Improving Web 2.0 Opinion Mining Systems Using Text Normalisation Techniques

A basic task in opinion mining deals with determining the overall polarity orientation of a document about some topic. This has several applications such as detecting consumer opinions in on-line product reviews or increasing the effectiveness of social media marketing campaigns. However, the informal features of Web 2.0 texts can affect the performance of automated opinion mining tools. These are usually short and noisy texts with presence of slang, emoticons and lexical variants which make more difficult to extract contextual and semantic information. In this paper we demonstrate that the use of lexical normalisation techniques can be used to enhance polarity detection results by replacing informal lexical variants with their canonical version. We have carried out several polarity classification experiments using English texts from different Web 2.0 genres and we have obtained the best result with microblogs where normalisation contribution to the classification model can be up to 6.4%.