BackgroundOne of the most important processes in a machine learning-based natural language processing is to represent words. The one-hot representation that has been commonly used has a large size of vector and assumes that the features that make up the vector are independent of each other. On the other hand, it is known that word embedding has a great effect in estimating the similarity between words because it expresses the meaning of the word well. In this study, we try to clarify the correlation between various terms in the biomedical texts based on the excellent ability of estimating similarity between words shown by word embedding. Therefore, we used word embedding to find new biomarkers and microorganisms related to a specific diseases.MethodsIn this study, we try to analyze the correlation between diseases-markers and diseases-microorganisms. First, we need to construct a corpus that seems to be related to them. To do this, we extract the titles and abstracts from the biomedical texts on the PubMed site. Second, we express diseases, markers, and microorganisms’ terms in word embedding using Canonical Correlation Analysis (CCA). CCA is a statistical based methodology that has a very good performance on vector dimension reduction. Finally, we tried to estimate the relationship between diseases-markers pairs and diseases-microorganisms pairs by measuring their similarity.ResultsIn the experiment, we tried to confirm the correlation derived through word embedding using Google Scholar search results. Of the top 20 highly correlated disease-marker pairs, about 85% of the pairs have actually undergone a lot of research as a result of Google Scholars search. Conversely, for 85% of the 20 pairs with the lowest correlation, we could not actually find any other study to determine the relationship between the disease and the marker. This trend was similar for disease-microbe pairs.ConclusionsThe correlation between diseases and markers and diseases and microorganisms calculated through word embedding reflects actual research trends. If the word-embedding correlation is high, but there are not many published actual studies, additional research can be proposed for the pair.
Semantic Role Labeling (SRL) is to determine the relationship between predicates and their arguments in a sentence. Nowadays the most widely used semantic corpus for the SRL is Proposition Bank (PropBank) which is semantically annotated over the predicate and argument structure. But the Korean version of the PropBank could not be widely used because the corpus has a size limit. To solve the problem, we use another semantic tagged corpus, built by Sejong Plan, which is nationwide Korean corpus construction project. In this paper, we propose a method for automatically converting the roles between PropBank and Sejong Corpus, which has different semantic roles, to expand the size of annotated corpus. In the experiment, we convert 491 arguments automatically and about 78% of them show the agreement with manually annotated arguments.
One of the most important processes in a machine learning-based natural language processing module is to represent words by inputting the module. This can be accomplished by representing words in one-hot form with a large vector size without applying the concept of semantic similarity between words, or by word representation (embedding) with vectors to represent lexical similarity. In this study, classification performance of three word representation models (Word2Vec, canonical correlation analysis, and GloVe) is tested on a corpus that established using the abstracts of 204,674 biomedical articles published in PubMed. Categories include disease name, disease symptom, and ovarian cancer marker. The classification performance of each word representation model for each category is visualized by mapping the results in two-dimensional word representations using t-SNE.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.