on the course of the accident; Physical Feature Consistency Retrieval Mode: When users (such as damage assessors) are primarily focused on repair plans, adjust the configuration parameters to emphasize the specific form of damage.

[0069] Overall score calculation: The configuration parameters include the similarity score of the graph embedding vectors and the weight coefficients of the semantic-logical consistency metric; the formula for calculating the overall score based on dynamic weights is as follows: ; in, This indicates the overall score; This represents the similarity score of the graph embedding vectors; This represents a semantic and logical consistency metric. This represents the graph embedding similarity weights; This represents the weight of the semantic logical consistency metric.

[0070] In the liability-based model, the weight of graph similarity is significantly increased to ensure that the recommended cases are all cases with the same logic, such as "changing lanes and hitting a straight-going vehicle". For example, the weight of graph embedding vector similarity score can be set to 0.7, and the semantic logic consistency metric can be set to 0.3.

[0071] In the damage assessment mode, the weight of visual consistency is increased to ensure that the recommended cases are all cases with the same degree of damage, such as "severe dent of the left front door". For example, the weight of the graph embedding vector similarity score can be set to 0.3 and the semantic logical consistency metric can be set to 0.7.

[0072] Output: Candidate cases are sorted from highest to lowest based on their overall scores, and the final list of target cases is displayed to the user.

[0073] This invention endows the retrieval system with strong human-computer interaction adaptability through a dynamic weight re-ranking mechanism, changing the traditional static ranking method of search engines. It can intelligently adjust recommendation strategies based on the work priorities of personnel in different positions such as surveying, damage assessment, and claims processing. This not only significantly reduces the time users spend manually filtering irrelevant results but also makes the recommended results more practical and targeted in specific business scenarios, thereby improving the processing efficiency of the entire auto insurance claims process.

[0074] Furthermore, based on the fixed weight mode configured above, the weight can also be precisely measured based on the user's search behavior, specifically including... Construct and deploy a user intent recognition neural network model, which includes an input layer, a temporal feature extraction layer, and an intent probability output layer; Real-time collection of continuous interactive behavior sequence data generated by users on the search interface, the continuous interactive behavior sequence data includes mouse trajectory coordinates with timestamps, control hover duration, scroll speed change rate and click operation type; The continuous interaction sequence data is vectorized and input into the user intent recognition neural network model; The temporal feature extraction layer captures the temporal dependencies and contextual patterns of interactive behaviors, and the intent probability output layer calculates the probability distribution value of the current user's preference for "logical relevance retrieval" or "physical feature consistency retrieval". The probability distribution value is directly mapped to graph embedding vector similarity weights and cross-modal consistency metric weights, and the ranking result of the candidate case set is dynamically updated.

[0075] Data Acquisition and Sequence Construction: Recording a detailed flow of behavioral events. For example: {Time Point 1: Mouse quickly hovers over a text area; Time Point 2: Mouse slows down and hovers over the damaged area of ​​the image; Time Point 3: Image zoom-in operation is performed; Time Point 4: Image details are slowly scrolled through}.

[0076] The discrete actions described above are transformed into a high-dimensional vector sequence; the vector at each time step includes: region identifier (text area, image area, list area), action type (move, hover, click, scroll), duration, mouse movement speed, etc.

[0077] The architecture design of the user intent recognition neural network model includes: Input layer: Receives the above sequence of behavioral vectors; Temporal feature extraction layer: Employs a Long Short-Term Memory (LSTM) network or Gated Recurrent Unit (GRU) structure to capture the implicit logic of the sequence of actions; for example, "reading the text first and then repeatedly looking at the pictures" and "only looking at the pictures and not the text" represent completely different concerns; the model memorizes and analyzes the contextual dependencies of the user's operation flow through a recurrent neural network structure.

[0078] Fully connected layer and attention layer: After extracting temporal features, an attention mechanism layer is introduced to automatically learn which actions in the behavior sequence are most critical for intent judgment and assign them higher weights.

[0079] Output layer: Employs a normalized exponential function to output probability values ​​in two dimensions: logical attention probability and physical attention probability.

[0080] The model training and inference process includes: Training data acquisition: Construct a dataset through historical logs; select the conversation records where users ultimately adopted a specific case. If the user ultimately adopted a case with "similar facts but different pictures", it is labeled as "logical concern"; if the user adopted a case with "identical injuries", it is labeled as "physical concern".

[0081] Inference application: When running online, the model reads the behavior sequence over a period of time at specific time slices, outputs the current weight suggestions in real time, and seamlessly adjusts the sorting of the list below.

[0082] This invention achieves a qualitative leap in the depth and accuracy of user intent recognition by introducing a temporal behavioral neural network. Its beneficial effects are mainly reflected in the following three aspects: First, it overcomes the limitations of linear rules in representing complex human behavior. Traditional rule-based judgments are too rigid and cannot understand the micro-operational patterns corresponding to complex psychological states such as hesitation, comparison, and rapid scanning. This solution utilizes the powerful modeling capabilities of recurrent neural networks for time-series data to capture extremely subtle non-linear features in user operations. For example, the "rhythm of mouse movement speed" often implies the user's degree of hesitation or focus, which cannot be defined by simple rules.

[0083] Secondly, it achieves intent tracking from static snapshots to dynamic streams, where the user's search intent changes dynamically within a session. The model of this solution is based on real-time reasoning of continuous sequences, which can keenly perceive the real-time shift of the user's focus between legal logic and physical evidence, and adjust the weights in milliseconds to ensure that the recommendation results are always synchronized with the user's current focus of thought.

[0084] Finally, it possesses self-evolving adaptive capabilities; the neural network model can continuously iterate through online learning, and as users with different experience levels use it, it can automatically learn the unique operational fingerprints of different groups, adapting to changes in business scenarios without rewriting rules, greatly reducing system maintenance costs and improving the level of intelligence in user experience.

[0085] Example 2: This invention proposes a car insurance case retrieval system based on semantic analysis, the structure of which is as follows: Figure 3 As shown, the system mainly runs on a high-performance computing server cluster and is logically divided into five core processing modules: data acquisition module, text parsing module, preliminary matching module, consistency measurement module, and sorting optimization module.

[0086] The data acquisition module acquires multimodal input data for the cases to be retrieved, including unstructured text data and visual modal data. This module is responsible for connecting to the insurance company's claim database and acquiring multimodal input data for the cases to be retrieved in real time. This mainly includes unstructured text data, such as claim summaries and investigation records, as well as visual modal data, such as on-site investigation photos and detailed vehicle damage images.

[0087] The text parsing module performs structured parsing on unstructured text data, extracts entity objects and attribute parameters, transforms the attribute parameters into spatiotemporal semantic sequence data based on a preset action semantic mapping table, and constructs an event evolution graph structure that represents the evolution logic of the accident; the event evolution graph structure includes physical entity nodes and virtual state nodes generated from spatiotemporal semantic sequence data. The text parsing module includes the construction of a natural language processing model and an event evolution graph structure; it integrates a natural language processing model based on a transformer architecture. The natural language processing model architecture is implemented as follows: Input embedding layer: Using a pre-trained transformer model (BERT) as the base, the text description in the input report text sequence is transformed into a high-dimensional word vector. This vector not only contains the meaning of the word itself, but also integrates the contextual information of the whole sentence through a self-attention mechanism.

[0088] Context feature extraction layer: The output of the embedding layer is fed into a bidirectional long short-term memory network; this network contains two data streams, forward and backward, which scan the text from left to right and from right to left, respectively. The generated hidden state vectors are concatenated to capture long-distance semantic dependencies, such as the logical relationship between "our vehicle" at the beginning of the sentence and "damaged state" at the end of the sentence.

[0089] Label decoding layer: Finally, entity and action labels are output through the conditional random field layer; this layer learns the transition probabilities between labels and forces the model to follow grammatical logic, such as "locative words" must be followed by "verbs", so as to accurately extract "physical entities", "action predicates" and "locative modifiers".

[0090] Graph-structured execution: Based on the extracted parameters, the module performs discretization deduction using a pre-defined action semantic mapping table to generate an event evolution graph containing virtual state nodes. During this process, the module calculates the semantic distance of actions, generates state evolution weights, and stores them in the temporal edges.

[0091] Furthermore, the specific implementation of the natural language processing model includes: Pre-trained model selection and fine-tuning: A Chinese BERT-base model was used for supervised fine-tuning in the field of auto insurance. The fine-tuning training set consisted of 1,000 manually annotated claim texts from the insurance claims database, of which 600 were used as the training set, 200 as the validation set, and 200 as the test set.

[0092] Entity and Action Annotation Standards: Defining Three Annotation Categories: Physical entities include our vehicles, third-party vehicles, and fixed facilities; Action predicates include major actions such as "lane change" (lateral movement), "sudden braking" (longitudinal deceleration), "turning" (angle change), "rear-end collision" (longitudinal approach and contact), and "scratching" (lateral contact). Location modifiers include multiple location markers such as "front left," "right side," and "rear."

[0093] The model configuration parameters include: the maximum input sequence length for BERT is 512 tokens; the hidden layer dimension of the bidirectional long short-term memory network is 256; the fine-tuning learning rate is 2e-5; the batch size is 32; and the training epochs are 10.

[0094] Action semantic mapping table: Establish a mapping function to map each extracted action predicate to a behavior state vector; for example, going straight is mapped to [1,0,0,0,0], turning left is mapped to [0,1,0,0,0], turning right is mapped to [0,0,1,0,0], and braking is mapped to [0,0,0,1,0].

[0095] The preliminary matching module encodes the event evolution graph structure into the graph neural network model, generates graph embedding vectors for the cases to be retrieved, performs vector similarity search in a pre-set graph vector database, and obtains a preliminary matching candidate case set based on the similarity score. This module is the system's search engine, responsible for filtering logically similar subsets from a massive amount of historical cases.

[0096] Its core algorithm is a graph isomorphic network based on the attention mechanism, and the model architecture implementation includes: Feature aggregation logic: Input the event evolution graph into the graph neural network; when passing messages, perform gating aggregation operation and use the state evolution weight in the temporal edge as the flow control coefficient; if the weight of an edge is close to 0, it means that the vehicle state has not changed, so the transmission of neighbor information is blocked; if the weight is close to 1, it means that a sudden braking or collision has occurred, so the transmission of neighbor information is amplified. Node adaptive scoring: A globally trainable query vector is matched with the features of each node to calculate the attention coefficient; since key nodes have already fused with high-intensity edge weight information in the previous step, their attention coefficients will be automatically amplified; Graph embedding generation: Finally, the module aggregates all the weighted node features into a fixed-dimensional graph embedding vector; Retrieval Execution: The module uses an approximate nearest neighbor search algorithm (such as the hierarchical navigation small-world algorithm) to retrieve the K historical cases that are closest to the current graph embedding vector in the pre-set graph vector database, forming a candidate set.

[0097] The consistency measurement module performs vectorization processing on visual modal data, extracts visual semantic feature vectors that represent the characteristics of damaged morphology and location, and calculates the semantic logical consistency measurement value between the visual semantic feature vectors and the spatiotemporal semantic sequence data. This module is the system's logic validator, used to determine whether the text description matches the image display; The visual feature extraction process includes: Region localization: First, the region of interest of the damaged parts of the vehicle is automatically cropped using a mask region convolutional neural network; Morphological encoding: Texture flow analysis is performed on the cropped region; the system calculates the orientation histogram of pixel gradients and its information entropy. If the information entropy is low, it is determined to be ordered linear scratching, corresponding to tangential force; if the information entropy is high, it is determined to be disordered aggregated deformation, corresponding to normal force. This feature is mapped into a visual semantic feature vector containing "tangential force probability" and "normal force probability".

[0098] Furthermore, morphological encoding uses Mask R-CNN for instance segmentation, with ResNet-101 as the backbone network. It is fine-tuned on a vehicle damage dataset (containing 5000 manually labeled vehicle damage photos) and trained for a total of 50 epochs. The output mask is used to accurately crop the damaged area.

[0099] Gradient calculation and histogram statistics: For the cropped region of interest, the Sobel operator is applied to calculate the horizontal and vertical gradients respectively. Based on the horizontal and vertical gradients, the gradient magnitude and gradient direction are further identified. The direction is quantized into 16 directional intervals, each covering 22.5 degrees. The number of pixels falling in each direction is counted to form a 16-dimensional histogram. Texture distribution discrimination: Calculate the information entropy of the histogram to determine whether it is a linear scratching pattern, corresponding to tangential force; or a clustered deformation pattern, corresponding to normal impact. For linear scratching patterns, locate the main direction and obtain the probability of lateral and longitudinal force; for clustered deformation patterns, determine the probability of lateral and longitudinal force based on the texture.

[0100] The cross-modal alignment implementation process includes: Unified spatial mapping: Using two independent fully connected layer networks, the "spatiotemporal semantic sequence vector" and "visual semantic feature vector" are projected onto a shared feature space of the same dimension, and layer normalization is performed. Consistency calculation: Calculate the cosine similarity between the two in the shared space and normalize it to a value between 0 and 1, which serves as a semantic and logical consistency measure.

[0101] The ranking optimization module re-ranks the candidate case set based on the weighted fusion result of the similarity score of the graph embedding vector and the semantic logical consistency metric, and outputs the ranked target case list.

[0102] This module is responsible for presenting the final results. To achieve intelligent sorting, a user intent recognition neural network is integrated into the module.

[0103] The implementation of the user intent recognition neural network includes: Behavior serialization: The system front-end collects user interaction data such as mouse trajectory, hover duration, and scrolling speed in real time, and converts it into a sequence of behavior vectors with timestamps. Specifically, the types of actions displayed by the user through the mouse include mouse movement, hovering, clicking, and scrolling; the mouse trajectory is the normalized position of the mouse coordinates relative to the screen; the mouse hovering area is identified as the text area, image area, and list area; the mouse click action includes the user's click interval time; the scrolling speed is the normalized scrolling speed; all of the above data has been normalized.

[0104] Temporal modeling: The sequence is input into a Long Short-Term Memory (LSTM) network, and the model recursively processes each time step, memorizing the user's operation rhythm; the input layer receives the behavioral sequence and uses a bidirectional LSTM as the temporal feature extraction layer; the hidden layer dimension of the forward LSTM is 128, and the output dimension is 128; the hidden layer dimension of the backward LSTM is 128, and the output dimension is 128; the dimension after bidirectional concatenation is 256.

[0105] Attention-weighted: An attention mechanism is introduced to identify the most representative operation actions and generate a comprehensive feature vector of intent.

[0106] Probability output: By using a normalized exponential function, output the probability distribution of whether the user currently prefers "logical association" or "physical characteristics"; Dynamic reordering: Based on the above probability values, the module dynamically adjusts the weight coefficients of "graph embedding vector similarity" and "semantic logical consistency metric", calculates the comprehensive score, reorders the candidate case set in descending order, and pushes the final list of target cases to the user interface.

[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for retrieving auto insurance cases based on semantic analysis, characterized in that, include: Acquire multimodal input data for the cases to be retrieved, including unstructured text data and visual modal data; Unstructured text data is parsed in a structured manner to extract entity objects and attribute parameters. Based on a preset action semantic mapping table, the attribute parameters are transformed into spatiotemporal semantic sequence data, and an event evolution graph structure representing the evolution logic of an accident is constructed. The event evolution graph structure includes physical entity nodes and virtual state nodes generated from spatiotemporal semantic sequence data. The event evolution graph structure is input into the graph neural network model for encoding, generating graph embedding vectors for the cases to be retrieved, and performing vector similarity search in a pre-set graph vector database to obtain a preliminary set of candidate cases based on the similarity score; Visual modal data is vectorized to extract visual semantic feature vectors that represent the morphology and location of damage, and the semantic logical consistency metric between the visual semantic feature vectors and the spatiotemporal semantic sequence data is calculated. Based on the weighted fusion result of the similarity score of graph embedding vectors and the semantic logical consistency metric, the candidate case set is reordered and the sorted target case list is output.

2. The method for retrieving auto insurance cases based on semantic analysis according to claim 1, characterized in that: The process of constructing an event evolution graph structure specifically includes: The natural language processing model is used to identify dynamic behavior descriptions in the unstructured text data and extract action predicates and location modifiers. Establish an action semantic mapping table to transform action predicates and location modifiers into behavior state vectors and relative location identifiers; Set a logical evolution step size, perform discretization state deduction on the entity object based on the behavior state vector and relative orientation identifier, and generate spatiotemporal semantic sequence data containing continuous logical temporal sequence; instantiate the state features in the spatiotemporal semantic sequence data into virtual state nodes in the graph structure, and establish temporal connection edges between physical entity nodes and virtual state nodes to form the event evolution graph structure.

3. The method for retrieving auto insurance cases based on semantic analysis according to claim 2, characterized in that: The data definition of the event evolution graph structure includes: Node layer data definition: includes a first type of node that identifies physical entities and stores entity ID and status attributes; includes a second type of node that identifies virtual spatiotemporal states and stores the logical timing and relative orientation status bits; Edge layer data definition: includes temporal edges, used to link the state changes of the same entity node under continuous logical temporal sequence, storing the state evolution weight; includes spatial relationship edges, used to link physical entity nodes and virtual spatiotemporal state nodes, storing the spatial topology relationship type; By defining data at the node and edge layers, graph index entries reflecting dynamic data relationships are constructed in the database.

4. The method for retrieving auto insurance cases based on semantic analysis according to claim 1, characterized in that: The process of performing vector similarity search in a pre-defined graph vector database specifically includes: An attention-based graph isomorphic network is used as an encoder to perform message passing and aggregation operations on the event evolution graph structure, generating a fixed-dimensional graph embedding vector. In the graph vector database, an approximate nearest neighbor search algorithm is used to retrieve the K historical case vectors that are closest to the Euclidean distance, using the graph embedding vector as the query key. The retrieval process is based on the topological isomorphism of the graph structure to filter out a subset of cases with similar node connection patterns and data evolution paths.

5. The method for retrieving auto insurance cases based on semantic analysis according to claim 1, characterized in that: The process of generating visual semantic feature vectors includes: Pre-trained deep neural network models are used to extract features from visual modal data to generate feature maps containing texture distribution information. The feature map is used to identify regions of interest and locate the deformation texture features of the damaged parts of the vehicle. Morphological encoding is performed on the deformation texture features to extract the distribution pattern and stress semantic features of the damaged area, and then encoded into a visual semantic feature vector. The visual semantic feature vector is used to characterize the logical attributes of physical impact and damage in a high-dimensional semantic space, serving as a semantic verification parameter in cross-modal retrieval.

6. The method for retrieving auto insurance cases based on semantic analysis according to claim 1, characterized in that: The calculation of semantic logical consistency metrics specifically includes: Extract text semantic vectors representing dynamic behavior from spatiotemporal semantic sequence data; construct a unified hidden layer feature mapping space and map the visual semantic feature vectors and the text semantic vectors into the unified hidden layer feature mapping space; calculate the vector similarity between the mapped visual semantic feature vectors and the text semantic vectors. By normalizing the vector similarity, a semantic logical consistency metric is obtained, which represents the degree of physical and logical consistency between the text description and the visual presentation.

7. The method for retrieving auto insurance cases based on semantic analysis according to claim 1, characterized in that: Reordering the candidate case set specifically includes dynamic weighted query processing; Specifically, this involves parsing the intent of the retrieval request and determining the query configuration parameters; If the query configuration parameters point to logical relevance retrieval, the weight coefficient of the graph embedding vector similarity in the ranking function is increased; if the query configuration parameters point to physical semantic consistency retrieval, the weight coefficient of the semantic logical consistency metric in the ranking function is increased; a comprehensive score is calculated based on the adjusted weight coefficients, and the candidate case set is sorted in descending order according to the comprehensive score.

8. A car insurance case retrieval system based on semantic analysis, characterized in that, include: The data acquisition module acquires multimodal input data of the cases to be retrieved, including unstructured text data and visual modal data; The text parsing module performs structured parsing on unstructured text data, extracts entity objects and attribute parameters, transforms the attribute parameters into spatiotemporal semantic sequence data based on a preset action semantic mapping table, and constructs an event evolution graph structure that represents the evolution logic of the accident; the event evolution graph structure includes physical entity nodes and virtual state nodes generated from spatiotemporal semantic sequence data. The preliminary matching module encodes the event evolution graph structure into the graph neural network model, generates graph embedding vectors for the cases to be retrieved, performs vector similarity search in a pre-set graph vector database, and obtains a preliminary matching candidate case set based on the similarity score. The consistency measurement module performs vectorization processing on visual modal data, extracts visual semantic feature vectors that represent the characteristics of damaged morphology and location, and calculates the semantic logical consistency measurement value between the visual semantic feature vectors and the spatiotemporal semantic sequence data. The ranking optimization module re-ranks the candidate case set based on the weighted fusion result of the similarity score of the graph embedding vector and the semantic logical consistency metric, and outputs the ranked target case list.

Images