CN114580389B - Chinese medical field causal relation extraction method integrating radical information - Google Patents

Abstract

The invention discloses a Chinese text causal relation extraction method in the medical field integrating radical information, which relates to the technical field of data mining and comprises the following steps: acquiring a Chinese text data set in the medical field through a web crawler, preprocessing the acquired data, converting English terms in the text into Chinese by adopting a translation technology, acquiring radicals of all characters by utilizing an online Xinhua dictionary, performing incremental training on the radicals by utilizing a Word2Vec architecture to obtain radical characteristic representation, and then taking a radical characteristic vector as input of a causality extraction model to extract causality from the data set to obtain a causality entity. The invention solves the problem of extracting the effective causal relation of the Chinese text data in the medical field. The causal entity of Chinese text data in the medical field can be obtained through the invention.

Description

A causal relationship extraction method in Chinese medical field integrating radical information

Technical Field

The present invention relates to causal relationship extraction in the medical field, and in particular to a causal relationship extraction method in the Chinese medical field integrating radical information.

Background Art

At present, the information construction in the medical field is steadily developing, and modern medical information systems have accumulated massive medical data. With the continuous accumulation of data, the use of natural language processing technology and deep learning methods to mine the rich information contained in medical text data has become a hot spot in the cross-study of the medical field and artificial intelligence. Text data in the medical field contains a large number of records of medical activities, including diseases, drugs, examinations and treatment results. This information is important clinical data. Accurate and efficient analysis and mining of it can provide theoretical and technical support for the establishment of medical knowledge bases and the construction of clinical diagnosis and treatment systems. However, medical text data has many different characteristics from traditional texts, such as containing a large number of English entity names, and semantics are highly correlated with radicals. These characteristics bring new challenges to causal extraction. At this time, a causal extraction method that can integrate radical information and enrich text semantic information is needed.

At present, the research on radical information is mainly focused on the field of named entity recognition. Chinese characters have the characteristics of being able to form words with a single word, and the radicals of Chinese characters often contain important information. The research on radical information is mainly to obtain radical features through conditional random field models, bidirectional long short-term memory network models, etc., integrate radical features into character features, enrich text semantic information, and obtain character feature vector representation that integrates radical information.

The obtained character feature representation of the fused radical information needs to be used as the input of the causal extraction model to obtain the causal entity. For the research on causal extraction, the commonly used methods are machine learning-based methods and deep learning-based methods. The machine learning method first models it as a multi-classification problem, extracts the feature vector, and then uses a supervised classifier for event extraction. With the hot research of neural networks, the application of neural network models to causal extraction can improve the accuracy and efficiency of causal extraction. However, the existing methods rarely consider the radical features of characters, resulting in incomplete semantic information acquisition, which brings risks to the application of causal extraction models in the medical field. This paper extracts causal relationships from medical text data by fusing radical information to improve the accuracy of causal extraction.

Summary of the invention

In order to solve the above problems, the purpose of the present invention is to provide a method for extracting causal relationships in the Chinese medical field by integrating radical information.

In order to achieve the above-mentioned purpose, the present invention provides a method for extracting causal relationships in the Chinese medical field by integrating radical information, which is carried out in the following steps:

Step 1: Data acquisition. Obtain the Chinese medical field text data set D = {D ₁ ,D ₂ ...D _n }, where D _i represents the i-th text, 1≤i≤n, and n is the total number of texts in the set D;

Step 2: Preprocess the acquired text data. The basic steps are as follows:

Step 2.1: Remove stop words, web page tags, etc. from the text and perform word segmentation;

Step 2.2: Extract the text into structured data and load it into the database;

Step 3: Convert the English professional terms in the text data into Chinese. The basic steps are as follows:

Step 3.1: Use ASCII code values to locate English professional terms in the data set;

Step 3.2: Use the translation interface to convert English professional terms into Chinese to obtain a data set containing only Chinese characters;

Step 4: Radical feature acquisition, the basic steps are as follows:

Step 4.1: Obtain the radicals of all characters in the dataset by querying the online Xinhua Dictionary. For Chinese characters without radicals, treat the characters themselves as words.

Step 4.2: Treat the radical as a word and use it as the input of the Word2Vec architecture. Perform incremental training on the radical to obtain the radical feature vector representation.

Step 5: Extracting causal relationships in the Chinese medical field by integrating radical information. The basic steps are as follows:

Step 5.1: Input layer: For the original text data in the Chinese medical field, the sentences are input into the BERT model to obtain character-level features, and the radicals are input into Word2Vec for incremental training to obtain radical feature representation;

Step 5.2: Receive character features and radical features, and output two embedding matrices by searching the embedding dictionary, and set the vector dimensions of characters and radicals to the same size. In this way, a Chinese character can be represented by two vector sequences, namely, a character sequence and a radical sequence;

Step 5.3: The representation layer combines character information with radical information to generate a comprehensive representation of the input text. The bidirectional long short-term memory network can capture the previous and next context information and the bidirectional semantic dependency. Consider concatenating the radical feature as a row vector after the character feature, encode the radical information into the character feature vector, pass the text through the BERT model and the Word2Vec architecture respectively, obtain the character feature and radical feature, and then concatenate these two independent feature vectors to obtain the text feature vector representation fused with the radical information.

Step 5.4: Take the final hidden state of the Bi-LSTM in the representation layer as the output and connect it to form a comprehensive representation. Then input it into the conditional random field model, use the Softmax function as the activation function, map each word to get the conditional probability; finally, use the BIO sequence labeling method to label the output text and get the final extraction result;

Step 6: Sequence labeling. Use sequence labeling to extract causal relationships. Each word in the sentence needs to be marked with a corresponding label. B-cause indicates the beginning of the cause event, B-effect indicates the beginning of the result event, I-cause indicates the middle word or ending word of the cause event, I-effect indicates the middle word or ending word of the result event, and the O label indicates that this word belongs to neither the cause event nor the result event. Perform probability calculation on the sentences in the prediction layer to obtain the causal label corresponding to each character and the causal entity.

The advantages and positive effects of the present invention are: the Chinese medical field causal relationship extraction method integrating radical information of the present invention can integrate character radical information, enrich text semantic information, and use it as the input of the causal relationship extraction model, thereby improving the accuracy of causal relationship extraction, and can be used in the medical field for tasks such as establishing a medical knowledge base and building an online consultation platform.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solution of the present invention, a brief introduction is given to the drawings required for use in the present invention.

FIG1 is a flow chart of a method for extracting causal relationships in the Chinese medical field by integrating radical information provided by the present invention;

FIG. 2 is a structural block diagram of a method for obtaining Chinese character radical features provided by the present invention.

FIG3 is a structural block diagram of a causal relationship extraction method provided by the present invention.

DETAILED DESCRIPTION

The present invention will be further described below:

The purpose of the present invention is to provide a method for extracting causal relationships in the Chinese medical field by integrating radical information. This is a method that achieves better extraction effects by integrating Chinese character radical information and enriching text semantic information on the basis of existing causal relationship extraction.

In conjunction with FIGS. 1 , 2 and 3 , a method for extracting causal relationships in the Chinese medical field by integrating radical information of the present invention is performed in the following steps:

Step 1: Data acquisition. Obtain the Chinese medical field text data set D = {D ₁ ,D ₂ ...D _n }, where D _i represents the i-th text, 1≤i≤n, and n is the total number of texts in the set D;

Step 2: Preprocess the acquired text data. The basic steps are as follows:

Step 2.1: Remove stop words, web page tags, etc. from the text and perform word segmentation;

Step 2.2: Extract the text into structured data and load it into the database;

Step 3: Convert the English professional terms in the text data into Chinese. The basic steps are as follows:

Step 3.1: Use ASCII code values to locate English professional terms in the data set;

Step 3.2: Use the translation interface to convert English professional terms into Chinese to obtain a data set containing only Chinese characters;

Step 4: Radical feature acquisition, the basic steps are as follows:

Step 4.1: Obtain the radicals of all characters in the dataset by querying the online Xinhua Dictionary. For Chinese characters without radicals, treat the characters themselves as words.

Step 4.2: Treat the radical as a word and use it as the input of the Word2Vec architecture. Perform incremental training on the radical to obtain the radical feature vector representation.

Step 5: Extracting causal relationships in the Chinese medical field by integrating radical information. The basic steps are as follows:

Step 5.1: Input layer: For the original text data in the Chinese medical field, the sentences are input into the BERT model to obtain character-level features, and the radicals are input into Word2Vec for incremental training to obtain radical feature representation;

Step 5.2: Receive character features and radical features, and output two embedding matrices by searching the embedding dictionary, and set the vector dimensions of characters and radicals to the same size. In this way, a Chinese character can be represented by two vector sequences, namely, a character sequence and a radical sequence;

Step 5.3: The representation layer combines character information with radical information to generate a comprehensive representation of the input text. The bidirectional long short-term memory network can capture the previous and next context information and the bidirectional semantic dependency. Consider concatenating the radical feature as a row vector after the character feature, encode the radical information into the character feature vector, pass the text through the BERT model and the Word2Vec architecture respectively, obtain the character feature and radical feature, and then concatenate these two independent feature vectors to obtain the text feature vector representation fused with the radical information.

Step 5.4: Take the final hidden state of the Bi-LSTM in the representation layer as the output and connect it to form a comprehensive representation. Then input it into the conditional random field model, use the Softmax function as the activation function, map each word to get the conditional probability; finally, use the BIO sequence labeling method to label the output text and get the final extraction result;

Step 6: Sequence labeling. Use sequence labeling to extract causal relationships. Each word in the sentence needs to be marked with a corresponding label. B-cause indicates the beginning of the cause event, B-effect indicates the beginning of the result event, I-cause indicates the middle word or ending word of the cause event, I-effect indicates the middle word or ending word of the result event, and the O label indicates that this word belongs to neither the cause event nor the result event. Perform probability calculation on the sentences in the prediction layer to obtain the causal label corresponding to each character and the causal entity.

In addition, the above embodiments are only used to illustrate the specific embodiments of the present invention rather than to limit it. Those skilled in the art should understand that some of the technical features therein may be replaced by equivalents, and these modifications and replacements also fall within the scope of protection of the present invention.

Claims (1)

1. A Chinese medical field causal relation extraction method integrating radical information is characterized by comprising the following steps:

Step 1: acquiring data, wherein a text data set D= { D ₁,D₂...D_n},D_i in the Chinese medical field represents an ith text, i is more than or equal to 1 and less than or equal to n, and n is the total number of texts in the set D;

Step 2: the method comprises the following basic steps of:

step 2.1: removing stop words and webpage labels in the text, and performing word segmentation;

step 2.2: extracting the text into structured data, and loading the structured data into a database;

Step 3: the English technical term in the text data is converted into Chinese, and the basic steps are as follows:

Step 3.1: locating English technical terms in the data set by using ASCII code values;

Step 3.2: converting English technical terms into Chinese by using a translation interface to obtain a data set only containing Chinese characters;

Step 4: radical feature acquisition, which basically comprises the following steps:

step 4.1: acquiring radicals of all characters in a data set by inquiring an online Xinhua dictionary, and regarding characters as words for Chinese characters without radicals;

step 4.2: taking the radicals as words, and performing incremental training on the radicals to obtain radical feature vector representation as input of a Word2Vec architecture;

step 5: the extraction of the causal relationship in the Chinese medical field integrating the radical information comprises the following basic steps:

Step 5.1: the input layer is used for inputting sentences into the BERT model to obtain character-level features for original text data in the Chinese medical field, and inputting radicals into Word2Vec to perform incremental training to obtain radical feature representation;

step 5.2: receiving character features and radical features, and outputting two embedding matrixes by searching an embedding dictionary, and setting vector dimensions of the characters and the radicals to be the same size, so that one Chinese character can be represented by two vector sequences, namely a character sequence and a radical sequence;

Step 5.3: the representation layer combines character information and radical information to generate comprehensive representation of an input text, the context information can be captured by utilizing a bidirectional long-short-term memory network, the bidirectional semantic dependence is captured, the radical characteristics are considered to be spliced into character characteristics, then the radical information is encoded into character characteristic vectors, the text is respectively passed through a BERT model and a Word2Vec architecture to obtain character characteristics and radical characteristics, and then the two independent characteristic vectors are spliced to obtain text characteristic vector representation of fused radical information;

Step 5.4: taking the final hidden layer of Bi-LSTM in the representation layer as output, and connecting the final hidden layer to form a comprehensive representation; then inputting the words into a conditional random field model, and mapping each word by adopting a Softmax function as an activation function to obtain conditional probability; finally, marking the output text by using a BIO sequence marking method to obtain a final extraction result;

Step 6: sequence labeling, namely performing causal relation extraction in a sequence labeling mode, namely marking corresponding labels on each word in a sentence, wherein B-cause represents the beginning of a cause event, B-effect represents the beginning of a result event, I-cause represents the middle word or end word of the cause event, I-effect represents the middle word or end word of the result event, O labels represent that the word does not belong to the cause event or the result event, performing probability calculation on sentences of a prediction layer to obtain causal labels corresponding to each character, and obtaining a causal entity.