CN115033703A - A relation extraction method for aquatic animals and disease texts - Google Patents

Abstract

The invention discloses a method for extracting entity semantic relationship of aquatic animal disease texts, comprising: collecting aquatic animal disease texts, labeling text data with a labeling tool, inputting the labelled data set into a BERT model, and automatically acquiring the semantic features of words , and represent and extract deep-level semantics, obtain the second text, embed the label information into the joint space of the words and labels of the second text, and carry out joint learning with each word, output the third text, and input the third text into the Bilstm model Carry out learning, obtain the correlation and context information of long-distance words, get the fourth text, send the fourth text to the Attention layer, reduce the loss of key information in the text sequence, obtain the fifth text, and input the fifth text into the CRF layer , and obtain the result of joint extraction of aquatic animal disease text entity relations. This method can effectively solve the problem of inaccurate overlapping relation extraction in text-level relation extraction.

Description

A relation extraction method for aquatic animals and disease texts

technical field

The invention relates to the technical field of aquatic disease prevention, and more particularly to a method for extracting entity semantic relations from aquatic animal disease texts.

Background technique

In the process of aquaculture, diseases of aquatic animals are a major factor affecting the economy of farmers. By combining the knowledge of aquatic animal diseases with computers, a knowledge map of aquatic animal diseases is constructed, so that farmers can get timely information when aquatic diseases occur. Accurate diagnosis, correct and appropriate treatment advice. Relation extraction is one of the important preliminary work of knowledge graph construction. It is of great significance to convert unstructured text into relational data in a unified format and to extract features from text data.

Zheng et al. first proposed a joint entity-relation extraction method based on a new labeling strategy. This method transforms the joint learning model including the two tasks of named entity recognition and relation classification into a sequence labeling problem, and achieves good results (ZHENG S, HAO Y, LU D, et al. Joint entity and relation extraction based on ahybrid neural network[J]. Neurocomputing, 2016, 257.). In the corpus of drug instructions, Zhang Yukun et al. combined convolutional neural network with support vector machine and conditional random field to construct a joint neural network model, and achieved good results (Zhang Yukun, Liu Maofu, Hu Huijun. Based on the joint neural network model Chinese medical entity classification and relation extraction [J]. Computer Engineering and Science, 2019, 41(06):1110-1118.). In the field of rice diseases, insects and weeds, Shen Liyan et al. designed a joint extraction algorithm for the entity relationship between rice diseases, insects and weeds and pesticides based on a new labeling model that combines dual long-term and short-term memory networks with an attention mechanism. The technical problems of clarifying the boundaries and the existence of a large number of multi-relationships between the pesticides and the entities of pests and diseases have achieved good results (Shen Liyan, Jiang Haiyan, Hu Bin, et al. Joint extraction algorithm for the relationship between rice pests, pests and pesticide entities [J]. Journal of Nanjing Agricultural University, 2020, 43(06):1151-1161.). In the financial field, Tang Xiaobo et al. proposed a new sequence labeling mode based on the characteristics of financial texts and constructed a BERT-based entity relationship joint extraction model in the financial field, which realized the identification of overlapping relationships between entities in financial texts, and the F value reached 54.3% (Tang Xiaobo, Liu Zhiyuan. Research on Joint Extraction of Text Sequence Labeling and Entity Relationship in Financial Field [J]. Information Science, 2021,39(05):3-11.). In the medical field, Cao Mingyu et al. proposed a neural network-based method for joint extraction of drug entities and relationships, using a new labeling model to convert the joint extraction of drug entities and relationships into an end-to-end sequence labeling task, with an F value of 67.3% (Cao Mingyu, Yang Zhihao, Luo Ling, et al. Joint extraction of drug entities and relationships based on neural network [J]. Computer Research and Development, 2019, 56(07): 1432-1440.). However, these methods above are limited in capturing entity semantic information in long-span sentences, and they cannot extract some new and effective features from discourse-level relational examples.

SUMMARY OF THE INVENTION

The embodiment of the present invention provides a kind of entity semantic relation extraction method of aquatic animal disease text, including:

Collect aquatic animal disease texts and construct an aquatic animal disease corpus;

Use the labeling tool to label the text data set Use the labeling tool to label the text data;

Input the marked data set into the BERT model, automatically obtain the semantic features of words, express and extract deep semantics, and obtain the second text;

Tag embedding is performed on the second text, the tag information is embedded in the joint space of the words and tags of the second text, and joint learning is performed with each word, and the third text is output;

Input the third text of joint learning into the Bilstm model for learning, further semantically encode the output information of the learned label embedding layer, obtain the correlation and context information of long-distance words, and obtain the fourth text;

The fourth text is sent to the Attention layer, and the useful information is processed with concentrated attention in a large amount of information, the loss of key information in the text sequence is reduced, and the fifth text is obtained;

Input the fifth text into the CRF layer to obtain the final predicted label sequence, and then obtain the result of joint extraction of the entity relationship of aquatic animal disease text.

Further, it also includes data preprocessing on the collected aquatic animal disease texts, including:

Data crawling of aquatic disease websites on the Internet by using Python statements;

Integrate data from literature and books;

Clean useless data.

Further, it also includes dividing the corpus in the corpus into two parts, one part is the training set and the other is the test set, and the text data in the training set is annotated with an annotation tool.

Further, the labeling method for labeling the text data in the training set with labeling tools includes:

The label of the disease is set as a fixed label, B-H-1 represents the head of the entity, and I-H-1 represents the middle part of the entity;

Entity tags all use HB to represent the head of the entity element, HI to represent the middle part of the entity element, and O to indicate that the element does not belong to any entity.

Further, the steps of obtaining the final predicted label sequence include:

Set the input sequence X=(X ₁ , X ₂ , . . . , X _n );

Obtain the output probability matrix P of the Attention layer;

The labeling sequence Y=(Y ₁ , Y ₂ , . . . , Y _n ) output by the CRF layer;

Calculate the predicted sequence score S(X, Y) according to the following formula, and the sequence with the highest score is the final output sequence;

Among them, A _yi,yi+1 represents the probability that the transition matrix in the probability is transferred from the label Y _i to the label Y _i+1 , and P _{i, yi} represents the probability that X _i is labelled as Y _i .

The embodiment of the present invention provides a kind of entity semantic relation extraction method of aquatic animal disease text, compared with prior art, its beneficial effect is as follows:

In order to solve the problem of inaccurate extraction results in text relation extraction in the textual level aquatic products field, the present invention proposes a method for extracting entity semantic relation of aquatic animal disease text, which can effectively solve the inaccurate extraction of overlapping relation in text level relation extraction The problem. Compared with the existing relationship extraction method, the present invention obtains the best performance on the aquatic disease text data set, and improves both the accuracy rate and the recall value, and on this basis, effectively improves the aquatic disease text relationship extraction. the F1 value.

Description of drawings

Fig. 1 is a kind of schematic flowchart of the entity semantic relationship extraction method of aquatic animal disease text provided by the embodiment of the present invention;

Fig. 2 is an annotation result diagram of a method for extracting entity semantic relations of aquatic animal disease texts provided by an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

Referring to Figures 1-2, an embodiment of the present invention provides a method for extracting entity semantic relations from aquatic animal disease text, the method comprising:

A kind of entity semantic relation extraction method of aquatic animal disease text of the present invention, is to send the aquatic animal disease text to be extracted into the network and extract, it is characterized in that described network is constructed according to the steps shown in Figure 1:

Step 1. Collect aquatic animal disease texts, perform data cleaning and data preprocessing on the collected aquatic animal disease texts, and form an aquatic animal disease corpus; by using Python statements to crawl the aquatic disease website on the Internet, the literature and books are collected. For some aquatic animal disease corpus obtained from public accounts and websites, there are some characters, symbols, etc. that are not related to aquatic animal diseases in the crawled pages. Some noise words, such as "previous page", "next page", etc. After converting the downloaded document PDF format into WORD format, there will be typos in the text. It is necessary to talk about these interference characters, interference words, and typo removal and modification. After the integration, the aquatic animal disease corpus is formed.

Step 2. Divide the corpus into two parts, one part is the training set and the other is the test set, and the text data in the training set is annotated with an annotation tool; the annotation method is that this paper studies the relationship between aquatic diseases, so both diseases and The relationship between other entity elements, the label of the disease is set as a fixed label, B-H-1 represents the entity head, and I-H-1 represents the middle part of the entity. The rest of the entity labels use HB to represent the head of the entity element, HI to represent the middle part of the entity element, and O to represent that the element does not belong to any entity. In the labeling, the first letter of the Chinese pinyin is used as the corresponding entity label; the entity role label is represented by numbers "1" and "2", which represent the order of entities in the extracted triples. In this paper, only triples are extracted, and the provisions for the entity role labels of different relationships are: In the aquatic disease field corpus applied in this paper, the entity roles except diseases are represented by "1", and the entity roles of other relationships are represented by "1". 2" to mark. This paper uses H+B/I to represent the overlapping relationship between role 1 and role 2. The labeling results are shown in Figure 1.

Step 3: Use the BERT model to pre-train the marked data set, automatically obtain the semantic features of the words, and express and extract the deep-level semantics;

Step 4: Perform label embedding on the text after pre-training, embed the label information into the joint space of words and labels, and learn with each word. The label information directly contains the information related to the relationship, making full use of the label information can learn better sentence representation and indirectly improve the accuracy of relationship extraction;

Step 5: Send the jointly learned data into the Bilstm model, learn the output information of the label embedding layer for further semantic encoding, and better obtain the correlation and context information of long-distance words;

Step 6: Send the text data learned by the Bilstm model into Attention, which can focus on processing useful information in a large amount of information and reduce the loss of key information in the text sequence;

Step 7: Step 7: Finally, the text data is sent to the CRF model, and the CRF layer improves the accuracy of label prediction by learning the constraints between labels, thereby obtaining the final predicted label sequence.

Let the input sequence be X=(X1, X2,..., Xn), the probability matrix output by the Attention layer is P, and the output label sequence Y=(Y1, Y2,..., Yn). According to formula (1), the predicted sequence score S(X, Y) is calculated, and the sequence with the highest score is the final output sequence.

Among them, Ayi,yi+1 indicates that the transition matrix in the probability is transferred from the label Yi to the probability of labeling Yi+1, and Pi, Yi indicates the probability that Xi is labelled as Yi.

Table 1 shows a comparison diagram of the joint extraction effect of the practice method of the present invention and the entity relationship of the prior art. Table 1 shows the joint extraction results of BiLSTM+CRF model, BERT+BiGRU+CRF model, BERT+BiLSTM+CRF model, and this model from top to bottom. The comparison results show that the methods of BiLSTM+CRF and other models have insufficient effect on the extraction of text-level relations. The model proposed by the present invention is more accurate, and solves the problem of text-level relationship extraction and a large number of overlapping relationships.

Table 1 Extraction result table

The above disclosures are only a few specific embodiments of the present invention. Those skilled in the art can make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the present invention. However, the embodiments of the present invention are not limited thereto. , any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims (5)

1. an entity semantic relation extraction method of aquatic animal disease text, is characterized in that, comprises: Collection of aquatic animal disease texts; Use annotation tools to annotate text data; Input the marked data set into the BERT model, automatically obtain the features of the word semantics, express and extract the deep-level semantics, and obtain the second text; Perform label embedding on the second text, embed the label information into the joint space of the words and labels of the second text, and perform joint learning with each word, and output the third text; Input the third text of joint learning into the Bilstm model for learning, further semantically encode the output information of the learned label embedding layer, obtain the correlation and context information of long-distance words, and obtain the fourth text; Send the fourth text to the Attention layer, focus on processing useful information in a large amount of information, reduce the loss of key information in the text sequence, and obtain the fifth text; Input the fifth text into the CRF layer to obtain the final predicted label sequence, and then obtain the result of joint extraction of semantic relations of aquatic animal disease text entities. 2. The entity semantic relation extraction method of a kind of aquatic animal disease text as claimed in claim 1, is characterized in that, also comprises carrying out data preprocessing to the aquatic animal disease text collected, and it comprises: Data crawling of aquatic disease websites on the Internet by using Python statements; Integrate data from literature and books; Clean useless data. 3. the entity semantic relation extraction method of a kind of aquatic animal disease text as claimed in claim 1, is characterized in that, also comprises the corpus in the corpus is divided into two parts, one part is training set and the other part is test set, adopts labeling tool to Annotate the text data in the training set. 4. the entity semantic relation extraction method of a kind of aquatic animal disease text as claimed in claim 3, is characterized in that, the labeling method that described adopts labeling tool to label the text data in training set, comprising: The label of the disease is set as a fixed label, B-H-1 represents the head of the entity, and I-H-1 represents the middle part of the entity; Entity tags all use HB to represent the head of the entity element, HI to represent the middle part of the entity element, and O to indicate that the entity element does not belong to any entity. 5. a kind of entity semantic relation extraction method of aquatic animal disease text as claimed in claim 1 is characterized in that, the described step of obtaining final predicted label sequence, comprises: Set the input sequence X=(X ₁ , X ₂ , . . . , X _n ); Obtain the output probability matrix P of the Attention layer; The labeling sequence Y=(Y ₁ , Y ₂ , . . . , Y _n ) output by the CRF layer; Calculate the predicted sequence score S(X, Y) according to the following formula, and the sequence with the highest score is the final output sequence;

Among them, A _yi,yi+1 represents the probability that the transition matrix in the probability is transferred from the label Y _i to the label Y _i+1 , and P _{i, yi} represents the probability that X _i is labelled as Y _i .

Images