Semantic understanding and automatic evidence extraction method and system for medical data appeal text

Abstract

The invention relates to the field of automatic evidence extraction, and discloses a semantic comprehension and automatic evidence extraction method and system for a medical appeal text, and the method comprises the steps: constructing a three-stage encoder related to the medical data appeal text, carrying out the semantic analysis and feature extraction from a character stage, an entity stage and a document stage, and obtaining multi-granularity semantic information, fusing the multi-granularity semantic information, and establishing a multi-granularity semantic fusion structure model; based on the multi-granularity semantic fusion structure obtained by training, constructing a dynamic space-time evidence graph to dynamically adjust the edge weight; through intelligent mining of the key evidence chain, multi-modal evidence is verified and contradictions are eliminated.

Description

Technical Field

[0001] The present invention relates to the technical field of automated evidence extraction, and particularly to a semantic understanding and automated evidence extraction method and system for pharmaceutical data appeal texts. Background Art

[0002] In the prior art, pharmaceutical data appeal texts usually contain complex medical terms, entity relationships, and temporal information. Traditional text processing methods may be difficult to comprehensively understand and analyze this information. In the medical appeal scenario, the complexity of the text includes, but is not limited to, multi-level semantics, temporal contradictions, and cross-modal conflicts, which pose higher requirements for the model. Due to problems such as single-granularity parsing, static knowledge dependence, and evidence chain breakage, the misjudgment rate of traditional models remains high. The prior art cannot handle complex texts. Traditional graph neural networks (GNNs) fix the weights of entity relationships. For example, the weight of the "drug - indication" edge is constantly 0.9 and cannot reflect the time decay effect, such as the failure risk after postoperative antibiotic use for more than 24 hours, and ignore pharmacokinetic parameters, such as the impact of the half-life on the association strength. Therefore, there is a need to provide a semantic understanding and automated evidence extraction method and system for pharmaceutical data appeal texts. Summary of the Invention

[0003] The purpose of the present invention is to provide a semantic understanding and automated evidence extraction method and system for pharmaceutical data appeal texts. To solve the above problems of the prior art, the present invention is achieved through the following technical solutions:

[0004] In a first aspect, a semantic understanding and automated evidence extraction method for pharmaceutical data appeal texts includes the following steps:

[0005] Construct a three-level encoder for pharmaceutical data appeal texts, and perform semantic analysis and feature extraction from the character level, entity level, and document level to obtain multi-granularity semantic information;

[0006] Fuse the multi-granularity semantic information to establish a multi-granularity semantic fusion structure model;

[0007] Based on the trained multi-granularity semantic fusion structure, construct a dynamic spatio-temporal evidence graph to dynamically adjust the edge weights;

[0008] Through intelligent mining of key evidence chains, verify and eliminate contradictions in multi-modal evidence.

[0009] In a second aspect, the semantic understanding and automated evidence extraction system for pharmaceutical data appeal texts provided by the present invention specifically includes the following modules:

[0010] Data collection module: collect relevant data of pharmaceutical data appeal texts;

[0011] Feature extraction module: Semantically analyze and extract features from relevant data of medical data appeal texts at the character level, entity level, and document level respectively;

[0012] Semantic fusion module: Fuse multi-granularity semantic information to establish a multi-granularity semantic fusion structure model;

[0013] Dynamic evidence graph construction module: Based on the trained multi-granularity semantic fusion structure, construct a dynamic spatio-temporal evidence graph and dynamically adjust the edge weights;

[0014] Evidence verification and optimization module: Verify multi-modal evidence and eliminate contradictions through intelligent mining of key evidence chains.

[0015] Advantages of the present invention:

[0016] 1. Through semantic analysis and feature extraction using encoders at three different granularities of character level, entity level, and document level, the entity-level encoder identifies medical entities and constructs a graph structure to mine the relationship information between entities; the document-level encoder highlights important sentences and information through a hierarchical attention mechanism, captures the temporal features of the disease course evolution, fuses multi-granularity features, and can more accurately reflect the semantics of the text than a single-granularity feature extraction method; the granularity feature fusion mechanism integrates multi-level information from the fine-grained character level to the coarse-grained document level, reduces the risk of overfitting, thereby improving the generalization ability and stability of the model, and accurately performs feature extraction and semantic analysis on medical data appeal texts with different formats and different expression habits; through multi-granularity feature extraction and fusion, and coding mechanisms specifically designed for different granularities, it can better process complex texts, improve the model's understanding ability of medical data appeal texts, and thus more accurately complete related tasks;

[0017] 2. Generate composite vectors through a multi-granularity semantic fusion structure, integrate character-level, entity-level, and document-level features, combine cosine similarity and pharmacokinetics, dynamically adjust the edge weights, retain semantic associations, quantify the time decay effect, achieve an all-round semantic representation of medical events, introduce the Boolean value of the recommended path in the NCCN guidelines, integrate the authority of clinical guidelines into the evidence chain construction, and through dynamic spatio-temporal evidence graph and key evidence chain mining, realize multi-dimensional correlation analysis, temporal logic verification, and guideline-driven reasoning of medical data, provide technical support for constructing a reliable and interpretable medical evidence system, and achieve high-precision and high-efficiency automated evidence extraction in the medical data appeal scenario. Brief description of the drawings

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0019] Figure 1 is a flowchart of the steps of the method for semantic understanding and automated evidence extraction of medical data appeal texts provided in Embodiment 1 of the present invention;

[0020] Figure 2 is a schematic structural diagram of the system for semantic understanding and automated evidence extraction of medical data appeal texts provided in Embodiment 2 of the present invention. Detailed implementation manners

[0021] In order to enable those skilled in the art to better understand the solutions of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all 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] Embodiment 1

[0023] As Figure 1 shown, the method for semantic understanding and automated evidence extraction of medical data appeal texts provided in the embodiments of the present invention specifically includes the following steps:

[0024] Step 1: Construct a three-level encoder for medical data appeal texts, perform semantic analysis and feature extraction from the character level, entity level, and document level respectively, and fuse multi-granularity semantic information to establish a multi-granularity semantic fusion structure model;

[0025] In Step 1:

[0026] Specifically, the specific process of constructing the character-level encoder for medical data appeal texts is as follows:

[0027] Perform convolution calculation on the collected medical data appeal text data using the Hybrid Convolutional Neural Network (Hybrid-CNN), adjust the parameters of the one-dimensional convolutional kernel, and perform max pooling operations to optimize the extraction effect of character-level features;

[0028] It should be noted that the hybrid convolutional neural network represents a model design method that combines different types of convolutional operations or structures, aiming to improve the model performance, enhance the feature extraction ability, and optimize the computational efficiency by integrating multiple convolutional methods. The one-dimensional convolutional kernel represents a filter or weight matrix used to perform convolutional operations on one-dimensional data. Max pooling represents selecting the maximum value within a local area of a given feature map, usually a rectangular window, as the output.

[0029] Specifically, based on the hybrid convolutional neural network capturing the morphological features of medical terms such as drug names, medical terms have specific morphological patterns. For example: "Aspirin Enteric-coated Tablets" or "Cefixime Capsules". By analyzing the features at the character level, the meaning in the medical data appeal text can be better understood.

[0030] Specifically, character-level encoding is performed on the medical data appeal text to obtain the encoding vector corresponding to each character. If the input text has a characters and the encoding vector dimension of each character is d, then the input data can be represented as an a×d matrix X.

[0031] Based on capturing character features of different lengths, w different sizes of one-dimensional convolutional kernels are used. The sizes of the one-dimensional convolutional kernels are preset as: k1, k2, k3,..., k w ; For each one-dimensional convolutional kernel of size k i , where i ∈ {1, 2, 3,..., w}, specifically represents a matrix W i of size k i × d, where d is the dimension of the character encoding vector.

[0032] Exemplarily, one-dimensional convolutional kernels of different sizes are used. For example: k1 = 3, k2 = 4, k3 = 5, where k1, k2, and k3 are the sizes of the one-dimensional convolutional kernels, and convolutional operations are performed on the character sequence; each one-dimensional convolutional kernel captures character features of different sizes. For example, the one-dimensional convolutional kernel of size 3 captures the common suffix feature "-pirin" of drug names.

[0033] Exemplarily, for the text "The patient had discomfort after taking aspirin" in the medical data appeal text, after character-level encoding, a character sequence is obtained. A one-dimensional convolutional kernel of size 3 is used for convolutional operations to simplify the character encoding into a simple numerical vector representation corresponding to each character; after encoding the character sequence, the encoding vectors of the characters are as shown in Table 1:

[0034] character encoding vector suffer from [0.1,0.2,0.3] patient [0.4,0.5,0.6] take [0.7,0.8,0.9] use [0.1,0.1,0.1] a [0.2,0.2,0.2] sp [0.3,0.3,0.3] ir [0.4,0.4,0.4] in [0.5,0.5,0.5] after [0.6,0.6,0.6] appear [0.7,0.7,0.7] uncomfortable [0.8,0.8,0.8] convolution position [0.9,0.9,0.9] convolution result vector [1.0,1.0,1.0]

[0035] Table 1 Statistical table of encoding vectors corresponding to characters

[0036] Using a one-dimensional convolutional kernel of size 3, the weight matrix W of the one-dimensional convolutional kernel is: Slide the one-dimensional convolution kernel on the character sequence, and take the encoding vector of three consecutive characters each time to perform convolution operation with the one-dimensional convolution kernel;

[0037] For each size k i The one-dimensional convolution kernel W i , perform convolution operation on the input matrix X, the one-dimensional convolution kernel slides on X, and takes consecutive k i The encoding vector of the characters and W i Perform matrix multiplication and add a bias term b i , and perform nonlinear transformation through the activation function GELU function, setting the one-dimensional convolution kernel W i Perform convolution operation at the jth position and obtain the convolution result c i,j for: in Represents the input matrix X from row j to row j+k i -1 row submatrix, W i T Represents the one-dimensional convolution kernel W i The transposed matrix of size k i After the convolution operation, a one-dimensional convolution kernel with a length of n+k is obtained. i +1 convolution result vector C i ;

[0038] It should be noted that for each one-dimensional convolution kernel, in addition to the weight matrix, i.e. the convolution kernel itself, a corresponding bias term is set to increase the degree of freedom of the model, so that the model can flexibly fit the training data;

[0039] For example, when the one-dimensional convolution kernel is applied to the first three characters “患”, “者”, and “服”, the convolution result vector is calculated as follows: [0.1, 0.2, 0.3]*0.1+[0.4, 0.5, 0.6]*0.4+[0.7, 0.8, 0.9]*0.7=[0.66, 0.78, 0.9];

[0040] The one-dimensional convolution kernel slides over the entire character sequence. After the convolution operation, the convolution results are shown in Table 2:

[0041] pooling position pooling result vector 1 [0.66,0.78,0.9] 2 [0.5,0.6,0.7] 3 [0.4,0.5,0.6] 4 [0.3,0.4,0.5] 5 [0.2,0.3,0.4] 6 [0.1,0.2,0.3] 7 [0.05,0.1,0.15] 8 [0.02,0.03,0.04] 9 [0.01,0.02,0.03] 10 [0.005,0.01,0.015] 11 [0.002,0.003,0.004]

[0042] Table 2 Convolution result vector corresponding to the convolution position

[0043] Perform a maximum pooling operation on the convolution result to extract the most important features of each one-dimensional convolution kernel output. Use a pooling window of size N (N is a positive integer and N is greater than or equal to 2). The pooling window slides on the convolution result and takes the maximum value of each dimension in the window as the output each time.

[0044] It should be noted that the pooling window size is preset by the professionals in the field of the present invention based on historical experience and can be adjusted according to specific circumstances;

[0045] Exemplarily, when using a pooling window of size 2 and the pooling window acts on the first two convolution results [0.66, 0.78, 0.9] and [0.5, 0.6, 0.7], the maximum value is taken for each dimension to obtain the pooling result vector: [max(0.66, 0.5), max(0.78, 0.6), max(0.9, 0.7)] = [0.66, 0.78, 0.9]. The pooling window continues to slide, and the convolution result vectors calculated based on Table 2 are successively max-pooled to obtain all the pooling result vectors, as shown in Table 3:

[0046] Figure 1 Figure 2 1 [0.66,0.78,0.9] 2 [0.4,0.5,0.6] 3 [0.2,0.3,0.4] 4 [0.05,0.1,0.15] 5 [0.01,0.02,0.03]

[0047] Table 3 Statistical Table of Max-Pooling Result Vectors

[0048] It should be noted that through the max-pooling operation, the most important features in each pooling window are extracted from the convolution results, reducing the dimension of the data while retaining the key information. The obtained pooling results are used for subsequent feature splicing to finally obtain the character-level feature representation;

[0049] Specifically, based on the obtained encoding vector, convolution operations are performed using multiple one-dimensional convolution kernels of different sizes, and max-pooling operations are performed on the results of each one-dimensional convolution kernel. All the pooled features are spliced together to capture richer character-level features, obtaining the character-level feature representation C, whose dimension is denoted as d C ;

[0050] Second specifically, the specific process of constructing the entity-level encoder for the pharmaceutical data appeal text is as follows:

[0051] Using the named entity recognition NER technology to identify medical entities from the pharmaceutical data appeal text, including but not limited to: drug entity D, disease entity I, and symptom entity S;

[0052] It should be noted that the named entity recognition NER technology represents an important task in natural language processing, aiming to identify and classify specific types of entities from the text;

[0053] Taking the identified entities as nodes, edges between entities are constructed according to the semantic information in the text to form a graph structure G=(V, E), where V is the set of nodes and E is the set of edges;

[0054] Exemplarily, if the text mentions that "drug A can treat disease B", then an edge from drug A to disease B is added in the graph;

[0055] Use the Gated Graph Neural Network (Gated-GNN) for information propagation to update the feature representation of nodes. For node v i , its feature update formula is: Where is the feature representation of node v i at the t-th moment, represents the update gate, which is used to control the fusion ratio of the old feature and the new feature . ⊙ represents element-wise multiplication;

[0056] It should be noted that the Gated Graph Neural Network (Gated-GNN) represents a neural network model specifically designed for processing graph-structured data. In the scenario of processing medical data appeal texts, the text contains various medical entities, such as drugs, diseases, and symptoms. There are complex relationships between entities. The complex relationships between entities can be abstracted into a graph structure, where nodes represent entities and edges represent the relationships between entities. By processing this graph-structured data through the Gated Graph Neural Network (Gated-GNN), the inference of the relationships between medical entities is realized;

[0057] Based on the Gated Graph Neural Network, the feature representation of each entity is obtained, which contains the relationship information between entities. The average pooling of all entity features is performed to obtain the entity-level feature representation E entity , whose dimension is denoted as d E ;

[0058] Specifically, the specific process of constructing the document-level encoder for medical data appeal texts is as follows:

[0059] Based on the obtained medical data appeal text, through the Hierarchical Attention mechanism, important sentences and information are highlighted, the characteristics of the disease course evolution are captured, the time sequence is obtained, and the time series information of the disease course evolution is obtained;

[0060] It should be noted that the Hierarchical Attention mechanism is a model architecture specifically designed for processing long documents or sequence data. By applying the attention mechanism at different levels, important information in the text is captured, and different parts are assigned different weights;[[ID=�5]]

[0061] Specifically, the medical data appeal text is divided into x sentences, and each sentence is encoded using the pre-trained language model BERT; the attention mechanism is applied at the sentence level, and the importance weight α of each sentence is calculated through the formula where exp represents the exponential function, and u x represents the sentence s q , qFeature representation represents the sentence feature representation u q transpose of, v s represents the attention vector of the sentence;

[0062] Apply the attention mechanism at the document level, combine the attention results at the sentence level, calculate the feature representation of the entire document, through the formula obtain the document-level document-level feature representation D doc , whose dimension is denoted as d D ;

[0063] Fourth Specifically, the specific process of fusing multi-granularity semantic information and establishing a multi-granularity semantic fusion structure is as follows:

[0064] Based on the obtained character-level feature representation C, entity-level feature representation E entity and document-level feature representation D doc , generate a multi-granularity semantic fusion structure T through a multi-granularity feature fusion mechanism sem : T sem =[B char , B entity , B doc ∈R a×d×3 , where represents the number of characters, i.e., the length of the medical data appeal text, R represents the set of real numbers, B char represents the feature matrix of the character-level feature representation C, B entity represents the entity-level feature representation E entity 's feature matrix, B doc represents the feature matrix of the document-level feature representation D doc ;

[0065] It should be noted that the multi-granularity feature fusion mechanism represents a technology used to integrate features at different levels in a deep learning model. By combining multi-level information from fine-grained to coarse-grained, it improves the model's ability to understand input data and task performance. The semantic information of the multi-granularity semantic fusion structure allows the model to learn richer feature representations, better adapt to various changes when facing different types and domains of text data, reduce the risk of overfitting, and thus improve the generalization ability and stability of the model;

[0066] Based on the obtained multi-granularity semantic fusion structure T sem , capture the word formation rules of drug names, construct a "drug→indication→contraindication" dynamic knowledge graph, and establish a temporal causal chain of "medication→examination→symptom deterioration" to construct a complete evidence chain;

[0067] The technical solution of the embodiment of the present invention is as follows: semantic analysis and feature extraction are performed through encoders at three different granularities, namely character level, entity level, and document level, to comprehensively capture information in the medical data appeal text from different perspectives, analyze the word-formation characteristics of medical terms such as drug names, and better understand the meaning of medical terms; the entity-level encoder identifies medical entities and constructs a graph structure to mine the relationship information between entities; the document-level encoder highlights important sentences and information through a hierarchical attention mechanism, captures the temporal characteristics of the disease course evolution, and fuses multi-granularity features, enabling the model to learn richer and more comprehensive feature representations, and being more accurate in reflecting the semantics of the text than a single-granularity feature extraction method; the granularity feature fusion mechanism integrates multi-level information from the fine-grained character level to the coarse-grained document level, enabling the model to better adapt to text data of different types and domains, learn and reason using rich semantic information, reduce the risk of overfitting, thereby improving the generalization ability and stability of the model, and accurately performing feature extraction and semantic analysis for medical data appeal texts with different formats and expression habits; through multi-granularity feature extraction and fusion, and encoding mechanisms specifically designed for different granularities, complex texts can be better processed, the understanding ability of the model for medical data appeal texts can be improved, and relevant tasks can be completed more accurately;

[0068] Embodiment 2

[0069] As Figure 1 shown, the semantic understanding and automated evidence extraction method for medical data appeal texts provided by the embodiment of the present invention specifically includes the following steps:

[0070] Step 2: Based on the trained multi-granularity semantic fusion structure, construct a dynamic spatio-temporal evidence graph to dynamically adjust the edge weights, and through the intelligent mining of key evidence chains, verify and eliminate contradictions in multi-modal evidence;

[0071] In Step 2:

[0072] Specifically, the specific process of dynamically constructing a spatio-temporal evidence graph based on the trained multi-granularity semantic fusion structure is as follows:

[0073] Based on the obtained multi-granularity semantic fusion structure, construct a 256-dimensional composite vector, and through the weighted calculation formula g i =β1F char +β2F entity +β3F doc obtain the semantic composite vector g i , where F char represents the character-level feature vector, F entity represents the entity-level feature vector, F docIt represents the document-level feature vector, and β1, β2, and β3 represent the preset weight coefficients, with values of 0.4, 0.3, and 0.3 respectively;

[0074] It should be noted that

[0075] For example, the vector of the event “ceftriaxone 2 g used on postoperative day 3” contains drug name, surgery time window, and administration route features;

[0076] Based on the calculated semantic composite vector g i and g j , through the formula Calculate the cosine similarity s between semantic composite vectors sem , where g i and g j All are semantic composite vectors;

[0077] The cosine similarity s sem Compared with the preset strong correlation threshold, if the cosine similarity s sem If the value is greater than or equal to the preset strong correlation threshold, the semantic composite vector g i and g j The corresponding medical data complaint text is marked as strongly associated;

[0078] Based on drug metabolism kinetics, the half-life of the drug is obtained by the formula Calculate the time difference weight w time , where e represents the natural constant, λ represents the preset drug time decay factor, and t i It represents the actual half-life of the drug, t j It represents the theoretical half-life of the drug;

[0079] Based on the obtained time difference weight, the cosine similarity between the obtained semantic composite vectors, the time difference weight and the Boolean value of the recommended path of the latest version of the NCCN guidelines are multiplied to obtain the final edge weight w ij ;

[0080] The second specific process of verifying and eliminating contradictions in multimodal evidence through intelligent mining of key evidence chains is as follows:

[0081] Based on the multi-granularity semantic fusion structure, a unit time window is set to obtain medical events that occur within the same unit time window, and the cosine similarity between the semantic composite vectors obtained by the medical event analysis is obtained;

[0082] If there is a strong correlation between medical events A and B, set the time threshold specified in the guidelines, calculate the difference between the actual time of medical event A and the time threshold specified in the guidelines, and take the absolute value to obtain the time deviation;

[0083] The obtained time deviation is processed by taking the ratio with the time threshold specified in the guideline to obtain the time deviation degree;

[0084] The obtained time deviation degree is compared with a preset deviation threshold. If the time deviation degree is greater than or equal to the deviation threshold, it indicates a logical error and is marked as a logical error;

[0085] Based on the logical error obtained from medical event A, the logical error is fed back, and a strongly associated medical event B is obtained, and it is judged whether a logical error occurs;

[0086] If a logical error is obtained for medical event B, the evidence corresponding to medical events A and B is eliminated and an abnormal feedback is made. If no logical error exists, the evidence corresponding to the strongly associated medical event B is retained, the evidence corresponding to medical event A is eliminated, and medical event A is fed back for verification;

[0087] Exemplarily, comparing the text record "fever occurred on the 5th day of immunotherapy" with the test data "CRP elevation occurred on the 3rd day", it is obtained that the time deviation degree of the text record "fever occurred on the 5th day of immunotherapy" is greater than the deviation threshold, then a logical error is marked. It is obtained that the time deviation degree of the test data "CRP elevation occurred on the 3rd day" is less than the deviation threshold, then no logical error exists. The evidence corresponding to the test data "CRP elevation occurred on the 3rd day" is retained, the evidence corresponding to the text record "fever occurred on the 5th day of immunotherapy" is eliminated, and the text record "fever occurred on the 5th day of immunotherapy" is fed back for verification;

[0088] The technical solution of the embodiment of the present invention is: generating a composite vector through a multi-granularity semantic fusion structure, integrating character-level, entity-level, and document-level features, combining cosine similarity and pharmacokinetics, dynamically adjusting edge weights, retaining semantic associations, quantifying the time decay effect, realizing an all-round semantic representation of medical events, introducing the Boolean value of the NCCN guideline recommended path, integrating the authority of clinical guidelines into the evidence chain construction, and realizing multi-dimensional correlation analysis, temporal logic verification, and guideline-driven reasoning of medical data through dynamic spatio-temporal evidence graph and key evidence chain mining, providing technical support for constructing a reliable and interpretable medical evidence system, and realizing high-precision and high-efficiency automated evidence extraction in the scenario of pharmaceutical data appeal.

[0089] Embodiment 3

[0090] As Figure 2 It should be noted that there may be some inaccuracies in the original text such as "sp", "ir", "in" which seem incorrect. This translation is based on the best understanding of the provided content. shown, the semantic understanding and automated evidence extraction system for pharmaceutical data appeal text provided by the embodiment of the present invention specifically includes the following modules:

[0091] Data collection module: collecting relevant data of pharmaceutical data appeal text;

[0092] Feature extraction module: Semantically analyze and extract features from relevant data of pharmaceutical data appeal texts at the character level, entity level, and document level respectively;

[0093] Semantic fusion module: Fuse multi-granularity semantic information to establish a multi-granularity semantic fusion structure model;

[0094] Dynamic evidence graph construction module: Based on the trained multi-granularity semantic fusion structure, construct a dynamic spatio-temporal evidence graph and dynamically adjust the edge weights;

[0095] Evidence verification and optimization module: Verify multi-modal evidence and eliminate contradictions through intelligent mining of key evidence chains.

[0096] The above has described an embodiment of the present invention in detail, but the content described is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention; the above formulas are all dimensionless and take their numerical calculations. The formulas are obtained by collecting a large amount of data for software simulation to obtain a formula closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation and historical experience and can be adjusted according to the actual situation; all equal changes and improvements made according to the scope of the present invention application should still fall within the scope covered by the patent of the present invention.

Claims

1. A method for semantic understanding and automated evidence extraction of pharmaceutical data appeal texts, characterized in that It includes the following steps: Construct a three-level encoder for medical data appeal texts, perform semantic analysis and feature extraction from the character level, entity level, and document level respectively to obtain multi-granularity semantic information; Fuse the multi-granularity semantic information to obtain a semantic composite vector, and establish a multi-granularity semantic fusion structure model; Based on the trained multi-granularity semantic fusion structure, construct a dynamic spatio-temporal evidence graph to dynamically adjust the edge weights to obtain the final edge weights; Through the intelligent mining of key evidence chains, verify and eliminate contradictions in multi-modal evidence.

2. The semantic understanding and automated evidence extraction method for pharmaceutical data appeal texts according to claim 1, characterized in that The method for obtaining the character-level features is as follows: Use the Hybrid-CNN (Hybrid Convolutional Neural Network) to perform convolution calculations on the collected medical data appeal text data, adjust the parameters of the one-dimensional convolution kernel, and perform max pooling operations to optimize the extraction effect of character-level features.

3. The semantic understanding and automated evidence extraction method for pharmaceutical data appeal texts according to claim 1, characterized in that The method for obtaining the entity-level features is as follows: Use the named entity recognition (NER) technology to identify medical entities from the medical data appeal text, take the identified entities as nodes, construct edges between entities according to the semantic information in the text, and calculate and update the nodes to form a graph structure G=(V, E), where V is the set of nodes and E is the set of edges.

4. The semantic understanding and automated evidence extraction method for pharmaceutical data appeal texts according to claim 3, characterized in that, The method for obtaining the updated nodes is as follows: Use the Gated Graph Neural Network (Gated-GNN) for information propagation to update the feature representation of nodes. For node v i , its feature update formula is: where is the feature representation of node v i at the t-th moment, represents the update gate, which is used to control the fusion ratio of the old feature and the new feature , and ⊙ represents element-wise multiplication.

5. The semantic understanding and automated evidence extraction method for pharmaceutical data appeal texts according to claim 1, characterized in that The method for obtaining the entity-level feature representation is as follows: Based on the gated graph neural network, the feature representation of each entity is obtained, which contains the relationship information between entities. The average pooling of all entity features is performed to obtain the entity-level feature representation, and its dimension is denoted as d E 。 6. The semantic understanding and automated evidence extraction method for pharmaceutical data appeal texts according to claim 1, characterized in that, The method for obtaining the semantic composite vector is as follows: Based on the obtained multi-granularity semantic fusion structure, construct a 256-dimensional composite vector and calculate the cosine similarity, and obtain the semantic composite vector through a weighted calculation formula.

7. The semantic understanding and automated evidence extraction method for the pharmaceutical data appeal text according to claim 6, characterized in that The method for obtaining the cosine similarity is as follows: Based on the calculated semantic composite vector g i and g j , the cosine similarity between the semantic composite vectors is calculated through a formula; Compare the obtained cosine similarity with a preset strong correlation threshold to determine whether there is a strong correlation.

8. The semantic understanding and automated evidence extraction method for pharmaceutical data appeal texts according to claim 1, characterized in that, The method for obtaining the final edge weights is as follows: Based on pharmacokinetics, obtain the half-life of the drug, and calculate the time difference weight through a formula; Based on the obtained time difference weight, multiply the cosine similarity between the obtained semantic composite vectors, the time difference weight, and the boolean value of the recommended path in the latest version of the NCCN guidelines to obtain the final edge weights.

9. The semantic understanding and automated evidence extraction method for pharmaceutical data appeal texts according to claim 1, characterized in that The specific process of contradiction elimination is as follows: Based on the multi-granularity semantic fusion structure, set a unit time window, obtain the medical events that occur within the same unit time window, and obtain the cosine similarity between the semantic composite vectors obtained from the analysis of the medical events; Based on the obtained logical errors, feedback the logical errors, and obtain the medical events with strong correlations, and determine whether there are logical errors and eliminate contradictions.

10. The semantic understanding and automated evidence extraction system for pharmaceutical data appeal texts according to claim 1, characterized in that, It includes: Data collection module: Collect relevant data of medical data appeal texts; Feature extraction module: Perform semantic analysis and feature extraction on the relevant data of medical data appeal texts from the character level, entity level, and document level respectively; Semantic fusion module: Fuse the multi-granularity semantic information and establish a multi-granularity semantic fusion structure model; Dynamic evidence graph construction module: Based on the trained multi-granularity semantic fusion structure, construct a dynamic spatio-temporal evidence graph to dynamically adjust the edge weights; Evidence verification and optimization module: Through the intelligent mining of key evidence chains, verify and eliminate contradictions in multi-modal evidence.

Images