Modelos composicionais: análise e aplicação em previsões no mercado de ações

Abstract

Among several textual representation techniques in the literature, the distributed representation of words is standing out recently in many tasks of Natural Language Processing through its representations based on dense vectors of 𝑑 dimensions that can capture syntactic and semantic information of the words. Therefore, it’s expected that similar words regarding to syntactic and sematic are closer of each other in the vector space. However, while this representation is becoming effective to isolated words, there isn’t a consensus in the literature regarding to the best way to represent more complex structures, such as phrases and sentences. The trend of recent years is the use of compositional models that represents these complex structures through the composition of the representations of its constituent structures using some combination function. However, it’s known that the obtained results by this technique depends directly of the domain in which they are applied. In this work, we analyzed several compositional models applied to the domain of stock price prediction in order to identify which of these models better represent the financial news title for various machine learning methods to predict the index polarity of the S & P 500 stock exchange.