Medical event extraction method based on three-dimensional tensor
The event extraction model based on the three-dimensional tensor labeling method and the two-stage structural optimization strategy solves the problems of label combination expansion and difficulty in capturing semantic features in traditional methods, and improves the accuracy and efficiency of medical event extraction.
Patent Information
- Application Number
- CN202510715904.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-30
- Publication Date
- 2025-09-12
AI Technical Summary
Existing medical event extraction methods have problems with the expansion of the number of label combinations and limited model performance when dealing with complex overlapping and nested events. Especially in the overlapping event extraction task, it is difficult for the model to accurately capture semantic features.
A three-dimensional tensor labeling method is adopted to solve the ambiguity problem of many-to-many mapping between trigger words and arguments through a refined cell mapping mechanism. An event extraction model with a two-stage structure and joint optimization strategy is designed to improve labeling accuracy and parsing efficiency.
It significantly improves the annotation accuracy and parsing efficiency of complex event structures, and enhances the model's ability to recognize deep dependencies between trigger words and arguments.
Smart Images
Figure CN120633787A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of natural language processing technology, specifically to the field of information extraction technology, and more specifically to providing a medical event extraction method based on three-dimensional tensor. Background Art
[0002] Medical event extraction is a key task in medical text information extraction. Its core goal is to automatically extract event types, trigger words, and related argument information from large amounts of unstructured medical text. With the rapid growth of medical data such as electronic medical records, clinical reports, and medical literature, the demand for medical event extraction in areas such as medical information extraction, intelligent diagnosis, and personalized treatment is increasing. Extracting structured event information from electronic medical records and building a medical event database can not only significantly improve doctors' diagnostic efficiency but also provide a solid data foundation for intelligent diagnosis and personalized medicine. Furthermore, structured medical event data is also of great value for scientific research analysis, patient profiling, and disease progression tracking.
[0003] While various medical event extraction methods have achieved some success, they still face significant challenges in dealing with complex overlapping and nested events. Overlapping events occur when the same argument appears in multiple events, while nested events occur when an argument from one event also serves as a trigger in another. Traditional event extraction methods, such as sequence tagging strategies based on BERT-softmax and BERT-CRF, have achieved some success in general event extraction but still have significant limitations when dealing with such complex structures.
[0004] To address this issue, researchers have proposed table-filling methods in recent years and achieved superior performance. While these methods improve extraction coverage and flexibility, they still suffer from limitations in accuracy. Current table-filling methods typically use a two-dimensional event-role matrix for annotation. This matrix requires labeling multiple trigger-word-argument pairs, leading to a rapid expansion of the number of label combinations and a large number of invalid label combinations, which in turn increases decoding time and impacts model performance. Furthermore, in overlapping event extraction tasks, a single element in the matrix may correspond to multiple labels, making it difficult for the model training process to accurately capture semantic features, further limiting performance improvements.
[0005] In summary, this field urgently needs to design a medical event extraction method based on three-dimensional tensors to solve the above problems. Summary of the Invention
[0006] In view of this, the present invention provides a medical event extraction method based on three-dimensional tensors. First, a three-dimensional model labeling method is designed. Through a refined cell mapping mechanism, the ambiguity problem of many-to-many mapping between trigger words and arguments in traditional sequence labeling methods is resolved, thereby improving labeling accuracy and parsing efficiency. Second, a medical event extraction structural model based on a three-dimensional model is constructed. Through a functionally clear two-stage structure and a joint optimization strategy, the model's ability to capture the complex dependencies between event trigger words and arguments is improved.
[0007] Definition 1: Trigger word-argument pair three-dimensional tensor. Given a three-dimensional tensor Where n is the length of the input text, and m is the total number of possible combinations of trigger word and argument length. Each subject word-argument pair uniquely corresponds to an element in a three-dimensional tensor, labeled (s_h, a_h, [s_len, a_len]). The first and second dimensions represent the trigger word's starting position s_h and the argument's starting position a_h, respectively. The third dimension represents the combination of the trigger word length s_len and the argument length a_len.
[0008] Definition 2: Event type-argument role three-dimensional tensor. Given a three-dimensional tensor Where n represents the length of the input text, and l is the number of predefined event type-argument roles. Each event triple can be represented by one or more elements in the event type-argument role three-dimensional tensor. For example, Figure 2 The event triplet "(nodule, r1, inside spleen)" shown can be jointly represented by the four elements "(nodule, r1, spleen)", "(nodule, r1, inside)", "(node, r1, spleen)" and "(node, r1, inside)" in the event type-argument role three-dimensional tensor, where r1 is the predefined event type-argument role "location", indicating that the occurrence location of the nodule is in the spleen.
[0009] The technical route of the present invention is implemented in the following way: first, a pre-trained language model is used as an encoder to vectorize the input text. Secondly, a three-dimensional model of trigger word-argument pairs is constructed, the joint probability of all trigger word and argument combinations is calculated, and candidate trigger word-argument pairs are screened out according to a set threshold. Then, based on the above candidate pairs, a three-dimensional model of event type-argument role is constructed, an independent matrix is generated for each role category, and semantic representations are extracted through a shared encoder and average pooling mechanism, and the probability that the corresponding trigger word-argument pair belongs to the role is calculated. Finally, an event triple extractor is constructed, which comprehensively considers the matching probability of the trigger word-argument pairs and each role to generate the final structured event result.
[0010] This invention takes medical event extraction as an example, and the specific steps are as follows:
[0011] S1. Use the pre-trained language model as an encoder to vectorize the input text;
[0012] S2. Construct a three-dimensional model of trigger word-argument pairs, calculate the joint probability of all trigger word-argument combinations, and select candidate trigger word-argument pairs based on a set threshold;
[0013] S3: Based on the candidate trigger word-argument pairs obtained in S2, a three-dimensional model of event type-argument role is constructed. A separate matrix is generated for each role category. The semantic representation is extracted through a shared encoder and average pooling mechanism, and the probability of the corresponding trigger word-argument pair belonging to the role is calculated.
[0014] S4. Construct an event triple extractor, comprehensively consider the matching probability of trigger word-argument pairs and each role, and generate the final structured event result.
[0015] Furthermore, step S1 includes:
[0016] S11, add two special mark symbols at the beginning and end of the input text C to obtain C′;
[0017] S12. Input C′ into the pre-trained model to obtain the vectorized representation H:
[0018] H=PLM(C′),
[0019] Among them, PLM stands for pre-trained model.
[0020] Furthermore, step S2 includes:
[0021] S21. Define the three-dimensional model of trigger word-argument pairs as Construct a matrix set with a dimension of m, where m is the product of the maximum length of the trigger word and the argument in the training set. Each matrix For the i-th matrix, each element Represents the matching probability of the trigger word starting at position j and the argument starting at position k in the current context. The specific length is s_len and a_len represented by the i-th matrix. The calculation process of the representation is as follows:
[0022]
[0023] Among them, the trainable weight parameters are W1, W2 and b1, b2, R k The rotation matrix that introduces relative position information;
[0024] S22. Define the loss function of the trigger word-argument pair extractor Used to learn the parameters in the extractor:
[0025]
[0026] Among them, Ω = {(j,k)|1≤j,k≤n} is the set of all possible trigger word-argument pairs in the i-th matrix, Contains all true trigger-argument pairs;
[0027] S23. Set a probability threshold and use the trigger word-argument pair extractor to filter out candidate trigger word-argument pairs (s_h, a_h, [s_len, a_len]) with a probability higher than the threshold from the model;
[0028] S24. Convert the obtained (s_h, a_h, [s_len, a_len]) into a quadruple with clearer entity boundary information.
[0029] Furthermore, step S3 includes:
[0030] S31. Define the event type-argument role three-dimensional model as Where l represents the number of predefined role categories, a matrix set with dimension l is constructed, each matrix
[0031] S32. In each matrix M′, the candidate trigger word-argument pairs are annotated, and the semantic representation M′ of the trigger word-argument pairs in all matrices is extracted through a shared text encoder and average pooling mechanism. p , the calculation formula is as follows:
[0032]
[0033] Among them, M′ i,j Specifically corresponding to the vector of the i-th trigger word token and the j-th argument token;
[0034] S33, M' p Input to the multilayer perceptron to calculate the association probability Pr between the trigger word-argument pair and different roles, which is calculated as follows:
[0035]
[0036] Among them, W r and b r is a learnable weight matrix and σ represents the activation function.
[0037] Furthermore, step S4 includes:
[0038] S41. Define a triplet extractor, using the same loss function as S22 To optimize the extractor parameters;
[0039] S42. In the reasoning stage, only the roles with the highest scores exceeding the set threshold are retained, and they are combined with the corresponding trigger word-argument pairs to form the final structured event triples.
[0040] After adopting the above strategy, the positive effects of the present invention are:
[0041] (1) Aiming at the ambiguity problem caused by the many-to-many mapping between trigger words and arguments, this paper proposes an innovative three-dimensional tensor labeling method. Through a refined cell mapping mechanism, this method solves the ambiguity problem of the many-to-many mapping between trigger words and arguments in traditional sequence labeling methods, thereby improving labeling accuracy and parsing efficiency.
[0042] (2) Aiming at the deep dependency problem between trigger words and arguments in complex event structures, the present invention designs an event extraction model with a two-stage structure and a joint optimization strategy, which significantly enhances the model's ability to model and recognize complex semantic roles. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 This is a flow chart of the medical event extraction method based on three-dimensional tensor of the present invention.
[0044] Figure 2 This is a schematic diagram of the overall framework of the medical event extraction method based on three-dimensional tensor of the present invention DETAILED DESCRIPTION
[0045] In order to enable those skilled in the art to better understand the present invention and to make the above-mentioned objectives, technical solutions and advantages of the present invention more obvious and understandable, the present invention is further described in detail below in conjunction with embodiments and accompanying drawings.
[0046] See Figure 1 , Figure 1 It is a flow chart of the medical event extraction method based on three-dimensional tensors of the present invention, which includes: first, using a pre-trained language model as an encoder to vectorize the input text. Secondly, constructing a three-dimensional model of trigger word-argument pairs, calculating the joint probability of all trigger word and argument combination pairs, and screening out candidate trigger word-argument pairs based on a set threshold. Then, based on the above candidate pairs, constructing a three-dimensional model of event type-argument role, generating an independent matrix for each role category, and extracting semantic representations through a shared encoder and average pooling mechanism, and calculating the probability that the corresponding trigger word-argument pair belongs to the role. Finally, constructing an event triple extractor, comprehensively considering the matching probability of the trigger word-argument pairs and each role, and generating the final structured event result.
[0047] See Figure 2 , Figure 2 This is a schematic diagram of the overall framework of the medical event extraction method based on three-dimensional tensors of the present invention, and the specific steps include:
[0048] S1. Use the pre-trained language model as the encoder to vectorize the input text.
[0049] S11, when inputting text C=[c1,c2,...,c n ]Add two special mark symbols at the beginning and end <cls>and <sep>, we get C′=[ <cls>,c1,c2,...,c n , <sep>];
[0050] S12. Input C′ into the pre-trained model to obtain the vectorized representation H:
[0051] H=PLM([ <cls>,c1,c2,...,c n , <sep>]),
[0052] Among them, PLM stands for the pre-trained model BERT.
[0053] S2. Construct a three-dimensional model of trigger word-argument pairs, calculate the joint probability of all trigger word and argument combinations, and screen out candidate trigger word-argument pairs based on the set threshold.
[0054] S21. Define the three-dimensional model of trigger word-argument pairs as Construct a matrix set with a dimension of m, where m is the product of the maximum length of the trigger word and the argument in the training set. Each matrix For the i-th matrix, each element Represents the matching probability of the trigger word starting at position j and the argument starting at position k in the current context. The specific length is s_len and a_len represented by the i-th matrix. The calculation process of the representation is as follows:
[0055]
[0056] Among them, the trainable weight parameters are W1, W2 and b1, b2, R k The rotation matrix that introduces relative position information;
[0057] S22. Define the loss function of the trigger word-argument pair extractor Used to learn the parameters in the extractor:
[0058]
[0059] Among them, Ω = {(j,k)|1≤j,k≤n} is the set of all possible trigger word-argument pairs in the i-th matrix, Contains all true trigger-argument pairs;
[0060] S23. Set a probability threshold and use the trigger word-argument pair extractor to filter out candidate trigger word-argument pairs (s_h, a_h, [s_len, a_len]) with a probability higher than the threshold from the model;
[0061] S24. Convert the obtained (s_h, a_h, [s_len, a_len]) into a four-tuple (s_h, s_t, a_h, a_t) with clearer entity boundary information, where s_h and s_t represent the starting and ending character positions of the trigger word, respectively, and a_h and a_t represent the starting and ending character positions of the argument, respectively.
[0062] S3. Based on the candidate trigger word-argument pairs obtained in S2, a three-dimensional model of event type-argument role is constructed, an independent matrix is generated for each role category, and semantic representations are extracted through a shared encoder and average pooling mechanism to calculate the probability that the corresponding trigger word-argument pair belongs to the role.
[0063] S31. Define the event type-argument role three-dimensional model as Where l represents the number of predefined role categories, a matrix set with dimension l is constructed, each matrix
[0064] S32. In each matrix M′, the candidate trigger word-argument pairs (s_h, s_t, a_h, a_t) are labeled, and the semantic representation M′ of the trigger word-argument pairs in all matrices is extracted through a shared text encoder and average pooling mechanism. p , the calculation formula is as follows:
[0065]
[0066] Among them, M′ i,j Specifically corresponding to the vector of the i-th trigger word token and the j-th argument token;
[0067] S33, M' p Input to the multilayer perceptron to calculate the association probability P between the trigger word-argument pair and different roles r , calculated as follows:
[0068]
[0069] Among them, W r and b r is a learnable weight matrix and σ represents the sigmoid activation function.
[0070] S4. Construct an event triple extractor, comprehensively consider the matching probability of trigger word-argument pairs and each role, and generate the final structured event result.
[0071] S41. Define a triplet extractor, using the same loss function as S22 To optimize the extractor parameters;
[0072] S42. In the reasoning stage, only the roles with the highest scores exceeding the set threshold λ are retained, and they are combined with the corresponding trigger word-argument pairs to form the final structured event triples.
[0073] Specific embodiments of the present invention have been described above with reference to the accompanying drawings. However, those skilled in the art will appreciate that various modifications and substitutions may be made to the specific embodiments of the present invention without departing from the spirit and scope of the present invention. Such modifications and substitutions are intended to fall within the scope of the claims.< / sep> < / cls> < / sep> < / cls> < / sep> < / cls>
Claims
1. A medical event extraction method based on three-dimensional tensor, characterized by The following steps are involved: S1. Use the pre-trained language model as an encoder to vectorize the input text; S2. Construct a three-dimensional model of trigger word-argument pairs, calculate the joint probability of all trigger word-argument combinations, and select candidate trigger word-argument pairs based on a set threshold; S3: Based on the candidate trigger word-argument pairs obtained in S2, a three-dimensional model of event type-argument role is constructed. A separate matrix is generated for each role category. The semantic representation is extracted through a shared encoder and average pooling mechanism, and the probability of the corresponding trigger word-argument pair belonging to the role is calculated. S4. Construct an event triple extractor, comprehensively consider the matching probability of trigger word-argument pairs and each role, and generate the final structured event result.
2. A medical event extraction method based on three-dimensional tensor according to claim 1, characterized in that: The step S1 specifically includes: S11, when inputting text C=[c1,c2,...,c n ]Add two special mark symbols at the beginning and end <cls>and <sep>, we get C′=[ <cls>,c1,c2,...,c n , <sep> ]；< / sep> < / cls> < / sep> < / cls> S12. Input C′ into the pre-trained model to obtain the vectorized representation H: H=PLM([ <cls>,c1,c2,...,c n , <sep> ]),< / sep> < / cls> Among them, PLM stands for the pre-trained model BERT.
3. A medical event extraction method based on three-dimensional tensor according to claim 1, characterized in that: The step S2 specifically includes: S21. Define the three-dimensional model of trigger word-argument pairs as Construct a matrix set with a dimension of m, where m is the product of the maximum length of the trigger word and the argument in the training set. Each matrix For the i-th matrix, each element Represents the matching probability of the trigger word starting at position j and the argument starting at position k in the current context. The specific length is s_len and a_len represented by the i-th matrix. The calculation process of the representation is as follows: Among them, the trainable weight parameters are W1, W2 and b1, b2, R k The rotation matrix that introduces relative position information; S22. Define the loss function of the trigger word-argument pair extractor Used to learn the parameters in the extractor: Among them, Ω = {(j,k)|1≤j,k≤n} is the set of all possible trigger word-argument pairs in the i-th matrix, Contains all true trigger-argument pairs; S23. Set a probability threshold and use the trigger word-argument pair extractor to filter out candidate trigger word-argument pairs (s_h, a_h, [s_len, a_len]) with a probability higher than the threshold from the model; S24. Convert the obtained (s_h, a_h, [s_len, a_len]) into a four-tuple (s_h, s_t, a_h, a_t) with clearer entity boundary information, where s_h and s_t represent the starting and ending character positions of the trigger word, respectively, and a_h and a_t represent the starting and ending character positions of the argument, respectively.
4. A medical event extraction method based on three-dimensional tensor according to claim 1, characterized in that: The step S3 specifically includes: S31. Define the event type-argument role three-dimensional model as Where l represents the number of predefined role categories, a matrix set with dimension l is constructed, each matrix S32. In each matrix M′, the candidate trigger word-argument pairs (s_h, s_t, a_h, a_t) are labeled, and the semantic representation M′ of the trigger word-argument pairs in all matrices is extracted through a shared text encoder and average pooling mechanism. p , the calculation formula is as follows: Among them, M′ i,j Specifically corresponding to the vector of the i-th trigger word token and the j-th argument token; S33, M' p Input to the multilayer perceptron to calculate the association probability P between the trigger word-argument pair and different roles r , calculated as follows: Among them, W r and b r is a learnable weight matrix and σ represents the sigmoid activation function.
5. A medical event extraction method based on three-dimensional tensor according to claim 1, characterized in that: The step S4 specifically includes: S41. Define a triplet extractor, using the same loss function as S22 To optimize the extractor parameters; S42. In the reasoning stage, only the roles with the highest scores exceeding the set threshold λ are retained, and they are combined with the corresponding trigger word-argument pairs to form the final structured event triples.