NTNUSentEval at SemEval-2016 Task 4: Combining General Classifiers for Fast Twitter Sentiment Analysis

Proceedings of SemEval-2016, pages 103–108, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics NTNUSentEval at SemEval-2016 Task 4:∗ Combining General Classifiers for Fast Twitter Sentiment Analysis Brage Ekroll Jahren Valerij Fredriksen Bj¨orn Gamb¨ack Lars Bungum Department of Computer and Information Science Norwegian University of Science and Technology (NTNU) Sem Sælands vei 7–9, NO–7491 Trondheim, Norway {brageej,valerijf}@stud.ntnu.no {gamback,larsbun}@idi.ntnu.no Abstract The paper describes experiments on sentiment classification of microblog messages using an architecture allowing general machine learn- ing classifiers to be combined either sequen- tially to form a multi-step classifier, or in par- allel, creating an ensemble classifier. The sys- tem achieved very competitive results in the shared task on sentiment analysis in Twitter, in particular on non-Twitter social media data, that is, input it was not specifically tailored to. 1 Introduction As a growing platform for people to express them- selves on a global scale, Twitter has become exceed- ingly attractive as an information source. In addi- tion to text, a tweet comes with metadata such as the sender’s location and language, and hashtags, mak- ing it possible to quickly gather vast amounts of data regarding a specific product, person or event. With a working Twitter Sentiment Analysis system, compa- nies could get a feel of what consumers think of their products, or politicians could estimate their popular- ity amongst Twitter users in specific regions. However, tweets and other informal texts on so- cial media are quite different from texts elsewhere. They are short in length and contain a lot of abbre- viations, misspellings, Internet slang, and creative syntax. Although the relative occurrence of non- standard English syntax is fairly constant among many types of social media (Baldwin et al., 2013), ∗Thanks to Mikael Brevik, Jørgen Faret, Johan Reitan and Øyvind Selmer for their work on two previous NTNU systems. analysing such texts using traditional language pro- cessing systems can be problematic, primarily since the main common denominator of social media text is not that it is informal, but that it describes language in rapid change (Androutsopoulos, 2011; Eisenstein, 2013), so that resources targeted directly at social media language quickly become outdated. Twitter Sentiment Analysis (TSA) has been a rapibly growing research area in recent years, and a typical approach to TSA has been identified, using a supervised machine learning strategy, consisting of three main steps: preprocessing, feature extraction and classifier training. Preprocessing is used in order to remove noise and standardize the tweet format, for example, by replacing or removing URLs. De- sired features of the tweets are then extracted, such as sentiment scores using specific sentiment lexica or the occurrence of different emoticons. Finally, a classifier is trained on the extracted features. Since the machine learning algorithms used com- monly are supervised, sentiment-annotated data is a prerequisite for training — and the growth of the TSA research field can largely be attributed to the In- ternational Workshop on Semantic Evaluation (Sem- Eval) having run shared tasks on this theme since 2013 (Wilson et al., 2013), annually …

Content truncated to 500 words