Meta-level sentiment models for big social data analysis

People react to events, topics and entities by expressing their personal opinions and emotions. These reactions can correspond to a wide range of intensities, from very mild to strong. An adequate processing and understanding of these expressions has been the subject of research in several fields, such as business and politics. In this context, Twitter sentiment analysis, which is the task of automatically identifying and extracting subjective information from tweets, has received increasing attention from the Web mining community. Twitter provides an extremely valuable insight into human opinions, as well as new challenging Big Data problems. These problems include the processing of massive volumes of streaming data, as well as the automatic identification of human expressiveness within short text messages. In that area, several methods and lexical resources have been proposed in order to extract sentiment indicators from natural language texts at both syntactic and semantic levels. These approaches address different dimensions of opinions, such as subjectivity, polarity, intensity and emotion. This article is the first study of how these resources, which are focused on different sentiment scopes, complement each other. With this purpose we identify scenarios in which some of these resources are more useful than others. Furthermore, we propose a novel approach for sentiment classification based on meta-level features. This supervised approach boosts existing sentiment classification of subjectivity and polarity detection on Twitter. Our results show that the combination of meta-level features provides significant improvements in performance. However, we observe that there are important differences that rely on the type of lexical resource, the dataset used to build the model, and the learning strategy. Experimental results indicate that manually generated lexicons are focused on emotional words, being very useful for polarity prediction. On the other hand, lexicons generated with automatic methods include neutral words, introducing noise in the detection of subjectivity. Our findings indicate that polarity and subjectivity prediction are different dimensions of the same problem, but they need to be addressed using different subspace features. Lexicon-based approaches are recommendable for polarity, and stylistic part-of-speech based approaches are meaningful for subjectivity. With this research we offer a more global insight of the resource components for the complex task of classifying human emotion and opinion.