CN112101027A - Chinese named entity recognition method based on reading understanding - Google Patents
Chinese named entity recognition method based on reading understanding Download PDFInfo
- Publication number
- CN112101027A CN112101027A CN202010720804.4A CN202010720804A CN112101027A CN 112101027 A CN112101027 A CN 112101027A CN 202010720804 A CN202010720804 A CN 202010720804A CN 112101027 A CN112101027 A CN 112101027A
- Authority
- CN
- China
- Prior art keywords
- index
- document
- document level
- context
- entity
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 28
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 16
- 238000012545 processing Methods 0.000 claims abstract description 10
- 230000011218 segmentation Effects 0.000 claims abstract description 7
- 238000000605 extraction Methods 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 5
- 238000010276 construction Methods 0.000 claims description 3
- 238000003058 natural language processing Methods 0.000 abstract description 7
- 230000000694 effects Effects 0.000 abstract description 5
- 238000012549 training Methods 0.000 description 14
- 230000006870 function Effects 0.000 description 5
- 101001013832 Homo sapiens Mitochondrial peptide methionine sulfoxide reductase Proteins 0.000 description 4
- 102100031767 Mitochondrial peptide methionine sulfoxide reductase Human genes 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 238000002372 labelling Methods 0.000 description 3
- 238000001514 detection method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000008093 supporting effect Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000010835 comparative analysis Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 238000009966 trimming Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
- G06F40/295—Named entity recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention relates to a Chinese named entity recognition method based on reading understanding, and belongs to the technical field of natural language processing. The invention comprises the following steps: performing word segmentation processing on the document level linguistic data to obtain a document level sequence; acquiring a triple composed of a retrieval tag problem, a document level sequence entity and a document level sequence; taking the retrieval tag problem and the document sequence in the triple as input, and generating hidden output of the context information fused into the document level through a BERT coding layer; hidden output of the document level context information is merged into the long-distance context information through a convolutional neural network, semantic features of the long-distance context are obtained, semantic information of the whole document context is captured, and the semantic information is compressed into a feature map; and predicting all entities in the document through a prediction layer by utilizing semantic information of the whole document context, predicting the initial index and the termination index of the entities, and splicing the initial index and the termination index to generate the named entity. The invention can identify the entity in the document and has good identification effect.
Description
Technical Field
The invention relates to a Chinese named entity recognition method based on reading understanding, and belongs to the technical field of natural language processing.
Background
Named Entity Recognition (NER), also known as entity identification, entity segmentation and entity extraction, is a subtask of information extraction that aims to locate and classify named entity mentions in unstructured text into predefined categories such as people. Name, organization, location, medical code, time expression, quantity, monetary value, percentage, etc. This is a fundamental NLP research problem that has been studied for many years. NER is a fundamental key task in NLP. From the flow of natural language processing, NER can be regarded as one of the identification of unknown words in lexical analysis, and is a problem that the number of the unknown words is the largest, the identification difficulty is the largest, and the influence on the word segmentation effect is the largest. Meanwhile, the NER is also the basis of a plurality of NLP tasks such as relation extraction, event extraction, knowledge graph, machine translation, question-answering system and the like.
The NER task is not solved well at present. Although named entity recognition only achieves good effects in limited text types (mainly in news corpora) and entity categories (mainly including names of people, places and organizations), compared with other information retrieval fields, entity naming evaluation is less expected and is easy to generate overfitting, and the system for universally recognizing various types of named entities has poor performance. Especially in the named entity recognition in Chinese, the named entity recognition in Chinese has more challenges compared with English, and the problem which is not solved at present is more. The named entities in english have a relatively obvious formal notation, i.e. the first letter of each word in the entity is capitalized, so the identification of the boundaries of the entities is relatively easy and the task is focused on determining the category of the entity. Compared with English, the Chinese named entity recognition task is more complex, and compared with the entity category labeling subtasks, the recognition of the entity boundary is more difficult. Therefore, the research on Chinese named entity recognition is still a hot problem in the field of natural language processing, and is widely concerned by academia and industry.
Disclosure of Invention
The invention provides a Chinese named entity recognition method based on reading understanding, which is used for solving the problem that the existing recognition method can only recognize entities in sentences.
The technical scheme of the invention is as follows: the Chinese named entity recognition method based on reading understanding comprises the following steps:
step1, performing word segmentation processing on the document level linguistic data to obtain a document level sequence;
step2, splicing the retrieval tag problem, the document level sequence entity and the document level sequence to obtain a triple composed of the retrieval tag problem, the document level sequence entity and the document level sequence;
step3, taking the retrieval tag problem in the triple and the document level sequence as input, and generating hidden output of the context information fused into the document level through a BERT coding layer;
step4, hidden output of the document level context information is merged into is passed through a convolutional neural network, semantic features of the long-distance context are obtained, semantic information of the whole document context is captured, and the semantic information is compressed into a feature map;
step5, predicting all entities in the document through a prediction layer by utilizing semantic information of the whole document context, predicting a start index and an end index of the entities, and splicing the start index and the end index to generate the named entity.
Further, the Step2 includes the specific steps of:
step2.1, dividing the document level linguistic data into words to obtain a document level sequence X ═ X (X)1,x2,x3......xn) Where n represents the length of the sequence, the entities are extracted in a document-level sequence, and each entity is then labeled with a category label y, each represented as xstart,endAssuming that all entity class label sets of the document level sequence are Y, the class label Y corresponding to each entity belongs to Y;
step2.2, retrieval tag problem construction: the problem of constructing the retrieval label of each category label by using the label explanation, namely explaining the category label;
step2.3, converting the document level sequence with the category label into a triple (Question, Answer, Context), wherein Question is a retrieval label problem qyContext is a document level sequence X, and Answer is all document level sequence entities, class tagged entity Xstart,end={xstart,xstart+1,......,xend-1,xendIs a set of subsequences of the document level sequence X satisfying the category label Y, where the subscript "start, end" is used to denote the succession of elements in the sequence from index "start" to index "end", associated with the category label Y e Y;
generating natural language search tag question q by based on category tag yyObtaining a triplet (q)y,xstart,endX), i.e. a triplet (Question, Answer, Context).
Further, in Step3, the search tag problem and the document level sequence in the triplet are used as input, when the sequence length of the search tag problem and the document level sequence is larger than 512 bytes when the search tag problem and the document level sequence pass through the BERT coding layer, the search tag problem and the document level sequence are directly truncated, and the Step2-Step3 are executed by taking the truncated part as a new sample processing loop.
Further, in step4.1, the hidden output merged into the document level context information is input into the convolutional layer of the CNN convolutional neural network to obtain long-distance context semantic features, and the maximum pooling operation is applied to the context semantic features, so that a global feature vector which is fixed in size and independent from the input can be obtained.
Further, in Step5, the prediction layer uses two binary classifiers, one for predicting whether each class label is a start index, and the other for predicting whether each class label is an end index.
Further, the Step5 includes the specific steps of:
step5.1, processing the output representation of the context feature extraction layer, only retaining text information to obtain a representation matrix E'CNN;
Step5.2, then predicting a starting index and an ending index of the named entity through two binary classifiers in a prediction layer; the prediction layer firstly predicts the probability that each entity class label becomes a starting index and an ending index, wherein the probability that each entity class label becomes the starting index is predicted by the prediction layer as follows:
is a learned parameter, d represents the vector dimension of the last layer of the BERT coding layer, N represents the length of the document-level sequence, PstartEach row of (a) means that the probability distribution of each index is the starting or ending position of the entity given the retrieve tag problem;
and Step5.3, predicting the initial index and the end index by using the predicted probability that each entity class label becomes the initial index and the end index, and matching and splicing the predicted initial index and the corresponding end index to generate the named entity.
The invention has the beneficial effects that:
1. the method has better performance on the recognition task of the named entity of the Chinese document;
2. problem coding based on reading understanding is priori knowledge, and the included priori information has a good guiding effect on overall improvement of the named entity model;
3. the entity is usually in a document level text, context content information of a full-text sentence environment also has a certain supporting effect on entity identification, and effective information of the entity in a wider document level context environment is considered, so that the understanding of the model to the context is improved.
Drawings
FIG. 1 is a general model architecture diagram of the present invention;
FIG. 2 is a flow chart of the present invention;
fig. 3 is a BERT context coding diagram.
Detailed Description
Example 1: as shown in fig. 1-3, the method for recognizing a named entity in chinese based on reading understanding includes the following steps:
step1, collecting and sorting a public data set MSRA, and performing word segmentation processing on document level linguistic data to obtain a document level sequence;
in Step1, the public data set MSRA is collected and sorted, and for each type of entity in the MSRA data set, the problem of the retrieval tag of each type is constructed through 'annotation specification', the problem of the retrieval tag can be constructed manually by writing software, and can also be constructed by adopting other methods in the prior art. For example, if the annotator needs to label all entities with category labels as the location LOC, and the label description corresponding to the location LOC is "country, city, and mountain", the problem of the search label constructed by the label description corresponding to the location LOC is "find out abstract or concrete locations such as country, city, and mountain"; as shown in table 1:
table 1 illustrates an example of a search tag problem
Step2, splicing the retrieval tag problem, the document level sequence entity and the document level sequence to obtain a triple composed of the retrieval tag problem, the document level sequence entity and the document level sequence;
step2.1, the document level linguistic data is processed by word segmentation to obtain a document level sequence X ═ X (x1,x2,x3,......xn) Where n represents the length of the sequence, the entities are extracted in a document-level sequence, and each entity is then labeled with a category label y, each represented as xstart,endAssuming that all entity class label sets of the document level sequence are Y, the class label Y corresponding to each entity belongs to Y;
step2.2, retrieval tag problem construction: the problem of constructing the retrieval label of each category label by using the label explanation, namely explaining the category label;
step2.3, converting the document level sequence with the category label into a triple (Question, Answer, Context), wherein Question is a retrieval label problem qyContext is a document level sequence X, and Answer is all document level sequence entities, class tagged entity Xstart,end={xstart,xstart+1,......,xend-1,xendIs a set of subsequences of the document level sequence X satisfying the category label Y, where the subscript "start, end" is used to denote the succession of elements in the sequence from index "start" to index "end", associated with the category label Y e Y;
generating natural language search tag question q by based on category tag yyObtaining a triplet (q)y,xstart,endX), i.e. a triplet (Question, Answer, Context).
Step3, converting the sequence labeling problem into a reading understanding problem, taking the retrieval tag problem and the document level sequence in the triple as input, and generating hidden output of the document level context information by a BERT coding layer; the idea of reading understanding is introduced by the retrieval of the label problem and the text splicing input BERT pre-training model, the document level sequence context is coded, and the problem is introduced as prior information.
Further, in Step3, the search tag problem and the document level sequence in the triplet are used as input, when the sequence length of the search tag problem and the document level sequence is larger than 512 bytes when the search tag problem and the document level sequence pass through the BERT coding layer, the search tag problem and the document level sequence are directly truncated, and the Step2-Step3 are executed by taking the truncated part as a new sample processing loop.
In the conventional method, truncation is performed directly, and the portion beyond the length is discarded directly. In the method, the part with the excess length is selected and reserved and directly treated as a new sample, and in order to prevent entity words from being truncated, rules are also used, so that characters of the same entity are stored in the same sequence sample as much as possible. However, for the test sample, no entity tag is provided by default, direct truncation according to length is adopted, and the sequence beyond the length is still reserved as a new sample. The BERT then receives the combined character string and outputs a representation matrixWhere d is the vector dimension of the last layer of the BERT coding layer;
step4, hidden output of the document level context information is merged into a convolutional neural network (context feature convolutional layer in figure 1), semantic features of long-distance context are obtained, semantic information of the whole document context is captured, and the semantic information is compressed into a feature map;
further, in step4.1, the hidden output merged into the document level context information is input into the convolutional layer of the CNN convolutional neural network to obtain long-distance context semantic features, and the maximum pooling operation is applied to the context semantic features, so that a global feature vector which is fixed in size and independent from the input can be obtained.
Specifically, in order to prevent the loss of context information and effectively acquire the hierarchical features of a long-distance context, the semantic information of the whole context is captured, and valuable semantics are compressed into a feature map, so that the context adopts a convolution layer to acquire the context semantic features, and meanwhile, the maximum pooling operation is applied to the context semantic features, and a global feature vector which is fixed in size and independent from the input can be obtained.
Representation matrix obtained for BERT coding layerConvolution is carried out, then the output of the convolution layer is input into the maximum pooling layer, and the calculation formula is as follows:
ECNN=max(f(ωE+b))
wherein ECNNExtracting a representation matrix for obtaining the context characteristics, wherein b is a bias term, and f is a nonlinear activation function;
step5, predicting all entities in the document through a prediction layer by utilizing semantic information of the whole document context, predicting a start index and an end index of the entities, and splicing the start index and the end index to generate the named entity.
Further, in Step5, the prediction layer uses two binary classifiers, one for predicting whether each class label is a start index, and the other for predicting whether each class label is an end index.
Further, the Step5 includes the specific steps of:
step5.1, processing the output representation of the context feature extraction layer, only retaining text information to obtain a representation matrix E'CNN;
Step5.2, then predicting a starting index and an ending index of the named entity through two binary classifiers in a prediction layer; the prediction layer firstly predicts the probability that each entity class label becomes a starting index and an ending index, wherein the probability that each entity class label becomes the starting index is predicted by the prediction layer as follows:
is a learned parameter, d represents the vector dimension of the last layer of the BERT coding layer, N represents the length of the document-level sequence, PstartEach row of (a) means that the probability distribution of each index is the starting or ending position of the entity given the retrieve tag problem;
and Step5.3, predicting the initial index and the end index by using the predicted probability that each entity class label becomes the initial index and the end index, and matching and splicing the predicted initial index and the corresponding end index to generate the named entity.
In this context, there may be multiple entities of the same category. This means that multiple start indices can be predicted from the start index prediction model and multiple end indices can be predicted from the end index prediction model. A model is further required to match the predicted start index with its corresponding end index. In particular by applying argmax to PstartAnd PendWill obtain the predicted index, i.e. 0-1 sequences of length nAnd
to pairAre each 1, andeach of which is a 1, when satisfying i ≦ j of consecutive characters xi,jIs the predicted entity, PendThe probability of each entity class label becoming a termination index is predicted for the prediction layer.
In training the model of the present invention, the document-level series X and the tag sequence YstartAnd YendPairings, representing the true label of each entity, are the start index or end index of any entity. Thus, for the start andthe end index prediction defines the following two losses, i.e., a start position loss and an end position loss. Data imbalance problems arise due to start and end index prediction, i.e., given search tag problem qyAnd document level sequence X, the beginning or end of which is marked much less than the beginning or end of which is not marked. Therefore, a Loss function, namely, Focal location, of target detection in an image is cited, wherein the Focal location is mainly used for solving the problem that the proportion of positive and negative samples in one-stage target detection is seriously unbalanced. The loss function reduces the weight of a large number of simple negative samples in training, which can also be understood as a difficult sample mining. The definition is as follows:
where y is the label of the true sample and y' is the predictor. Finally, at the time of training, the lossy function is defined as follows:
Lstart=Focal Loss(ps,py)
Lstart=Focal Loss(pe,py)
Final loss=Lstart+Lstart
further, return loss, tuning the model.
The metric value F1 score widely applied to the fields of information retrieval and statistical classification is adopted as an evaluation index. The F1 score is an index used in statistics to measure the accuracy of the two-class model. The method simultaneously considers the accuracy rate and the recall rate of the classification model. The F1 score can be viewed as a harmonic mean of model accuracy and recall with a maximum of 1 and a minimum of 0. Therefore, our experiments used precision, recall, and F1 scores to evaluate the performance of our model.
In order to make the experiment more convincing, two types of baseline models are designed for experimental comparative analysis in order to verify the effectiveness of the reading understanding method. A method for binding BilsTM to CRF (hereinafter abbreviated as BiLSTM + CRF) and a method for binding BilsTM to CRF (hereinafter abbreviated as CNN + BilsTM + CRF) without pre-training are used. Another method for combining BilsTM with CRF (hereinafter BERT + BilsTM + CRF) is pre-trained using BERT sequence labeling trimming (hereinafter BERT) and BERT.
In the experiment, with pre-training, a BERT model is called as a roberta-base character-level Chinese model for Haugh training, the fine tuning parameters are set to be 128 in hidden layer size, 0.25 in dropout, 6 in training round and 5e-6 in learning rate; without pre-training, the hidden layer size is 128, dropout is 0.25, the training round is 6, and the learning rate is 5 e-6. For the baseline model, the training corpus is set to be character-type sentences, and the document-level data is divided into sentence representations.
The results of the identification of our proposed model and baseline model in the MSRA dataset are shown in table 2. It is clear that the method proposed by the present invention achieves optimal results for all data sets with respect to all baseline model P, R, F1 values. The method F1 of the invention reaches 95.72%, the value of F1 without pre-training reaches 85.88% and 87.23%, compared with the method for identifying the Chinese named entity without pre-training, the accuracy is improved by 8.01%, 6.21%, the recall rate is improved by 11.3%, 10.17%, the F1 is improved by 9.84%, 8.49%, and the accuracy and the recall rate are both greatly improved. Compared with the Chinese named entity recognition method with pre-training, the accuracy is improved by 1.11%, 0.3%, the recall rate is improved by 3.51%, 3.31%, and F1 is improved by 3.22%, 2.89%, the recall rate is greatly improved, and the accuracy is slightly improved.
The method for recognizing the Chinese named entity based on reading understanding has better performance on the Chinese named entity recognition task, and mainly has the following reasons: 1. reading and understanding problem codes are priori knowledge, and the priori information contained in the problem codes has a good guiding effect on overall improvement of the named entity model; 2. the entity is usually in a document level text, the context content information of the full-text sentence environment also has a certain supporting function on entity identification, and the invention considers the effective information of the entity in a wider document level context environment, thereby improving the understanding of the model on the context.
Table 2 named entity identification method control experiment F1 values
While the present invention has been described in detail with reference to the embodiments shown in the drawings, the present invention is not limited to the embodiments, and various changes can be made without departing from the spirit of the present invention within the knowledge of those skilled in the art.
Claims (6)
1. The Chinese named entity recognition method based on reading understanding is characterized by comprising the following steps of:
step1, performing word segmentation processing on the document level linguistic data to obtain a document level sequence;
step2, splicing the retrieval tag problem, the document level sequence entity and the document level sequence to obtain a triple composed of the retrieval tag problem, the document level sequence entity and the document level sequence;
step3, taking the retrieval tag problem in the triple and the document level sequence as input, and generating hidden output of the context information fused into the document level through a BERT coding layer;
step4, hidden output of the document level context information is merged into is passed through a convolutional neural network, semantic features of the long-distance context are obtained, semantic information of the whole document context is captured, and the semantic information is compressed into a feature map;
step5, predicting all entities in the document through a prediction layer by utilizing semantic information of the whole document context, predicting a start index and an end index of the entities, and splicing the start index and the end index to generate the named entity.
2. The reading understanding-based Chinese named entity recognition method of claim 1, wherein:
the specific steps of Step2 are as follows:
step2.1, dividing the document level linguistic data into words to obtain a document level sequence X ═ X (X)1,x2,x3,......xn) Where n represents the length of the sequence, the entities are extracted in a document-level sequence, and each entity is then labeled with a category label y, each represented as xstart,endAssuming that all entity class label sets of the document level sequence are Y, the class label Y corresponding to each entity belongs to Y;
step2.2, retrieval tag problem construction: the problem of constructing the retrieval label of each category label by using the label explanation, namely explaining the category label;
step2.3, converting the document level sequence with the category label into a triple (Question, Answer, Context), wherein Question is a retrieval label problem qyContext is a document level sequence X, and Answer is all document level sequence entities, class tagged entity Xstart,end={xstart,xstart+1,......,xend-1,xendIs a set of subsequences of the document level sequence X satisfying the category label Y, where the subscript "start, end" is used to denote the succession of elements in the sequence from index "start" to index "end", associated with the category label Y e Y;
generating natural language search tag question q by based on category tag yyObtaining a triplet (q)y,xstart,endX), i.e. a triplet (Question, Answer, Context).
3. The reading understanding-based Chinese named entity recognition method of claim 1, wherein:
in Step3, the search tag problem and the document level sequence in the triplet are used as input, when the sequence length of the search tag problem and the document level sequence is larger than 512 bytes when the search tag problem and the document level sequence pass through a BERT coding layer, the search tag problem and the document level sequence are directly truncated, and the truncated part is used as a new sample processing loop to execute Step2-Step3.
4. The reading understanding-based Chinese named entity recognition method of claim 1, wherein:
in Step4.1, the hidden output of the document level context information is input into the convolutional layer of the CNN convolutional neural network to obtain long-distance context semantic features, and the maximum pooling operation is applied to the context semantic features, so that the global feature vector which is fixed in size and independent of the input can be obtained.
5. The reading understanding-based Chinese named entity recognition method of claim 1, wherein:
in Step5, the prediction layer uses two binary classifiers, one for predicting whether each class label is a start index and the other for predicting whether each class label is an end index.
6. The reading understanding-based Chinese named entity recognition method of claim 1, wherein:
the specific steps of Step5 are as follows:
step5.1, processing the output representation of the context feature extraction layer, only retaining text information to obtain a representation matrix E'CNN;
Step5.2, then predicting a starting index and an ending index of the named entity through two binary classifiers in a prediction layer; the prediction layer firstly predicts the probability that each entity class label becomes a starting index and an ending index, wherein the probability that each entity class label becomes the starting index is predicted by the prediction layer as follows:
is a learned parameter, d represents the vector dimension of the last layer of the BERT coding layer, N represents the length of the document-level sequence, PstartMeans that the probability distribution of each index is the entity of a given retrieval tag problemA starting position or an ending position of;
and Step5.3, predicting the initial index and the end index by using the predicted probability that each entity class label becomes the initial index and the end index, and matching and splicing the predicted initial index and the corresponding end index to generate the named entity.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010720804.4A CN112101027A (en) | 2020-07-24 | 2020-07-24 | Chinese named entity recognition method based on reading understanding |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010720804.4A CN112101027A (en) | 2020-07-24 | 2020-07-24 | Chinese named entity recognition method based on reading understanding |
Publications (1)
Publication Number | Publication Date |
---|---|
CN112101027A true CN112101027A (en) | 2020-12-18 |
Family
ID=73749901
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010720804.4A Pending CN112101027A (en) | 2020-07-24 | 2020-07-24 | Chinese named entity recognition method based on reading understanding |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112101027A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113033210A (en) * | 2021-05-31 | 2021-06-25 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Drug potential side effect mining method based on social media data analysis |
CN113033206A (en) * | 2021-04-01 | 2021-06-25 | 重庆交通大学 | Bridge detection field text entity identification method based on machine reading understanding |
CN113626577A (en) * | 2021-07-01 | 2021-11-09 | 昆明理工大学 | Chinese cross-language news event element extraction method based on reading understanding |
CN114297987A (en) * | 2022-03-09 | 2022-04-08 | 杭州实在智能科技有限公司 | Document information extraction method and system based on text classification and reading understanding |
CN114328938A (en) * | 2022-03-16 | 2022-04-12 | 浙江卡易智慧医疗科技有限公司 | Image report structured extraction method |
CN114580422A (en) * | 2022-03-14 | 2022-06-03 | 昆明理工大学 | Named entity identification method combining two-stage classification of neighbor analysis |
CN115130435A (en) * | 2022-06-27 | 2022-09-30 | 北京百度网讯科技有限公司 | Document processing method and device, electronic equipment and storage medium |
CN115470354A (en) * | 2022-11-03 | 2022-12-13 | 杭州实在智能科技有限公司 | Method and system for identifying nested and overlapped risk points based on multi-label classification |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108829801A (en) * | 2018-06-06 | 2018-11-16 | 大连理工大学 | A kind of event trigger word abstracting method based on documentation level attention mechanism |
CN110309503A (en) * | 2019-05-21 | 2019-10-08 | 昆明理工大学 | A kind of subjective item Rating Model and methods of marking based on deep learning BERT--CNN |
CN110705272A (en) * | 2019-08-28 | 2020-01-17 | 昆明理工大学 | Named entity identification method for automobile engine fault diagnosis |
CN111046946A (en) * | 2019-12-10 | 2020-04-21 | 昆明理工大学 | Burma language image text recognition method based on CRNN |
CN111126068A (en) * | 2019-12-25 | 2020-05-08 | 中电云脑(天津)科技有限公司 | Chinese named entity recognition method and device and electronic equipment |
CN111177376A (en) * | 2019-12-17 | 2020-05-19 | 东华大学 | Chinese text classification method based on BERT and CNN hierarchical connection |
CN111190997A (en) * | 2018-10-26 | 2020-05-22 | 南京大学 | Question-answering system implementation method using neural network and machine learning sequencing algorithm |
-
2020
- 2020-07-24 CN CN202010720804.4A patent/CN112101027A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108829801A (en) * | 2018-06-06 | 2018-11-16 | 大连理工大学 | A kind of event trigger word abstracting method based on documentation level attention mechanism |
CN111190997A (en) * | 2018-10-26 | 2020-05-22 | 南京大学 | Question-answering system implementation method using neural network and machine learning sequencing algorithm |
CN110309503A (en) * | 2019-05-21 | 2019-10-08 | 昆明理工大学 | A kind of subjective item Rating Model and methods of marking based on deep learning BERT--CNN |
CN110705272A (en) * | 2019-08-28 | 2020-01-17 | 昆明理工大学 | Named entity identification method for automobile engine fault diagnosis |
CN111046946A (en) * | 2019-12-10 | 2020-04-21 | 昆明理工大学 | Burma language image text recognition method based on CRNN |
CN111177376A (en) * | 2019-12-17 | 2020-05-19 | 东华大学 | Chinese text classification method based on BERT and CNN hierarchical connection |
CN111126068A (en) * | 2019-12-25 | 2020-05-08 | 中电云脑(天津)科技有限公司 | Chinese named entity recognition method and device and electronic equipment |
Non-Patent Citations (5)
Title |
---|
JACOB DEVLIN等: "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding", 《ARXIV:1810.04805V2》 * |
JAKE BOUVRIE: "Notes on Convolutional Neural Networks", 《HTTP://WEB.MIT.EDU/JVB/WWW/PAPERS/CNN_TUTORIAL.PDF》 * |
PRATYAY BANERJEE等: "Knowledge Guided Named Entity Recognition for BioMedical Text", 《ARXIV:1911.03869V3》 * |
XIAOYA LI等: "A Unified MRC Framework for Named Entity Recognition", 《ARXIV:1910.11476V6》 * |
刘奕洋等: "基于机器阅读理解的中文命名实体识别方法", 《模式识别与人工智能》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113033206A (en) * | 2021-04-01 | 2021-06-25 | 重庆交通大学 | Bridge detection field text entity identification method based on machine reading understanding |
CN113033210A (en) * | 2021-05-31 | 2021-06-25 | 哈尔滨工业大学(深圳)(哈尔滨工业大学深圳科技创新研究院) | Drug potential side effect mining method based on social media data analysis |
CN113626577A (en) * | 2021-07-01 | 2021-11-09 | 昆明理工大学 | Chinese cross-language news event element extraction method based on reading understanding |
CN114297987A (en) * | 2022-03-09 | 2022-04-08 | 杭州实在智能科技有限公司 | Document information extraction method and system based on text classification and reading understanding |
CN114580422A (en) * | 2022-03-14 | 2022-06-03 | 昆明理工大学 | Named entity identification method combining two-stage classification of neighbor analysis |
CN114328938A (en) * | 2022-03-16 | 2022-04-12 | 浙江卡易智慧医疗科技有限公司 | Image report structured extraction method |
CN115130435A (en) * | 2022-06-27 | 2022-09-30 | 北京百度网讯科技有限公司 | Document processing method and device, electronic equipment and storage medium |
CN115130435B (en) * | 2022-06-27 | 2023-08-11 | 北京百度网讯科技有限公司 | Document processing method, device, electronic equipment and storage medium |
CN115470354A (en) * | 2022-11-03 | 2022-12-13 | 杭州实在智能科技有限公司 | Method and system for identifying nested and overlapped risk points based on multi-label classification |
CN115470354B (en) * | 2022-11-03 | 2023-08-22 | 杭州实在智能科技有限公司 | Method and system for identifying nested and overlapped risk points based on multi-label classification |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112101027A (en) | Chinese named entity recognition method based on reading understanding | |
CN113011533B (en) | Text classification method, apparatus, computer device and storage medium | |
CN109033307B (en) | CRP clustering-based word multi-prototype vector representation and word sense disambiguation method | |
CN112231447B (en) | Method and system for extracting Chinese document events | |
CN108268447B (en) | Labeling method for Tibetan named entities | |
CN113268995B (en) | Chinese academy keyword extraction method, device and storage medium | |
CN115599902B (en) | Oil-gas encyclopedia question-answering method and system based on knowledge graph | |
CN113505200A (en) | Sentence-level Chinese event detection method combining document key information | |
CN112434164B (en) | Network public opinion analysis method and system taking topic discovery and emotion analysis into consideration | |
CN115952292B (en) | Multi-label classification method, apparatus and computer readable medium | |
CN111581943A (en) | Chinese-over-bilingual multi-document news viewpoint sentence identification method based on sentence association graph | |
CN115168541A (en) | Chapter event extraction method and system based on frame semantic mapping and type perception | |
CN114970536A (en) | Combined lexical analysis method for word segmentation, part of speech tagging and named entity recognition | |
CN114881043A (en) | Deep learning model-based legal document semantic similarity evaluation method and system | |
CN114611520A (en) | Text abstract generating method | |
CN110245234A (en) | A kind of multi-source data sample correlating method based on ontology and semantic similarity | |
CN115587163A (en) | Text classification method and device, electronic equipment and storage medium | |
CN114996442A (en) | Text abstract generation system combining abstract degree judgment and abstract optimization | |
CN114943235A (en) | Named entity recognition method based on multi-class language model | |
CN115017404A (en) | Target news topic abstracting method based on compressed space sentence selection | |
CN114491062B (en) | Short text classification method integrating knowledge graph and topic model | |
CN113076750B (en) | Cross-domain Chinese word segmentation system and method based on new word discovery | |
CN117291192B (en) | Government affair text semantic understanding analysis method and system | |
CN112925886B (en) | Few-sample entity identification method based on field adaptation | |
CN115221871B (en) | Multi-feature fusion English scientific literature keyword extraction method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20201218 |