Análise de sentimento usando a representação distribuída de parágrafos para o português

Abstract

Distributed representation of words has been very discussed in the Natural Language Processing area (NLP). With the continuous growth of information on the Internet in recent decades, there is a need to in analysis tasks of this large volume of data to computers, tasks that were usually performed manually in order to make them more viable and efficient. Distributed representation of words consists of obtaining a richer modeling framework that considers relevant aspects like ordergin, semantics and compositionality of the words in a sentence. The difficulty increases when these sentences tend to grow in size, which is the case of texts with a large number paragraphs. Once you have all the sentences of a given text structured in vectors it is possible, for example, to sumarize an entire document, extract sentiment, recognize expressions, translate it into another language, among other various types of tasks. Recent studies, such as (LE; MIKOLOV, 2014) have presented techniques such as Word Vector and Paragraph Vector, which are able to take words, sentences and even paragraphs and distribute them into vectors. These techniques have shown significant gains in tasks such as Automatic Sentiment Analysis (SA) and Information Retrieval over traditional language models such as Bag-of-Words, N-gram and Skip-gram. This papaer aims to replicate the experiments made in SA task using a Brazilian Portuguese corpus.The experiments performed with the ReLi corpus in Brazilian Portuguese using the 10-fold Cross-validation method achieved a average combined accuracy of 82.99%. This higher than expected result was a consequence of an unequal number of sentences in this corpus. More experiments were performed with modified versions of the ReLi in attempt to make the number of sentences equal in the training and testing stages, resulting in a average combined accuracy 60.59% when the number of sentences with positive and negative polarity are equal.