Sentiment Analysis of News Articles in Spanish using Predicate Features

The automatic prediction of the course of action of agents involved in social or economic trends is an imperative challenge nowadays. However, it is a difficult task because stance or opinion is often spread throughout long, complex texts, such as news articles. The current study tests sentence predicates as features to automatically determine the writer’s stance in news articles. We capture the semantics and stance of the text by encoding features such as the attribute of copulative sentences, the predicate of transitive sentences, adjectival phrases, and the section of the article. Under the assumption that these features are informative enough to model the semantics of the text, each word sequence is disambiguated and assigned a sentiment value using weighting rules. Different experiments were run using either SentiWordNet and ML-Senticon to determine words’ sentiment. Feature vectors are automatically built to populate a database that is tested using two machine learning algorithms. An efficiency of 69% was achieved using a SVM with Gaussian kernel along with a feature selection strategy. This score outperformed the bag-of-words baseline in 12%. These results are promising considering that the sentiment analysis is performed on very complex texts written in Spanish.