Chinese nested named entity recognition method fusing font information
Through the method of fusing glyph information, combined with the crossed Biaffine dual affine decoding layer and the diagonal fusion convolutional neural network, the problem of poor effectiveness of the Chinese nested named entity recognition model is solved, and more efficient named entity recognition performance is achieved.
Patent Information
- Application Number
- CN202311851255.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-29
- Publication Date
- 2025-04-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing Chinese nested named entity recognition model is not effective when processing Chinese data, mainly due to the lack of obvious segmentation symbols in Chinese words, resulting in unclear semantic and boundary information.
Using the method of fused glyph information, a nested named entity in text is extracted through the crossed Biaffine dual affine decoding layer and a diagonal fusion convolutional neural network, combining character embedding and font embedding.
It improves the prediction performance of the Chinese nested named entity recognition model, and can more accurately identify and locate named entities in Chinese text.
Smart Images
Figure CN119862885A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a named entity recognition method, in particular to a Chinese nested named entity recognition method integrating glyph information. Background Art
[0002] Named entity recognition is an important task in the field of natural language processing and is the basis of downstream tasks. With the continuous deepening of research on named entity recognition, it is found that traditional flat named entity recognition tasks will ignore the nested entities in real applications. Therefore, nested named entity recognition tasks have become a hot topic in current research. The current best named entity recognition models are mostly focused on English data sets and Chinese data sets. Since Chinese words do not have obvious segmentation symbols, the semantics and boundary information are unclear, which in turn causes poor recognition of Chinese nested named entities. Therefore, how to improve the prediction performance of Chinese nested named entity recognition models is an important research topic. Summary of the invention
[0003] Purpose of the invention: The technical problem to be solved by the present invention is to provide a Chinese nested named entity recognition method that integrates glyph information in view of the deficiencies in the prior art.
[0004] In order to solve the above technical problems, the present invention discloses a Chinese nested named entity recognition method integrating glyph information, comprising the following steps:
[0005] Step 1: prepare a data set and preprocess the text data in the data set to obtain font embedding and character embedding;
[0006] Step 2, designing a named entity recognition model, wherein the named entity recognition model predicts and recognizes entities based on the font embedding and character embedding of the text data obtained in step 1;
[0007] Step 3: Use the data set prepared in step 1 to train the named entity recognition model;
[0008] Step 4: Use the trained named entity recognition model to perform named entity recognition and complete Chinese nested named entity recognition that integrates glyph information.
[0009] Furthermore, the design of the named entity recognition model described in step 2 specifically includes:
[0010] Step 2-1, design a cross-Biaffine double affine decoding layer, fuse the font embedding and the character embedding, and obtain two score matrices respectively, each value in the score matrix corresponds to the probability of whether each span corresponding to the text in the text data is an entity, and fuse the two score matrices to obtain a fused score matrix;
[0011] Step 2-2, designing a diagonal fusion convolutional neural network to obtain the spatial relationship between adjacent spans of the two score matrices;
[0012] Step 2-3: predict and identify the entity based on the fusion score matrix and the above spatial relationship.
[0013] Furthermore, the step 1 of preparing the data set and preprocessing the data in the data set specifically includes:
[0014] Step 1-1, obtaining a Chinese nested data set, wherein the data set at least includes: Chinese text and its corresponding annotated real entity labels;
[0015] Step 1-2, obtaining the font of each character in the above data set, that is, obtaining the font features corresponding to each Chinese character through the font dictionary;
[0016] Step 1-3: Use the existing pre-trained model to convert each Chinese character and its corresponding font features into character embedding E T and font embedding E R .
[0017] Furthermore, the span described in step 2-1 is to segment the text data, and the distance between the head and the tail of any possible entity is the span.
[0018] Furthermore, the cross-Biaffine dual affine decoding layer described in step 2-1 specifically includes the following steps:
[0019] Step 2-1-1, calculate the head and tail matrices corresponding to character embedding and font embedding respectively, as follows:
[0020] H s =Leaky ReLU(E T W s )
[0021] H e =Leaky ReLU(E T W e )
[0022] H Rs =Leaky ReLU(E R W s )
[0023] H Re =Leaky Re LU(E R W E )
[0024] Among them, H s represents the head matrix of character embedding, H erepresents the tail matrix of character embedding, H Rs Represents the head matrix of the font embedding, H Re represents the tail matrix of the font embedding; Leaky Re LU represents the Leaky Re LU activation function, W s and W e is a trainable parameter;
[0025] Step 2-1-2, then cross-input the head and tail matrices corresponding to the character embedding and the font embedding into the multi-head Biaffine double affine decoding layer to obtain two score matrices, as follows:
[0026] Score up =MHBiaffine(H s ,H Re )
[0027] Score down =MHBiaffine(H Rs ,H E )
[0028] Among them, MHBiaffine represents the multi-head bi-affine attention mechanism, Score up Represents the score matrix, Score down represents the score matrix;
[0029] Step 2-1-3, Score up With Score down The score matrix Score of the fusion glyph information and original text information is obtained by adding them together, as follows:
[0030] Score=Score up +Score down
[0031] Among them, Score is the fusion score matrix.
[0032] Furthermore, the spatial relationship described in step 2-2 is obtained by a method comprising the following steps:
[0033] Step 2-2-1, the score matrix Score up And the score matrix Score down Take the upper and lower diagonal parts respectively;
[0034] Step 2-2-2, splicing the upper and lower diagonal parts to obtain a splicing matrix;
[0035] Step 2-2-3, input the concatenated matrix into the three-layer convolutional layer structure of the diagonal fusion convolutional neural network, and output it as Score u, the output result matrix Score u That is, the score matrix Score up And the score matrix Score down The spatial relationship between adjacent spans.
[0036] Furthermore, the entity is predicted and identified according to the fusion score matrix and the above spatial relationship described in step 2-3, which specifically includes the following steps:
[0037] Step 2-3-1, use the fully connected layer to predict the probability P of whether all spans corresponding to the text are entities, as follows:
[0038] P=Sigmoid(W0·(Score+Score u )+b)
[0039] Among them, P represents the prediction probability matrix, W0 is the trainable parameter, Sigmoid is the activation function, and b represents the linear constant;
[0040] Step 2-3-2, take the average value of the symmetrical position of the lower triangular matrix and the upper triangular matrix as the final upper triangular matrix for prediction The details are as follows:
[0041]
[0042] Among them, P ij represents the predicted probability of the i-th row and j-th column in the predicted probability matrix, i≤j;
[0043] Step 2-3-3, finally use the upper triangular matrix Perform predictive recognition on entities.
[0044] Furthermore, the upper triangular matrix described in step 2-3-3 is finally used Predict and identify entities as follows:
[0045] The upper triangular matrix Mapping to named entities in actual text: First, define all spans whose scores do not exceed the judgment threshold as non-entity spans and delete them, and define other spans as entity spans; then judge the label category of the entity span, and select the entity label corresponding to the largest entity score as the entity category of the span.
[0046] Furthermore, the training of the named entity recognition model described in step 3 specifically includes:
[0047] Step 3-1, dividing a training set from the data set described in step 1 for training the named entity recognition model; Step 3-2, inputting the training set into the named entity recognition model designed in step 2, and performing entity prediction and recognition on the text in the data set;
[0048] Step 3-3, use the cross entropy loss function to calculate the loss Loss in the above prediction and recognition, i.e., the training process;
[0049] Step 3-4, optimizing the named entity recognition model according to the above loss Loss.
[0050] Furthermore, the cross entropy loss function described in step 3-3 is used to calculate the loss Loss in the above prediction and recognition, i.e., the training process, as follows:
[0051]
[0052] Among them, y ij Represents the true entity label annotated in the dataset corresponding to the i-th row and j-th column of the score matrix.
[0053] Beneficial effects:
[0054] The Chinese nested named entity recognition method integrating glyph information in the present invention combines Chinese fonts and original text information, and is used to solve the problem that the existing work integrating glyph information can only solve the problem of plane named entity recognition, while being able to improve the performance of the named entity recognition model. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.
[0056] Figure 1 It is the overall structural diagram of the present invention.
[0057] Figure 2 This is a structural diagram of the cross-Biaffine dual affine decoding layer of the present invention. DETAILED DESCRIPTION
[0058] The present invention provides a Chinese nested named entity recognition method integrating glyph information, such as Figure 1 As shown, the following steps are included:
[0059] Step 1: Dataset preparation and processing
[0060] Obtain the Chinese nested dataset X, which includes labeled real entity tags. The font composition of each Chinese character is obtained through the font dictionary. For characters that do not exist in the font dictionary, such as English characters, their font features are uniformly set to ○. For others, the font composition of each Chinese character is obtained through the font dictionary. Use the word embedding weights of the pre-trained model chinese-roberta-wwm-ext proposed by Yiming Cui et al. to convert each Chinese character and its corresponding font into character embeddings E T and font embeddings E R .
[0061] Construct a named entity recognition model as follows:
[0062] Step 2: A large amount of effective information is contained in the Chinese character glyphs. For example, most of the entities that often appear in diseases such as "pain" and "hemorrhoids" have the radical "疒", while the disassembled character information of organ names such as "lung" and "liver" often contains "月". Therefore, it can be considered that adding glyph feature information in the named entity recognition task is helpful for improving the model's performance. Therefore, the present invention proposes a cross Biaffine double affine decoding layer to fuse the font embeddings and character embeddings to obtain two score matrices Score up and Score down , as Figure 2 shown. The specific steps are as follows:
[0063] First, calculate the head and tail matrices H s 、H e 、H Rs 、H Re
[0064] H s =Leaky ReL U(E T W s )
[0065] H e =Leaky Re L U(E T W e )
[0066] H Rs =Leaky Re L U(E R W s )
[0067] H Re =Leaky Re L U(E R W E )
[0068] where Leaky Re L U represents the Leaky Re L U activation function, ET 、E R Respectively represent character embedding and font embedding, W s and W e is a trainable parameter.
[0069] Then H s , H e , H Rs , H Re Cross-input into the multi-head Biaffine double affine decoding layer to obtain two score matrices:
[0070] Score up =MHBiaffine(H s ,H Re )
[0071] Score down =MHBiaffine(H Rs ,H e )
[0072] Among them, MHBiaffine represents the multi-head biaffine attention mechanism, H s , H e , H Rs , H Re They respectively represent the head and tail matrices corresponding to character embedding and glyph embedding, and each value in the score matrix corresponds to the probability of whether the span in the text data is an entity.
[0073] Then Score up With Score down The score matrix Score is obtained by adding the fusion of glyph information and original text information:
[0074] Score=Score up +Score down .
[0075] Step 3: DFCNN (Diagonal Fusion Convolutional Neural Network) is used to model the spatial relationship between adjacent spans in the score matrix. The specific steps are as follows:
[0076] Score up and Score down The upper and lower diagonal parts are spliced and input into a neural network consisting of three convolutional layers to obtain the score. up and Score downThe spatial relationship between adjacent spans (reference: WU S, SONG X, FENG Z. MECT: Multi-metadata embedding based cross-transformer for Chinese named entity recognition [C] / / Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021: 1529-1539.), compared with the reference, this method can simultaneously obtain the spatial information contained in the upper and lower diagonal matrices, thereby obtaining better model performance.
[0077] Step 4: Training and prediction of named entity recognition model. First, train the model, then select the best model after training, and perform named entity recognition on the data in the test set. The specific steps are as follows:
[0078] In named entity recognition (NER), "span" refers to the start and end positions of the named entity in the identified text. When the model performs named entity recognition on the text, it determines the position of the start and end characters of the entity in the text, forming the range or span of an entity. For example, in the sentence "Cell reduction is closely related to the degree of lesions in the lungs and the degree of absorption of inflammatory lesions in the lungs", "lesions in the lungs" is a sym type entity, and its span can be expressed as the position from the first character to the last character. This span information helps the model accurately identify and locate the named entities in the text.
[0079] First, use the fully connected layer to predict the probability P of whether all spans (the starting and ending positions of entities in the text) are entities:
[0080] P=Sigmoid(W0·(Score+Score u )+b)
[0081] Among them, W0 is a trainable parameter, Sigmoid is an activation function, Score is a score matrix that integrates glyph information and original text information, and Score is a matrix that integrates glyph information and original text information. u It is the output of the last layer of the three convolutional layers in the diagonal fusion CNN.
[0082] Then use the cross entropy loss function to calculate the loss during training:
[0083]
[0084] Among them, y ij represents the real entity label in the dataset corresponding to the i-th row and j-th column of the score matrix, P ij Represents the model's predicted probability for each category in the i-th row and j-th column of the span score matrix.
[0085] When predicting entity probability, since the score obtained in the diagonal fusion CNN layer is a matrix symmetric about the main diagonal, P is no longer a symmetric matrix. Therefore, when calculating the final output, it is necessary to calculate the upper triangular part and the lower triangular part at the same time, and take the average of the symmetric positions of the lower triangular matrix and the upper triangular matrix as the upper triangular matrix used for prediction.
[0086]
[0087] Among them, i≤j, the final use Make predictions about entities.
[0088] Will When the score matrix is mapped to the named entities in the actual text, first, all spans whose scores do not exceed the judgment threshold are defined as non-entity spans, and the other spans are defined as entity spans. The non-entity spans are deleted during prediction, and then the label category of the entity span is judged, and the entity label corresponding to the largest entity score is selected as the entity category of the span.
[0089] Example:
[0090] The following is a specific practical case to explain the above technical solution in detail:
[0091] Initial data:"text":"Studies have confirmed that cell reduction is closely related to the degree of lung lesions and the absorption of inflammatory lesions in the lungs.","entities":[{"start_idx":10,"end_idx":10,"type":"bod","entity":"lung"},{"start_idx":10,"end_idx":13,"type":"sym","entity":"lung lesions"},{"start_idx":17,"end_idx":17,"type":"bod","entity":"lung"},{"start_idx":17,"end_idx":22,"type":"sym","entity":"lung inflammatory lesions"}]}
[0092] Obtain the font embeddings of the original dataset through the font dictionary: {"tokens": ["Research has confirmed that cell reduction is closely related to the degree of lung lesions and the absorption degree of inflammatory lung lesions."], "entity_mentions": [{"start": 10, "end": 11, "entity_type": "bod", "text": "lung"}, {"start": 10, "end": 14, "entity_type": "sym", "text": "lung lesions"}, {"start": 17, "end": 18, "entity_type": "bod", "text": "lung"}, {"start": 17, "end": 23, "entity_type": "sym", "text": "inflammatory lung lesions"}],"char_token":[[["stone","","open"]],[["hole","","nine"]],[["word","","right"]],[["roof","","head"]],[["○"]],[["silk","","field"]],[["meat","","package"]],[["ice","","salty"]],[["small","","slash"]],[["one","", "","one"]],[["meat","","market"]],[["gate","","person"]],[["disease","","C"]],[["also","","again"]],[["grain","","show"]],[["wide","","twenty","","again"]],[["nai","","right-falling stroke"]],[["meat","","market"]],[["gate","","person"]],[["fire","","fire"]],[["heart","","life"]],[["disease","","C"]],[["also","","again"]],[["mouth","","and"]],[["jiu","","lash"]],[["grain","","show"]],[["wide","","twenty","","again"]],[["roof","","must","","mountain"]],[["seven","","knife"]],[["wood","","eye"]],[["umbrella","","sky"]],[["○"]]]}
[0093] Use the word embedding weights of the Chinese-Roberta-WWM-EXT pre-trained model proposed by Yiming Cui et al. to convert each Chinese character and its corresponding glyph into a character embedding E T and font embedding E R , where E T index and E R The index represents the characters corresponding to each code:
[0094] E T : [101,4777,4955,6395,2141,8024,5301,5528,1121,2208,680,5511,1079,4567,1359,4923,2428,1350,5511,1079,4142,2595,4567,1359,1429,3119,4923,2428,2166,1147,4685,1068,511,102]
[0095] E T index:[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,0]
[0096] E R :[101,4767,2458,4954,736,6241,3633,100,1928,472,692,4506,5489,1259,1102,1496,2207,716,671,100,671,5489,2354,100,782,4567,688,771,1348,4897,1439,2408,2457,1348 ,718,100,5489,2354,100,782,4125,4125,2552,4495,4567,688,771,1348,1366,1350,100,100,4897,1439,2408,2457,1348,100,2553,2255,673,1143,3312,4680,100,1921,472,102]
[0097] E Rindex:[0,1,1,2,2,3,3,4,4,5,6,6,7,7,8,8,9,9,10,10,10,11,11,12,12,13,13,14,14,15,15,16,16,16,17,17 ,18,18,19,19,20,20,21,21,22,22,23,23,24,24,25,25,26,26,27,27,27,28,28,28,29,29,30,30,31,31,32,0]
[0098] The score matrix obtained by the invented method has a shape of 1*32*32*2, where 1 is the number of data sets, 32 is the sentence length, and 2 is the number of entity types:
[0099] {'scores':tensor([[[[1.0458,0.4844],[1.2314,-0.4975],[0.6121,-1.0001],...,[0.6931,- 0.2952],[0.6219,-0.2165],[0.6986,-0.8591]],[[1.4097,-0.1585],[0.3497,0.4498],[1.359 5,-0.5248],...,[0.5164,0.1519],[0.4313,0.3211],[0.4704,-0.1817]],[[1.7098,-0.5858],[1.1694,-0.9953],[0.6186,0.4171],...,[0.6011,0.0114],[0.4297,0.1762],[0.3377,-0.5997 ]],...,[[0.0289,-0.7709],[-0.2799,-0.4213],[0.4553,-0.3881],...,[-0.0128,0.8370],[1.1252,-0.5223],[0.4457,-0.6938]],[[0.1683,-0.8304],[0.1184,-0.9352],[0.7638,-0.6932 ],...,[1.1330,-1.0635],[0.2389,0.8373],[1.1178,-0.9479]],[[0.7442,-0.5093],[-0.0650,-0.6136],[0.4811,-0.4626],...,[0.9499,-0.6194],[0.8487,-0.8102],[0.2060,0.3694]]]]}
[0100] Then, the entity type with the highest score for each span exceeding the threshold (0.5) is selected as the final predicted entity type, and all spans whose scores do not exceed the threshold are predicted as non-entities.
[0101] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, the invention content of a Chinese nested named entity recognition method integrating glyph information provided by the present invention and some or all steps in each embodiment can be executed. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.
[0102] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on such an understanding, the technical solutions in the embodiments of the present invention can be essentially or partly contributed to the prior art in the form of computer programs, i.e., software products, which can be stored in a storage medium and include several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, an MCU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.
[0103] The present invention provides a method and idea for Chinese nested named entity recognition that integrates glyph information. There are many methods and approaches to implement the technical solution. The above is only a preferred implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention. All components not specified in this embodiment can be implemented using existing technologies.
Claims
1. A Chinese nested named entity recognition method integrating glyph information, characterized in that: The following steps are involved: Step 1: prepare a data set and preprocess the text data in the data set to obtain font embedding and character embedding; Step 2, designing a named entity recognition model, wherein the named entity recognition model predicts and recognizes entities based on the font embedding and character embedding of the text data obtained in step 1; Step 3: Use the data set prepared in step 1 to train the named entity recognition model; Step 4: Use the trained named entity recognition model to perform named entity recognition and complete Chinese nested named entity recognition that integrates glyph information.
2. A Chinese nested named entity recognition method integrating glyph information according to claim 1, characterized in that: Designing a named entity recognition model as described in step 2 specifically includes: Step 2-1, design a cross-Biaffine double affine decoding layer, fuse the font embedding and the character embedding, and obtain two score matrices respectively, each value in the score matrix corresponds to the probability of whether each span corresponding to the text in the text data is an entity, and fuse the two score matrices to obtain a fused score matrix; Step 2-2, designing a diagonal fusion convolutional neural network to obtain the spatial relationship between adjacent spans of the above two score matrices; Step 2-3: predict and identify the entity based on the fusion score matrix and the above spatial relationship.
3. The Chinese nested named entity recognition method integrating glyph information according to claim 2, characterized in that: Prepare the dataset as described in step 1 and preprocess the data in the dataset, including: Step 1-1, obtaining a Chinese nested data set, wherein the data set at least includes: Chinese text and its corresponding annotated real entity labels; Step 1-2, obtaining the font of each character in the above data set, that is, obtaining the font features corresponding to each Chinese character through the font dictionary; Step 1-3: Use the existing pre-trained model to convert each Chinese character and its corresponding font features into character embedding E T and font embedding E R .
4. The Chinese nested named entity recognition method integrating glyph information according to claim 3 is characterized in that: The span described in step 2-1 is to segment the text data, and the distance between the head and the tail of any possible entity is the span.
5. The Chinese nested named entity recognition method integrating glyph information according to claim 4, characterized in that: The cross-Biaffine dual affine decoding layer described in step 2-1 specifically includes the following steps: Step 2-1-1, calculate the head and tail matrices corresponding to character embedding and font embedding respectively, as follows: A s =Leaky ReLU(E T W s ) A e =Leaky ReLU(E T W e ) A Rs =Leaky ReLU(E R W s ) A Re =Leaky ReLU(E R W e ) Among them, H s represents the head matrix of character embedding, H e represents the tail matrix of character embedding, H Rs Represents the head matrix of the font embedding, H Re represents the tail matrix of the font embedding; Leaky ReLU represents the Leaky ReLU activation function, W s and W e is a trainable parameter; Step 2-1-2, then cross-input the head and tail matrices corresponding to the character embedding and the font embedding into the multi-head Biaffine double affine decoding layer to obtain two score matrices, as follows: Score up =MHBiaffine(H s ,H Re ) Score down =MHBiaffine(H Rs ,H e ) Among them, MHBiaffine represents the multi-head bi-affine attention mechanism, Score up Represents the score matrix, Score down represents the score matrix; Step 2-1-3, Score up With Score down The score matrix Socre that integrates the glyph information and the original text information is obtained by adding them together, as follows: Score=Score up +Score down Among them, Score is the fusion score matrix.
6. The Chinese nested named entity recognition method integrating glyph information according to claim 5, characterized in that: The spatial relationship described in step 2-2 is obtained by the method comprising the following steps: Step 2-2-1, convert the score matrix Socre up And the score matrix Score down Take the upper and lower diagonal parts respectively; Step 2-2-2, splicing the upper and lower diagonal parts to obtain a splicing matrix; Step 2-2-3, input the concatenated matrix into the three-layer convolutional layer structure of the diagonal fusion convolutional neural network, and output it as Score u , the output result matrix Score u That is, the score matrix Socre up And the score matrix Score down The spatial relationship between adjacent spans.
7. The Chinese nested named entity recognition method integrating glyph information according to claim 6 is characterized in that: The prediction and recognition of entities according to the fusion score matrix and the above spatial relationship described in step 2-3 specifically includes the following steps: Step 2-3-1, use the fully connected layer to predict the probability P of whether all spans corresponding to the text are entities, as follows: P=Sigmoid(W0·(Score+Score u )+b) Among them, P represents the prediction probability matrix, W0 is the trainable parameter, Sigmoid is the activation function, and b represents the linear constant; Step 2-3-2, take the average value of the symmetrical position of the lower triangular matrix and the upper triangular matrix as the final upper triangular matrix for prediction The details are as follows: Among them, P ij represents the predicted probability of the i-th row and j-th column in the predicted probability matrix, i≤j; Step 2-3-3, finally use the upper triangular matrix Perform predictive recognition on entities.
8. The Chinese nested named entity recognition method integrating glyph information according to claim 7, characterized in that: The final upper triangular matrix described in step 2-3-3 is used Predict and identify entities as follows: The upper triangular matrix Mapping to named entities in actual text: First, define all spans whose scores do not exceed the judgment threshold as non-entity spans and delete them, and define other spans as entity spans; then judge the label category of the entity span, and select the entity label corresponding to the largest entity score as the entity category of the span.
9. The Chinese nested named entity recognition method integrating glyph information according to claim 8, characterized in that: The training of the named entity recognition model described in step 3 specifically includes: Step 3-1, dividing a training set from the data set described in step 1 for training the named entity recognition model; Step 3-2, inputting the training set into the named entity recognition model designed in step 2, and performing entity prediction and recognition on the text in the data set; Step 3-3, use the cross entropy loss function to calculate the loss Loss in the above prediction and recognition, i.e., the training process; Step 3-4, optimizing the named entity recognition model according to the above loss Loss.
10. The Chinese nested named entity recognition method integrating glyph information according to claim 9, characterized in that: The cross entropy loss function described in step 3-3 is used to calculate the loss Loss in the above prediction recognition, i.e., training process, as follows: Among them, y ij Represents the true entity label annotated in the dataset corresponding to the i-th row and j-th column of the score matrix.