This chapter presents a software tool that helps the Competitive Intelligence process by collecting and analyzing texts published on the Internet. The goal is to automatically analyze indicators of the sentiment present in texts about a certain theme, whether positive or negative. The sentiment analysis is made through a probabilistic process over keywords present in the texts, using as reference a task ontology with positive and negative words defined with a degree of confidence.
TopAutomatic Sentiment Analysis is the analysis of affective states made by machines over documents written in natural language, based on emotion categories like positive and negative (Gregory, et al., 2006). Other authors utilize synonyms as “opinion mining” (Pang & Lee, 2008). This kind of work is applied for analysis of customers´ satisfaction about products and services, for identification of attitudes on reports of employees and so on.
Ma et al. (2005) presents experiments with sentiment analysis over chat messages. Emotions are estimated based on keywords present in the text. The work of Genereux and Evans (2006) classifies weblogs according to their affective content, using a two-dimensional space with axes “positive or negative” and “active or passive.”
Godbole et al. (2007) assigns scores indicating positive or negative opinion to each distinct entity in texts from newspapers and blogs. They expand a small candidate list of positive and negative words into a sentiment dictionary using path-based analysis of synonyms and antonyms in WordNet.
Pang and Lee (2008) presents a survey about the main techniques used for opinion mining and sentiment analysis in texts, besides concerning other issues in this theme. They expose the difficulty of analyzing sentiments based only on words present in the text. They report an accuracy of 60% for the use of lists of keywords created by humans within a straightforward classification policy. In contrast, “word lists of the same size but chosen based on examination of the corpus’ statistics achieves almost 70% accuracy” and “applying machine learning techniques based on unigram models can achieve over 80% in accuracy.”
Ortony, Clore, and Colins (Ortony, et al., 1988) presents a psychological and cognitive model of emotions, containing emotions resulting from the description of cognitive processes and real world interpretations by a human subject. The model contains 22 kinds of emotions: “happy for,” “resentment,” “gloating,” “pity,” “joy,” “distress,” “pride,” “shame,” “admiration,” “reproach,” “love,” “hate,” “hope,” “fear,” “satisfaction,” “fears-confirmed,” “relief,” “disappointment,” “gratification,” “remorse,” “gratitude,” and “anger.” Table 1 shows the emotions and their corresponding adjectives.