PUT at SemEval-2016 Task 4: The ABC of Twitter Sentiment Analysis

Proceedings of SemEval-2016, pages 126–132, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics PUT at SemEval-2016 Task 4: The ABC of Twitter Sentiment Analysis Mateusz Lango and Dariusz Brzezinski and Jerzy Stefanowski Institute of Computing Science, Poznan University of Technology 60-965 Pozna´n, Poland {mlango,dbrzezinski,jstefanowski}@cs.put.edu.pl Abstract This paper describes a classification system that participated in SemEval-2016 Task 4: Sentiment Analysis in Twitter. The proposed approach competed in subtasks A, B, and C, which involved tweet polarity classification, tweet classification according to a two-point scale, and tweet classification according to a five-point scale. Our system is based on an en- semble consisting of Random Forests, SVMs, and Gradient Boosting Trees, and involves the use of a wide range of features including: n- grams, Brown clustering, sentiment lexicons, Wordnet, and part-of-speech tagging. The proposed system achieved 14th, 6th, and 3rd place in subtasks A, B, and C, respectively. 1 Introduction In recent years, sentiment analysis (Liu, 2012) has become a common yardstick for many new text min- ing algorithms. This trend is a direct result of the rapid growth of social media, where users express their views and opinions regarding a wide range of topics. As a result, social networks like Twitter have become a crucial resource in product design, assessing marketing campaigns, and detecting news bursts (Liu, 2012; Mathioudakis and Koudas, 2010). However, while the merits of resources such as Twitter are evident, there are several difficulties with the use of social media data. In contrast to classi- cal sentiment analysis methods, which were origi- nally designed for dealing with well-written prod- uct reviews, texts from social media often contain misspellings, letter substitutions, ambiguities, non- standard abbreviations, and improper use of gram- mar (Sarker et al., 2015). Furthermore, resources such as Twitter generate thousands of new texts per second and introduce challenges characteristic for stream processing (Krempl et al., 2014). Moreover, the limited length of these texts makes classical n- gram feature vectors extremely sparse, which in turn hinders generalization abilities of classification al- gorithms. Finally, sentiments are usually unevenly distributed (Kiritchenko et al., 2014), resulting in class imbalance and, therefore, additional difficul- ties for classifiers (He and Garcia, 2009). To promote research in this area, Task 4 of SemEval-2016 was devoted to sentiment analysis in Twitter. The task consisted of five subtasks in- volving standard classification, ordinal classifica- tion, and distribution estimation; for a more detailed description see (Nakov et al., 2016). In this paper, we present our approach to learn a classification system which participated in subtasks A, B, and C of SemEval-2016 Sentiment Analysis in Twitter. The proposed approach combines Ran- dom Forests, Support Vector Machines, and Gradi- ent Boosting Trees, trained on a wide range of lex- ical and semantic features including: n-grams, k- grams, Brown clustering, sentiment lexicons, Senti- WordNet, and part of speech tagged 1-grams. These components were carefully combined and optimized to create a separate version of the system for each of the tackled subtasks. In the following sections, we describe each …

Content truncated to 500 words