Graph Embeddings para Node Classification em representação baseada em grafos de frases em linguagem natural

Abstract

Due to the large amount of works developed in the biomedical field and the availability ofhuge databases on biomedical entities, including proteins, genes and viruses, it comesthe need to be able to automatically index such human knowledge bases.Such need has led to the development and computational tools to assist the researcherin the recovery of specific information involving certain proteins and their relations. Inthis context, two of the main problems in the biomedical area involving techniques of Text Mining most investigated are the Named Entity Recognition (NER) and RelationExtraction.This work focuses on the first problem that serves as a basis for the second, i.e., first wehave to identify and classify the entities and then, with the identified/classified entities,identify the existing relations between them, if any. The approach adopted in this paperis based on the recent techniques of supervised/non-supervised learning of deep neural networks, or Deep Learning (DL). In particular, the problem of NER is investigated usingrecent techniques of dense feature representation using DL.At first, the sentences from a biomedical corpus are represented as graphs thanks tothe generation of annotations (metadata) generated automatically by natural language processing tools, such as tokenization, syntactic parsing, etc. These graphs are thenimported into a graph-based database so that various queries submitted to this data base can be optimized in order to extract both lexical and syntactic attributes (or features) ofthe entities (or nodes) present in the graphs. The information generated in the previousstep is used as input Deep Learning-based algorithms called Graph Embedding (GE)that map the representation of graph nodes (entity) in a dense vector representation(vector of real numbers) that has several properties of interest for this search. Finally,such dense representation of features) are employed as input for supervised machine learning algorithms.This work presents an experimental study where some of the existent algorithms of GEare compared, along with several types of sentence representation based on graphs,and their impacts on the task of entity classification (NER), or node classification. Theexperimental results are promising, reaching more than 90% accuracy in the best cases