Uma proposta de avaliação de embeddings de palavras por similaridade

Abstract

The representation of human text in numerical information is a difficult task that has been investigated and used on the field of Artificial Intelligence. In this context, the use of vector representations through Word Embeddings and Sentence Embeddings for Natural Language Processing tasks has become very common in the state of art, it is also a well-studied knowledge in the area of data sciences. Using them as hidden layers in Deep Learning models greatly reduces costs and improves performance for most tasks. However, as there are many such algorithms and pre-trained embeddings, it is not an obvious choice for a data scientist which model best fits their context. Therefore, many strategies seek to evaluate the performance of such models, intrinsically and extrinsically, so that some intuition or conclusion about them can be used to assist in the process of creating a statistical or Deep Learning model in Natural Language Processing. Thus, seeking to develop a way to intrinsically evaluate some of the classic Word Embeddings, this work proposes an evaluation of pre-trained Word Embeddings through word similarity, investigating the semantic similarity and association between learned words by embeddings and basing them on pre-annotated datasets of synonym words.