GU-MLT-LT: Sentiment Analysis of Short Messages using Linguistic Features and Stochastic Gradient Descent

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 328–332, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics GU-MLT-LT: Sentiment Analysis of Short Messages using Linguistic Features and Stochastic Gradient Descent Tobias G¨unther University of Gothenburg Olof Wijksgatan 6 41255 G¨oteborg, Sweden email@tobias.io Lenz Furrer University of Zurich Binzm¨uhlestrasse 14 8050 Z¨urich, Switzerland lenz.furrer@gmail.com Abstract This paper describes the details of our system submitted to the SemEval-2013 shared task on sentiment analysis in Twitter. Our approach to predicting the sentiment of Tweets and SMS is based on supervised machine learning tech- niques and task-specific feature engineering. We used a linear classifier trained by stochas- tic gradient descent with hinge loss and elas- tic net regularization to make our predictions, which were ranked first or second in three of the four experimental conditions of the shared task. Furthermore, our system makes use of social media specific text preprocessing and linguistically motivated features, such as word stems, word clusters and negation handling. 1 Introduction Sentiment analysis, also known as opinion min- ing, is a research field in the area of text min- ing and natural language processing, which inves- tigates the automated detection of opinions in lan- guage. In written text, an opinion is a person’s atti- tude towards some topic, pronounced by verbal (e.g. choice of words, rhetorical figures) or non-verbal means (e.g. emoticons, emphatic spelling). More formally, Liu (2012) defines an opinion as the quin- tuple (ei, aij, sijkl, hk, tl) where “ei is the name of an entity, aij is an aspect of ei, sijkl is the sentiment on aspect aij of entity ei, hk is the opinion holder, and tl is the time when the opinion is expressed by hk. The sentiment sijkl is positive, negative, or neu- tral, or expressed with different strength/intensity levels [...]. When an opinion is on the entity itself as a whole, the special aspect GENERAL is used to denote it. [...] ei and aij together represent the opinion target” (Liu, 2012). With the massively growing importance of social media in everyday life, being able to automatically find and classify attitudes in written text allows for estimating the mood of a large group of people, e.g. towards a certain event, service, product, matter of fact or the like. As working with the very short and informal texts typical for social networks poses chal- lenges not encountered in more traditional text gen- res, the International Workshop on Semantic Evalu- ation (SemEval) 2013 has a shared task on sentiment analysis in microblogging texts, which is detailed in Wilson et al. (2013). The task requires sentiment analysis of Twitter1 and SMS messages and com- prises two subtasks, one of which deals with deter- mining the sentiment of a given message fragment depending on its context (Task A) and one on over- all message polarity classification (Task B). We treat both tasks as document-level senti- ment classification tasks, which we define ac- cording to …

Content truncated to 500 words