On Monday, November 27, Professor Diego Burgos, Associate Professor of Spanish, Department of Spanish and Italian, Wake Forest University, will present:
From Words to Feelings: A Journey to Opinion Mining
Have you ever ended up feeling supportive or opposing of public figures or policies after reading the news? Such scenario of persuasiveness is not uncommon in natural languages due to the rhetoric effect of some texts. However, persuasive is not how one would normally tag news articles that are claimed to be objective such as those in the economy or politics sections of dailies. This talk presents the methodology and experimentation for opinion mining on a news article database. The study aims at automatically determining the stance of the newswriter with regard to the topic he or she writes about.
I will first describe how we populated our database with 27 linguistic features, and will report on the results of a three-stage series of classification experiments with five machine learning algorithms. In a first phase, we tackle the problem as a classification task into three classes (i.e., positive articles, negative articles, neutral articles) with an efficiency of 48% for the best model. The second phase excludes the neutral class with the best model reaching 68%. In the third phase, we classify into three classes again, but using feature selection and dimension reduction techniques, and reached 54%. By excluding the neutral class again in this phase, we obtained 72% of efficiency using principal component analysis, which also showed a better balance of sensibility and specificity.
I will conclude with some remarks on what makes this classification task particularly challenging. Then I will close with an outline of ongoing and future work.