KR102461295B1 - A method for normalizing biomedical names - Google Patents

Abstract

Building training data by processing a known database and unclassified data about biomedical terms, constructing a regularization model by applying a word representation based on deep learning to the built training data, and the and assigning an identifier by applying the test data to which the identifier is to be assigned to the constructed normalization model.

Description

{A method for normalizing biomedical names}

The present invention relates to a method for normalizing biomedical related subject names.

In biomedical literature, the technology for identifying entity names (so-called named entity recognition, NER) technology is an essential element in extracting the knowledge we want from various documents. Text mining using biomedical literature In the field, it is important to normalize the identified name to a standardized identifier after the entity name recognition technique is applied.

In general, many entity name canonicalization methods rely on specific domain dictionaries to solve the problem of identifying synonyms and abbreviations. However, dictionaries do not cover various object names except for some object names such as genes. Recently, since biomedical literature is rapidly accumulating, a neural network-based algorithm that integrates a large amount of data is being applied to various text mining techniques.

McCandless, M., Hatcher, E., Gospodnetic, O.: Lucene in Action: Covers Apache Lucene 3.0. Manning Publications Co., New York (2010) Mikolov, T., Chen, K., Corrado, G., Dean, J.: Ecient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013) Leaman, R., Dogan, R.I., Lu, Z.: Dnorm: disease name normalization with pairwise learning to rank. Bioinformatics, 474 (2013)

We propose an accurate normalization method for various entity names than the existing entity name normalization method.

Building training data by processing the existing database and unclassified data on biomedical terms, constructing a regularization model by applying a deep learning-based word representation to the built training data; and assigning an identifier by applying test data to which an identifier is to be assigned to the configured normalization model.

According to an embodiment of the present invention, more accurate entity name normalization than the existing entity name normalization method is possible.

1 schematically shows a method for normalizing an identifier of a biomedical entity name according to an embodiment of the present invention.
2 is a block diagram illustrating an algorithm according to an embodiment of the present invention in more detail.
3 shows a comparison of the performance of the model proposed in the present invention and the existing entity name normalization model, DNorm, with and without an abbreviation resolution step.
4 shows a comparison of performances between models for plant name normalization proposed in the present invention.
5 is a flowchart illustrating a biomedical term identification method according to an embodiment of the present invention.

Hereinafter, the embodiments disclosed in the present specification will be described in detail with reference to the accompanying drawings, but the same or similar components are assigned the same reference numerals regardless of reference numerals, and redundant description thereof will be omitted. The suffixes "module" and "part" for components used in the following description are given or mixed in consideration of only the ease of writing the specification, and do not have distinct meanings or roles by themselves. In addition, in describing the embodiments disclosed in the present specification, if it is determined that detailed descriptions of related known technologies may obscure the gist of the embodiments disclosed in this specification, the detailed description thereof will be omitted. In addition, the accompanying drawings are only for easy understanding of the embodiments disclosed in this specification, and the technical idea disclosed herein is not limited by the accompanying drawings, and all changes included in the spirit and scope of the present invention , should be understood to include equivalents or substitutes.

Terms including ordinal numbers such as first, second, etc. may be used to describe various elements, but the elements are not limited by the terms. The above terms are used only for the purpose of distinguishing one component from another.

When a component is referred to as being “connected” or “connected” to another component, it may be directly connected or connected to the other component, but it is understood that other components may exist in between. it should be On the other hand, when it is said that a certain element is "directly connected" or "directly connected" to another element, it should be understood that the other element does not exist in the middle.

The singular expression includes the plural expression unless the context clearly dictates otherwise.

In the present application, terms such as “comprises” or “have” are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, but one or more other features It is to be understood that this does not preclude the possibility of the presence or addition of numbers, steps, operations, components, parts, or combinations thereof.

1 schematically shows a method for normalizing an identifier of a biomedical entity name according to an embodiment of the present invention.

In general, a natural language refers to a language that occurs naturally, such as a language used for communication. The opposite concept is artificial language. For example, the natural language refers to Korean, English, and the like, and the artificial language may be a programming language.

Recently, in order to computerize vast amounts of data, a technology is needed to process natural language, convert it into a language that computers can understand, and vice versa. In other words, it is necessary to match the natural language with a specific symbol and digitize it. The digitized data is stored on a server, etc., and the conversion process is reversed, so that it can be expressed in a language that can be understood by humans.

Here, before describing the normalization method according to an embodiment of the present invention, some terms will be described.

Named entity recognition refers to a technique for recognizing, extracting, and classifying a word (entity name) corresponding to a predefined person, company, place, time, etc. from a document. For example, in the case of disease named entity recognition, when a sentence includes the words “cancer”, “tumor”, and “prostate cancer”, the disease is an entity type in which the above words are predefined. ) can be recognized as a corresponding word.

Named entity normalization refers to assigning an appropriate unique identifier or identification code to a recognized entity name. For example, if the word “cancer” is recognized as a disease name through the disease name recognition technique, an identifier “MESH:D009369”, which is a previously known “MESH” database identifier, can be assigned to “cancer” through normalization. Even when identically recognized entities are expressed using different types of words, the same identifier may be assigned. For example, even when a word is a full name or a word is an abbreviation of the word, the meaning is the same, and the same identifier may be assigned.

Because general NER recognizes predefined common nouns (person, place, time, etc.), there are cases in which it is not necessary to assign an individual ID. However, there is a limit in that it is difficult to apply NER without an individual identifier when it is necessary to distinguish various biomedical entity names.

To overcome this, a dictionary-based approach has been proposed. However, the dictionary-based normalization method is a method of normalizing using a dictionary related to a specific domain, and has a limitation in that it cannot be applied to words not in the dictionary.

As another overcoming method, a regularization method using machine learning has been proposed. As a regularization method using machine learning, methods such as a rule-based method, a feature-based method, and TF-IDF have been proposed. However, in the case of the regularization method using machine learning, there is a limitation in that new words such as abbreviations of a new method cannot be normalized.

Accordingly, the biomedical entity name normalization method of the present invention is largely performed through two steps as shown in FIG. 1 .

The first stage is the word representation stage. Normalize the biomedical literature and the existing DB through a deep learning-based processing method. In this case, the word representation method based on deep learning used is preferably the Word2Vec method. This is a deep learning method under the premise that words with a similar distribution have similar meanings.

The second step is the normalization step. The first step is to apply the word to be identified to the model constructed through the first step. Specifically, the normalization step is a step of assigning an identifier to a word by applying candidate words that have not yet been normalized in the biomedical literature or the like to the normalization model constructed in the first step. The identifier assigned to the word is configured in a form that is easy to convert into a computer language and can be easily used to identify a biomedical term.

2 is a block diagram illustrating an algorithm according to an embodiment of the present invention in more detail.

First, in the training stage, a regularization model is constructed using the existing database and biomedical literature.

1) Integration of information from the training data set

Before constructing the training corpus and word vectors for all words in the unclassified data, we describe how the information in the object dictionary and training corpus is integrated. Hereinafter, the name of the biomedical entity in the sentence is referred to as a mention.

In the present invention, the sentences of the training corpus and unclassified data are replaced with the concept of synonyms and training corpus in the dictionary. For example, if "cancer" is mentioned in a sentence, create a new sentence replacing "cancer" with synonyms such as "neoplasms", "tumor", "tumors", "tumour", or "tumours".

In addition, root extraction and transformation methods of disease names are added. The lexical variation is obtained through the stemming analyzer of Apache Lucene that implements the porter stemming algorithm (Prior Art Document 1). For example, if "metabolism" is mentioned in a sentence, the common variations of "metabolic", including "metabolic", "metabolite", and "metabolize", and the root ("metabole") are replaced to make a new sentence. .

When a mention includes several words, in the present invention, each word is connected with an underscore ("_") to create one word. For example, if a mention of "breast cancer" is identified from a sentence, a new sentence is created that replaces "breast cancer" with one single word, "breast_cancer". In order to increase the coverage of entities represented in the vector space, diseases or plant names in the entity dictionary that are not included in the training data are added to the training data.

2) Word representations (Word2Vec)

Mikolov developed Word2Vec (Prior Art Document 2), a neural network approach for computing vector representations of words. Vectors can be constructed using two algorithms: a continuous bag-of-word (CBOW) model and a skip-gram model. The CBOW model trains word representations by predicting words in a sentence using the surrounding words, and the skip-gram model trains word representations by predicting the surrounding words of a word in the input layer. In Word2Vec, words are expressed as vectors with hundreds of dimensions, and words with related meanings have a high probability of having similar values in the vector space. The vector w _t for the word located at the t-th position in the sentence is calculated by maximizing the average log probability as shown in Equations 1 (CBOW equation) and Equation 2 (skip-gram equation).

는 문장에서 c번째 단어의 주변에 있는 단어 벡터들이고, T는 토큰의 수이다. 본 발명에서는 모델을 트레이닝하기 위해 CBOW 및 skip-gram 알고리즘에서 단어 주변의 윈도우 사이즈 및 단어의 벡터 크기의 몇가지 옵션을 적용한다. 그리고 난 뒤, 본 발명에서는 개발 셋을 사용하여 최선의 옵션을 선택한다.From here,

are the word vectors around the c-th word in the sentence, and T is the number of tokens. In the present invention, several options of the window size around the word and the vector size of the word are applied in the CBOW and skip-gram algorithms to train the model. Then, in the present invention, the best option is selected using the development set.

Since the amount of unclassified data is huge, it is difficult to manually process the entity name. In the present invention, unclassified data is constructed using the existing NER system. Modified evidence sentences were constructed, as described above, by replacing biomedical entity mentions with pre-existing training sets and synonym concepts.

For example, "VWS" is an abbreviation for "van der woude syndrome" through an abbreviation dictionary, and it is known that the word has synonyms such as "lip pits". In the training data, the sentence appears similar to "affected males and females are equally likely to transmit VWS."

Through the rules, the word "VWS" is changed in a sentence to "van der woude syndrome", "van_der_woude_syndrom", "lip pits", or "lip_pits". Therefore, the following sentences can be additionally obtained from the basic line sentence.

(1) "affected males and females are equally likely to transmit van der woude syndrome,"

(2) "affected males and females are equally likely to transmit van_der_woude_syndrome"

(3) "affected males and females are equally likely to transmit lip pits"

(4) "affected males and females are equally likely to transmit lip_pits"

In the present invention, four semi-supervised learning models are proposed. Each model constructs a vector set (V) representing words in the vector space by applying Word2Vec to the training corpus and unlabeled data set. Each learning model is as follows.

(1) semi-supervised learning with unlabeled data of "all abstracts" (hereafter referred to as "SSL-all abstracts"),

(2) semi-supervised learning with unlabeled data of "entity-specific abstracts" ("SSL-entity abstracts");

(3) semi-supervised learning with unlabeled data of "evidence sentences" ("SSL-evidences"),

(4) semi-supervised learning with unlabeled data of "modified evidence sentences" ("SSL-modified evidences").

In addition to the above four models, the present invention further comprises two additional models.

(5) semi-supervised model that used only modified evidence sentences without the training corpus ("SSL-only modified evidences").

(6) supervised learning model with the training corpus ("SL-only training data").

Then, in the testing phase, the terms to be identified are applied to the model constructed in the training phase.

1) Prediction for normalizing biological entities

As shown in FIG. 2 , in the testing phase, as an example, abstracts of the NCBI disease corpus and plant corpus are used to test the normalization model. Biomedical mentions are extracted from the abstract. If the extracted mention exactly matches the concept name, it is assigned to the corresponding concept ID, and no further normalization steps are performed. Next, an abbreviation step is applied, in which the abbreviation is changed to the original long word using an abbreviation vowel dictionary or an abbreviation system.

For normalization, the test mention is mapped to the concept by calculating the cosine similarity between the vector of the test mention and the vector of all possible concepts in the object dictionary. Then, a word with a high cosine similarity is considered as a candidate concept. Mention m and candidate concept c are represented by vectors v _m and v _c , respectively. If mention m constitutes a single token, such as “cancer” or “tumor”, a vector for a single token in the vector set (V) is assigned to v _m . When mention m includes multiple tokens, v _m is assigned to the average of vectors for tokens in mention as shown in Equation 3 below.

Here, v _mi is the vector of the i-th token in the mention, and n is the number of tokens. If v _mi is not included in the aforementioned vector set V, v _mi is assigned to zero, and an average vector v _m is calculated using Equation 3 above. At this time, the concept of multiple tokens is converted into a single token using underscores in the training phase. After mentions of biomedical entities are expressed as vectors, the concept of high cosine similarity to the vector (v _m ) of biomedical entities with respect to the word vector (v _c ) included in the vector set V is a normalization concept. Recommended.

2) Evaluation metric of regularization tool

In order to measure the performance of the disease name normalization tool, a concept mapped manually in a test corpus is compared with a concept predicted according to an algorithm according to an embodiment of the present invention. Table 1 shows examples of fully qualified disease names in the NCBI test set.

ranks Candidate names cosine similarity *
*
*
*
*
*



* One
2
3
4
5
6
7
8
9
10 COMPLEMENT_COMPONENT_7_DEFICIENCY
complement_component_7_defici
c7_defici
complement_component_7_deficiency
C7_DEFICIENCY
c7_deficiency
antibodi_defici_syndrom
Immunologic_Deficiency_Synmdromes
immunolog_defici_syndrom
c7d 0.559244
0.554464
0.549911
0.540654
0.533657
0.525014
0.510718
0.499981
0.492753
0.491925

"C7 defect" is a synonym for "COMPLEMENT COMPOENET 7 DEFICIENCY" as a disease mention in the NCBI disease test corpus, and the corresponding concept identifier is "OMIM:610102". For a given mention, other names are ranked according to their cosine similarity to the mention in the vector representation.

Since the concept identifier contains several disease synonyms, the asterisk in the first column indicates that they are synonyms for the concept identifier, indicating that they are correctly suggested answers. Candidates ranked first, second, third, fourth, fifth, sixth, and tenth in Table 2 are correct results.

The performance of the regularization model for all mentions in the test set for each rank threshold is measured. For a given rank threshold, a predicted name (or its concept ID) that ranks higher than the threshold is considered positively predicted. True positive (TP) is a correct positive prediction, false positive (FP) is an incorrect positive prediction, and false negative (FN) is a mention that is not positively predicted. If the extracted mention exactly matches the concept name, only a single concept ID has been assigned, and it is judged to be properly qualified. Thus, when calculating performance for each rank threshold, this exact match is treated as a TP. 3 shows a candidate list and examples of TP, FP and FN. The precision (p), the recovery rate (r) and the F-score (f) are expressed as in Equation (4).

Figure 3 shows a comparison of the performance of the model proposed in the present invention and the existing regularization model DNorm (Prior Art Document 3) in the presence and absence of an abbreviation resolution step.

In Figure 3, "SSL-modified evidence" shows superior performance in both cases, "SL-only training data" shows a better effect than "SSL-only modified evidence".

3 shows that normalization accuracy is improved when unclassified data is combined with training data. "SSL-modified evidences" gives the best performance.

4 shows a comparison of performance between models for plant name normalization proposed in the present invention.

As shown in Figure 4, "SSL-modified plant evidence" shows the best performance. Unlike disease name or disease name normalization results, "SSL-only modified evidence" shows better performance than "SL-only training data". Since an abbreviation dictionary cannot be used, plant names are expressed in multiple types depending on their context, region or language. And plant name or plant name normalization shows lower accuracy than disease name normalization.

5 is a flowchart illustrating a biomedical term identification method according to an embodiment of the present invention.

The algorithm described in FIG. 5 may be programmable, and the corresponding program may be implemented in an entity name normalization system.

The entity name normalization system constructs training data (S101).

In this case, the entity name normalization system may construct training data by processing the existing database and unclassified data. The entity name identification system may include constructing additional data by substituting unclassified terms using synonyms in the dictionary. In addition, the entity name normalization system may include a method of constructing additional data by extracting and transforming roots from unclassified terms. Also, when multiple words are included in the mention, a method of generating a single word by connecting each word with an underscore may be included.

The entity name normalization system configures a normalization model by applying a deep learning-based word expression to the constructed training data (S103). In this case, it is preferable to use the Word2Vec algorithm as the deep learning-based algorithm used.

In this case, the configured regularization model may be plural.

The entity name normalization system identifies biomedical terms by applying test data to the configured normalization model (S105).

Here, the test data refers to biomedical data to which an identifier is to be assigned. The entity name normalization system assigns an identifier to data to which an identifier is not assigned according to a normalization model, so that digital databaseization of a corresponding term can be easily performed.

The present invention described above can be implemented as computer-readable code on a medium in which a program is recorded. The computer-readable medium includes all kinds of recording devices in which data readable by a computer system is stored. Examples of computer-readable media include Hard Disk Drive (HDD), Solid State Disk (SSD), Silicon Disk Drive (SDD), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc. There is this. Accordingly, the above detailed description should not be construed as restrictive in all respects but as exemplary. The scope of the present invention should be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the present invention are included in the scope of the present invention.

Claims (5)

In the object name normalization method in which each step is performed by a computing device,
processing known databases and unclassified data on biomedical terms to construct training data;
constructing a regularization model by applying a word representation based on deep learning to the constructed training data; and
applying the test data to which the identifier is to be assigned to the configured normalization model to give an identifier, and identifying the biomedical term with the normalized model to which the identifier is assigned;
The step of constructing the training data is,
A method of constructing additional data using synonyms in a known dictionary, a method of constructing additional data using root extraction and transformation, Concatenate with underscores to form a single word.
How to normalize object names. delete delete delete According to claim 1,
The deep learning-based word expression is
Word2Vec algorithm
How to normalize object names.

Images