thecerealkiller at SemEval-2016 Task 4: Deep Learning based System for Classifying Sentiment of Tweets on Two Point Scale

Proceedings of SemEval-2016, pages 100–102, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics thecerealkiller at SemEval-2016 Task 4: Deep Learning based System for Classifying Sentiment of Tweets on Two Point Scale Vikrant Yadav Amazon.in Hyderabad, India vikrantiitr1@gmail.com Abstract In this paper, we propose a deep learning sys- tem for classification of tweets on a two-point scale. Our architecture consists of a multilay- ered recurrent neural network having gated re- current units. The network is pre-trained with a weakly labeled dataset of tweets to learn the sentiment specific embeddings. Then it is fine tuned on the given training dataset of the task 4B in SemEval-2016. The network does very little pre-processing for raw tweets and no post-processing at all. The proposed sys- tem achieves 3rd rank on the leaderboard of task 4B. 1 Introduction Task 4 of SemEval-2016 (Nakov et al., 2016) - Sentiment Analysis in Twitter- turned out to be the most popular task of SemEval-2016. Among its sub- tasks, sub-task B - Tweet classification according to a two-point scale - was the 2nd most popular sub- task. A total of 19 teams participated in it including ours. In this paper, we propose a multilayerd RNN architecture for classifying tweets on a two-point scale, namely, positive and negative. We first pre- train the network with a weakly labeled corpus of tweets, where labels are assigned based on the senti- ment of the emoticon present in the tweets. It helps the network to learn sentiment specific embeddings of the words in the tweet. The network is then fine- tuned on the dataset of tweets provided as part of the sub-task 4B along with the training and development dataset of SemEval-2013 task 2. In the second section, we describe our architec- ture. In third section, we explain our approach to train the network. After that, we describe the exper- imental setup and statistical properties of the data used. In the end, we discuss our results on the test dataset. We stood 3rd on the final leaderboard for sub-task 4B. 2 Proposed System In this section, we elaborate our proposed architec- ture. The network is fed pre-processed tweets as in- put and it predicts the binary label of the tweets. It has 3 RNN layers each with gated recurrent units. Then a sum layer is added to sum the hidden states of the last recurrent layer over time. After that, a dense layer is present followed by a single sigmoid unit. Now, we describe the components of our architec- ture in brief. 2.1 Pre-processing The tweets are pre-processed to remove any punctu- ation if present. All URLs are encoded into a URL token. All the user accounts mentioned in a tweet are encoded as USER token. The tweet is converted to lower-case before feeding to the network. Note that we don’t remove any stop-words as they define use- ful relationships between words and phrases. This amount of pre-processing turned out to be sufficient for the network to learn useful semantic …

Content truncated to 500 words