senti.ue-en: an approach for informally written short texts in SemEval-2013 Sentiment Analysis task

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 508–512, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics senti.ue-en: an approach for informally written short texts in SemEval-2013 Sentiment Analysis task Jos´e Saias DI - ECT - Universidade de ´Evora Rua Rom˜ao Ramalho, 59 7000-671 ´Evora, Portugal jsaias@uevora.pt Hil´ario Fernandes Cortex Intelligence Rua Sebasti˜ao Mendes Bolas, 2 K 7005-872 ´Evora, Portugal hilario.fernandes@cortex-intelligence.com Abstract This article describes a Sentiment Analysis (SA) system named senti.ue-en, built for participation in SemEval-2013 Task 2, a Twitter SA challenge. In both challenge sub- tasks we used the same supervised machine learning approach, including two classifiers in pipeline, with 22 semantic oriented features, such as polarized term presence and index, and negation presence. Our system achieved a better score on Task A (0.7413) than in the Task B (0.4785). In the first subtask, there is a better result for SMS than the obtained for the more trained type of data, the tweets. 1 Introduction This paper describes the participation of a group led by Universidade de ´Evora’s Computer Science Department in SemEval-2013 Task 2 (Wilson et al., 2013), using senti.ue-en system. Having previous experience in NLP tasks, such as ques- tion answering (Saias, 2010; Saias and Quaresma, 2012), this was the authors first attempt to imple- ment a system for Sentiment Analysis (SA) in En- glish language. We have a recent work (Fernandes, 2013) involving SA but it is geared towards Por- tuguese language, and thought for regular text. It was based on rules on the outcome of linguistic anal- ysis, which did not work well for tweets, because the morphosyntactic analyzer misses much, due to the abundance of writing errors, symbols and abbrevia- tions. Moreover, in that work we began by detecting named entities and afterwards classify the sentiment being expressed about them. For SemEval the goal is different, being target-independent. In both A and B subtasks, systems must work on sentiment polar- ity, in a certain context or full message, but the target entity (or the opinion topic) will not appear in the output. Thus, we have decided that senti.ue-en system would be implemented from scratch, for En- glish language and according to the objectives of this challenge, in particular the Task B. 2 Related Work Microblogging and social networks are platforms where people express opinions. In recent years many papers have been published on social me- dia content SA. Pang et al. (2002) applied machine learning based classifiers for sentiment classification on movie reviews. Their experimental results using Naive Bayes, Maximum Entropy, and Support Vec- tor Machines (SVM) algorithms achieved best re- sults with SVM and unigram presence as features. Some target-dependent approaches are sensitive to the entity that is receiving each sentiment. A sen- tence can have a positive sentiment about an entity and a negative for another. Such classification can be performed with rules on the occurrence of nouns, verbs and adjectives, as done in (Nasukawa and Yi, …

Content truncated to 500 words