UofL at SemEval-2016 Task 4: Multi Domain word2vec for Twitter Sentiment Classification

Proceedings of SemEval-2016, pages 164–170, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics UofL at SemEval-2016 Task 4: Multi Domain word2vec for Twitter Sentiment Classification Omar Abdelwahab, Adel Elmaghraby Computer Science and Engineering University of Louisville, Louisville, KY {omar.abdelwahab,adel}@louisville.edu Abstract In this paper, we present a transfer learning sys- tem for twitter sentiment classification and compare its performance using different fea- ture sets that include different word represen- tation vectors. We utilized data from a different source domain to increase the performance of our system in the target domain. Our approach was based on training various word2vec mod- els on data from the source and target domains combined, then using these models to calculate the average word vector of all the word vectors in a tweet observation, then input the average word vector as a feature to our classifiers for training. We further developed one doc2vec model that was trained on the positive, nega- tive and neutral tweets in the target domain only. We then used these models in calculating the average word vector for every tweet in the training set as a preprocessing step. The final evaluation results show that our approach gave a prediction accuracy on the Twitter2016 test dataset that outperformed two teams that were among the top 10 in terms of AvgF1 scores. 1 Introduction Twitter sentiment analysis deals with classifying the polarity of a tweet as positive or negative or neutral. We have participated in Semeval 2016 Twitter Sen- timent Analysis subtask-A. Where we had to predict the polarity of a whole tweet rather than part of a tweet. We have started off with feature engineering and data preprocessing. Then we have divided our task into five classification tasks. Then we worked on creating our feature sets that we later used to se- lect the best classifier/feature set combination for each classification task as will be explained further in the paper. In section 2 we explain our transfer learning ap- proach, system components, data preprocessing, ag- gregation approach, word2vec training and doc2vec training. In section 3, we present our experiments. Then in section 4, we review and discuss our results. Finally in section 5, we make conclusions and out- line some future work. 2 Transfer Learning Approach Our transfer learning approach was based on using data from an additional source domain to supple- ment the provided target domain twitter data to train our word2vec models. The resultant models were later used in calculating the average word2vec vec- tor of each tweet in our twitter training set as will be explained in section 2.4. We have also developed four binary classifiers and one arbiter three class la- bel classifier. The four binary classifiers as well as the fifth arbiter classifier were a combination of lo- gistic regression and SVM classifiers as we will show in the following sections. 2.1 System Components and Data Sets In this section, we will specify the system compo- nents that were used in predicting the final results that …

Content truncated to 500 words