CUFE at SemEval-2016 Task 4: A Gated Recurrent Model for Sentiment Classification

Proceedings of SemEval-2016, pages 52–57, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics CUFE at SemEval-2016 Task 4: A Gated Recurrent Model for Sentiment Classification Mahmoud Nabil1, Mohamed Aly2 and Amir F. Atiya3 1,2,3Computer Engineering, Cairo University, Egypt 2Visual Computing Center, KAUST, KSA 1mah.nabil@cu.edu.eg 2mohamed@mohamedaly.info 3amir@alumni.caltech.edu Abstract In this paper we describe a deep learn- ing system that has been built for SemEval 2016 Task4 (Subtask A and B). In this work we trained a Gated Recurrent Unit (GRU) neural network model on top of two sets of word embeddings: (a) general word embed- dings generated from unsupervised neural lan- guage model; and (b) task specific word em- beddings generated from supervised neural language model that was trained to classify tweets into positive and negative categories. We also added a method for analyzing and splitting multi-words hashtags and appending them to the tweet body before feeding it to our model. Our models achieved 0.58 F1-measure for Subtask A (ranked 12/34) and 0.679 Recall for Subtask B (ranked 12/19). 1 Introduction Twitter is a huge microbloging service with more than 500 million tweets per day1 from different lo- cations in the world and in different languages. This large, continuous, and dynamically updated content is considered a valuable resource for researchers. However many issues should be taken into account while dealing with tweets, namely: (1) informal lan- guage used by the users; (2) spelling errors; (3) text in the tweet may be referring to images, videos, or external URLs; (4) emoticons; (5) hashtags used (combining more than one word as a single word); (6) usernames used to call or notify other users; (7) 1http://internetlivestats.com/ twitter-statistics/ spam or irrelevant tweets; and (8) character limit for a tweet to 140 characters. This poses many chal- lenges when analyzing tweets for natural language processing tasks. In this paper we describe our sys- tem used for SemEval 2016 (Nakov et al., 2016b) Subtasks A and B. Subtask A (Message Polarity Classification) requires classifying a tweet’s senti- ment as positive; negative; or neutral,. Subtask B (Tweet classification according to a two-point scale) requires classifying a tweet’s sentiment given a topic as positive or negative. Our system uses a GRU neu- ral network model (Bahdanau et al., 2014) with one hidden layer on top of two sets of word embeddings that are slightly fine-tuned on each training set (see Fig. 1). The first set of word embeddings is con- sidered as general purpose embeddings and was ob- tained by training word2vec (Mikolov et al., 2013) on 20.5 million tweets that we crawled for this pur- pose. The second set of word embeddings is consid- ered as task specific set, and was obtained by train- ing on a supervised sentiment analysis dataset using another GRU model. We also added a method for analyzing multi-words hashtags by splitting them and appending them to the body of the tweet before feeding it to the GRU model. In our experiments we tried both keeping the word embeddings static during …

Content truncated to 500 words