UniPI at SemEval-2016 Task 4: Convolutional Neural Networks for Sentiment Classification

Proceedings of SemEval-2016, pages 220–224, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics UniPI at SemEval-2016 Task 4: Convolutional Neural Networks for Sen- timent Classification Giuseppe Attardi, Daniele Sartiano Dipartimento di Informatica Università di Pisa Italy {attardi,sartiano}@di.unipi.it Abstract The paper describes our submission to the task on Sentiment Analysis on Twitter at SemEval 2016. The approach is based on a Deep Learning architecture using convolu- tional neural networks. The approach used on- ly word embeddings as features. The submis- sion used embeddings created from a corpus of news articles. We report on further experi- ments using embeddings built for a corpus of tweets as well as sentiment specific word em- beddings obtained by distant supervision. 1 Introduction Up until recently, the typical approaches to senti- ment analysis of tweets were based on classifiers trained using several hand-crafted features, in par- ticular lexicons of words with an assigned polarity value. At the SemEval 2015 task 10 on Sentiment Analysis of Twitter (Rosenthal et al., 2015), most systems relied on features derived from sentiment lexicons. Other important features included bag-of- words features, hash-tags, handling of negation, word shape and punctuation features, elongated words, etc. Moreover, tweet pre-processing and normalization were an important part of the pro- cessing pipeline. Quite significantly, the top scoring system in subtask A: Phrase-Level Polarity (Moschitti and Severyn, 2015) was instead based on the use of a convolutional neural network, which used word embeddings as its only features. Word embeddings were created by unsupervised learning from a col- lection of 50 million tweets, using the SkipGram model by Mikolov et al, (2013). The tweets used for training were collected by querying the Twitter API on the presence of a set of emoticons repre- senting positive or negative sentiment. The win- ning team achieved an F1 of 84.79 on the Twit- ter2015 test set. The team participated with a simi- lar approach also to subtask B: Message-Level Po- larity, achieving the second best score with an F1 of 64.59. The fourth F1 score of 64.17 was achieved also by a system exploiting word embed- dings by INESC-ID. The top scoring system in- stead consisted in an ensemble combining four Twitter sentiment classification approaches that participated in previous editions, with an F1 of 64.84. We decided to explore a similar approach for tackling SemEval 2016 task 4 on Sentiment Analy- sis in Twitter (Nakov et al., 2016). 2 Approach The classifier is based on a Deep Learning archi- tecture consisting of the following layers: 1. a lookup layer for word embeddings 2. a convolutional layer with relu activation 3. a maxpooling layer 4. a dropout layer 5. a linear layer with tanh activation 6. a softmax layer 220 The architecture is illustrated in Figure 1. The first layer looks up the embedding for words, i.e. vec- tors in Rd. In our experiments we set d = 300. The convolutional layer consists of multiple convolu- tional filters of sliding windows of various sizes. These are then combined through …

Content truncated to 500 words