Geração de vetores de sentido para o português

Abstract

Numerical vector representations are able to represent from words to meanings, in a low-dimensional continuous space. These representations are based on distributional modeling, where the context in which the word occurs is taken into account for vector generation. The word representations, known as word embeddings or word vectors (Word2vec, FastText, Wang2vec and Glove), which have been widely used until now, have an important limitation: they produce a single vector representation for each word, ignoring the fact that ambiguous words can represent different meanings (different contexts). This mixture of meanings can be a problem for many applications. For example, in a language comprehension task, using the vector of an ambiguous word as "bank", all possible meanings --such as financial institution, blood bank, or furniture item --will be mixed into a single numerical vector, causing an erroneous semantic interpretation of the sentence in which it occurs. Over the last few years, representations of meanings, known as sense embeddings or sense vectors, have proven to be able to model syntactic and semantic knowledge and have been used in NLP applications. By being able to transform the various meanings of an ambiguous word into numerical vectors, sense vectors can be applied to Word Sense Disambiguation (WSD). Thus, this work generated and evaluated sense vectors for Portuguese (PT-BR and PT-EU), and showed that they overcome traditional vectors in intrinsic and extrinsic NLP tasks, since they are capable of dealing with lexical ambiguity. To the best of our knowledge, this is the first work to address the geneation and evaluation of sense vectors for Portuguese.