FBM: Combining lexicon-based ML and heuristics for Social Media Polarities

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 483–489, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics FBM: Combining lexicon-based ML and heuristics for Social Media Polarities Carlos Rodr´ıguez-Penagos, Jordi Atserias, Joan Codina-Filb`a, David Garc´ıa-Narbona, Jens Grivolla, Patrik Lambert, Roser Saur´ı Barcelona Media Av. Diagonal 177, Barcelona 08018 Corresponding author: carlos.rodriguez@barcelonamedia.org Abstract This paper describes the system implemented by Fundaci´o Barcelona Media (FBM) for clas- sifying the polarity of opinion expressions in tweets and SMSs, and which is supported by a UIMA pipeline for rich linguistic and sen- timent annotations. FBM participated in the SEMEVAL 2013 Task 2 on polarity classifi- cation. It ranked 5th in Task A (constrained track) using an ensemble system combining ML algorithms with dictionary-based heuris- tics, and 7th (Task B, constrained) using an SVM classifier with features derived from the linguistic annotations and some heuristics. 1 Introduction We introduce the FBM system for classifying the polarity of short user-generated text (tweets and SMSs), which participated in the two subtasks of SEMEVAL 2013 Task 2 on Sentiment Analysis in Twitter. These are: Task A. Contextual Polarity Dis- ambiguation, and Task B. Message Polarity Classifi- cation. The former aimed at classifying the polarity of already identified opinion expressions (or cues), whereas the latter consisted in classifying the polar- ity of the whole text (Wilson et al., 2013). The literature agrees on two main approaches for classifying opinion expressions: using supervised learning methods and applying dictionary/rule- based knowledge (see (Liu, 2012) for an overview). Each of them on its own has been used in work- able systems, and a principled combination of both of them can yield good results on noisy data, since generally one (dictionaries/rules) offers good preci- sion while the other (ML) is able to discover unseen examples and thus enhances recall. FBM combined both approaches in order to bene- fit from their respective strengths and compensating as much as possible their weaknesses. For Task A we used linguistic (lexical and syntactic) annotations to implement both types of approaches. On the one hand, we built machine learning classifiers based on Support Vector Machines (SVMs) and Conditional Random Fields (CRFs). On the other, we imple- mented a basic classification system mainly based on polarity dictionaries and negation information, as well as simple decision tree-like heuristics extracted from the training data. For task B we trained an SVM classifier using some of the annotations from Task A. The paper first presents the process of data com- pilation and preprocessing (section 2), and then de- scribes the systems for Tasks A (section 3) and B (section 4). Results and conclusions are discussed in the last section. 2 Data Compilation and Processing 2.1 Making data available The corpus of SMSs was provided to the partici- pants by the organizers of the task. As for the corpus of tweets, legal restrictions on twitter data distribu- tion required the participants to download the tex- tual contents of the …

Content truncated to 500 words