uOttawa: System description for SemEval 2013 Task 2 Sentiment Analysis in Twitter

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 380–383, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics uOttawa: System description for SemEval 2013 Task 2 Sentiment Analysis in Twitter Hamid Poursepanj, Josh Weissbock, and Diana Inkpen School of Electrical Engineering and Computer Science University of Ottawa Ottawa, K1N6N5, Canada {hpour099, jweis035, Diana.Inkpen}@uottawa.ca Abstract We present two systems developed at the Uni- versity of Ottawa for the SemEval 2013 Task 2. The first system (for Task A) classifies the po- larity / sentiment orientation of one target word in a Twitter message. The second system (for Task B) classifies the polarity of whole Twitter messages. Our two systems are very simple, based on supervised classifiers with bag-of- words feature representation, enriched with in- formation from several sources. We present a few additional results, besides results of the submitted runs. 1 Introduction The Semeval 2013 Task 2 focused on classifying Twitter messages (“tweets”) as expressing a posi- tive opinion, a negative opinion, a neutral opinion, or no opinion (objective). In fact, the neutral and objective were joined in one class for the require- ments of the shared task. Task A contained target words whose sense had to be classified in the con- text, while Task B was to classify each text into one of the three classes: positive, negative, and neutral/objective. The training data that was made available for each task consisted in annotated Twitter message. There were two test sets for each task, one composed of Twitter messages and one of SMS message (even if there was no specific training data for SMS messages). See more details about the datasets in (Wilson et al., 2013). 2 System Description We used supervised learning classifiers from We- ka (Witten and Frank, 2005). Initially we extracted simple bag-of-word features (BOW). For the sub- mitted systems, we also used features calculated based on SentiWordNet information (Baccianella et al., 2010). SentiWordNet contains positivity, negativity, and objectivity scores for each sense of a word. We explain below how this information was used for each task. As classifiers, we used Support Vector Ma- chines (SVM) (SMO and libSVM from Weka with default values for parameters), because SVM is known to perform well on many tasks, and Multi- nomial Naive Bayes (MNB), because MNB is known to perform well on text data and it is faster than SVM. 2.1 Task A Our system for Task A involved two parts: the expansion of our training data and the classifica- tion. The expansion was done with information from SentiWordNet. Stop words and words that appeared only once in the training data were fil- tered out. Then the classification was completed with algorithms from Weka. As mentioned, the first task was to expand all of the tweets that were provided as training data. This was doing using Python and the Python NLTK library, as well as SentiWordNet. SentiWordNet provides a score of the sentient state for each word …

Content truncated to 500 words