Finki at SemEval-2016 Task 4: Deep Learning Architecture for Twitter Sentiment Analysis

Proceedings of SemEval-2016, pages 149–154, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics Finki at SemEval-2016 Task 4: Deep Learning Architecture for Twitter Sentiment Analysis Dario Stojanovski, Gjorgji Strezoski, Gjorgji Madjarov, Ivica Dimitrovski Faculty of Computer Science and Engineering Ss. Cyril and Methodius University Rugjer Boshkovikj 16 1000 Skopje, Republic of Macedonia {stojanovski.dario, strezoski.g}@gmail.com {gjorgji.madjarov, ivica.dimitrovski}@finki.ukim.mk Abstract In this paper, we present a novel deep learning architecture for sentiment analysis in Twitter messages. Our system finki, employs both convolutional and gated recurrent neural net- works to obtain a more diverse tweet repre- sentation. The network is trained on top of GloVe word embeddings pre-trained on the Common Crawl dataset. Both neural net- works are used to obtain a fixed length rep- resentation of variable sized tweets, and the concatenation of these vectors is supplied to a fully connected softmax layer with dropout regularization. The system is evaluated on benchmark datasets from the Sentiment Anal- ysis in Twitter task of the SemEval 2016 chal- lenge where our model achieves best and sec- ond highest results on the 2-point and 5-point quantification subtasks respectively. Despite not relying on any hand-crafted features, our system manages the second highest average rank on the considered subtasks. 1 Introduction Twitter sentiment analysis is an area of Natural Lan- guage Processing (NLP) dealing with the classifi- cation of sentiment polarity in Twitter messages. Most of the approaches to this problem are generally based on hand crafted features and sentiment lexi- cons (Mohammad et al., 2013; Pak and Paroubek, 2010). These features are then used as input to classifying algorithms such as, Support Vector Ma- chines (SVM) and naive Bayes classifier. However, such approaches require extensive domain knowl- edge, are laborious to define, and can lead to incom- plete or over-specific features. Deep learning methods for sentiment analysis, on the other hand, handle the feature extraction automatically which provides for robustness and adaptability. Notably, most popular deep learning methods are convolutional neural networks (CNN), which have been shown to achieve state-of-the-art results (Kim, 2014; dos Santos and Gatti, 2014), though some works propose different models such as Recursive Neural Tensor Network (Socher et al., 2013). Recurrent neural networks (RNN) are intuitive ar- chitectures for NLP as they inherently take into ac- count the ordering of words in the text as opposed to CNNs which take only a small limited context window. However, to our knowledge, these net- works have not been applied to sentiment analysis in Twitter messages. Le et al. (Le and Zuidema, 2015) report state-of-the-art results with Long Short Term Memory (LSTM) networks on binary and fine-grained classification on the Stanford Senti- ment Treebank dataset. In this paper, we present a novel deep learning ar- chitecture for sentiment classification and quantifi- cation in Twitter messages. The model consists of a convolutional and a gated recurrent neural network (GRNN). Both neural networks are used to model a suitable representation of a tweet. The feature representations output from the networks are fused and fed to …

Content truncated to 500 words