Learning temporal sentiment from business news : a computational approach

Abstract

This thesis concerns is about the analysis of textual sentiment using computational approaches. Textual sentiment reflects the opinions and attitudes expressed through textual communications. With the advent of the Internet and World Wide Web, individuals are being exposed to an ever-growing amount of sentiment through various channels; such information can have powerful effect on how people make decisions. Because of this, computational approaches to sentiment analysis has attracted increasing interest from researchers in the past two decades. The goal of sentiment analysis is to determine whether a piece of text conveys a positive or negative evaluation on a certain topic. Being able to identify and comprehend sentiment expressed by the public at a large scale can be very useful to many applications. One such application is to use sentiment conveyed in business news to explain the behaviours of financial markets. In this thesis, I have proposed an extended model for describing the interactions between news sentiment and financial markets. This extended model recognises two different types of news sentiment: (i) the retrospective news sentiment, which refers to the sentiment that is associated with news stories that recount past events in the market; and (ii) the prospective news sentiment, which refers to the textual sentiment in business and financial news that is associated with speculations and projections about the future developments of the market. A method that can automatically extract sentiment-laden language patterns for the temporal sentiment classes was developed. The method models the texts in news articles as a mixture of lexical items distributions, where each distribution characterises one of the temporal sentiment classes. A supervised maximisation expectation algorithm was derived to infer these distributions given a corpus and a market performance measure as the target. The supervision was enforced by implementing different conditional probability distributions for returns given the underlying temporal sentiment class of the news article. The method was implemented by a system comprising two components, CiCui and TSMiner, which was used to conduct a case study on firm-level data. The results of the study suggested it is mainly changes in the market that leads to changes in the news sentiment, and retrospective news sentiment is in general more prevalent in business news. Also noted was that word dependencies seem to be able to capture prospective sentiment than unigrams. It has been shown that the method proposed in this thesis slightly outperforms content analysis methods used with either the GI or Loughran and McDonald's sentiment dictionary.