Automatic sentiment polarity analysis of financial news across languages and markets.

The increasing amount of sentiments disseminated by traditional and social media and their impact on human societies including financial markets made the automatic detection and analysis of sentiments an important research area in academia and industry. In this thesis, techniques rooted in corpus linguistics and language for special purposes (LSP) have been used to develop a method for the automatic sentiment polarity analysis of financial news. Contrary to the existing practice in sentiment analysis where manual analysis is usually required to construct sentiment lexicons, the novelty introduced here is that the method can automatically extract a set of domain specific keywords and a set of polarity bearing words from training corpora comprising financial news across languages. The method requires only general news corpora for extracting keywords and historical market data time-series for extracting polar words. The method also automatically labels the extracted polar words as positive or negative and extracts the patterns in which these words are used. The language dependence of the use of polarity laden words was examined by looking at such patterns of usage in two typologically distinct languages: English and Arabic. Despite the cultural and linguistic differences, the patterns of words used in expressing sentiments have remarkable similarities. The method has been evaluated using two evaluation corpora comprising financial news in English and Arabic using a tri-partite strategy. First, human volunteers were asked to label the polarity of all the sentences in the evaluation corpora. Second, a widely used polarity lexicon, the General Inquirer, was used to automatically label the polarity of the same sentences. Third, LoLo, a system implementing the method, was used to label the polarity of the same sentences as well using automatically extracted polar words and patterns from training corpora. LoLo' s results were closer to that of humans when analysing the full content of financial news items and when analysing only lead sentences, in both English and Arabic, than that of the General Inquirer based analysis