SENSEI-LIF at SemEval-2016 Task 4: Polarity embedding fusion for robust sentiment analysis

Proceedings of SemEval-2016, pages 202–208, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics SENSEI-LIF at SemEval-2016 Task 4: Polarity embedding fusion for robust sentiment analysis Mickael Rouvier Aix-Marseille Universit´e CNRS, LIF UMR 7279 13000, Marseille, France mickael.rouvier@lif.univ-mrs.fr Benoit Favre Aix-Marseille Universit´e CNRS, LIF UMR 7279 13000, Marseille, France benoit.favre@lif.univ-mrs.fr Abstract This paper describes the system developed at LIF for the SemEval-2016 evaluation cam- paign. The goal of Task 4.A was to iden- tify sentiment polarity in tweets. The system extends the Convolutional Neural Networks (CNN) state of the art approach. We ini- tialize the input representations with embed- dings trained on different units: lexical, part- of-speech, and sentiment embeddings. Neural networks for each input space are trained sep- arately, and then the representations extracted from their hidden layers are concatenated as input of a fusion neural network. The system ranked 2nd at SemEval-2016 and obtained an average F1 of 63.0%. 1 Introduction This paper describes the system developed at LIF for the SemEval-2016 sentiment analysis evaluation task (Nakov et al., 2016). The goal of our partic- ipation was to apply approaches developed for the European FP7 project SENSEI 1 based on the study of human conversations according to feelings, opin- ions, emotions of the participants, in corpora such as transcripts of telephone speech and web comments. We have participated in Subtask A: sentiment analysis at the message level. It consists in deter- mining the message polarity of each tweet in the test set. The sentiment polarity classification task is set as a three-class problem: positive, negative and neu- tral. The sentiment analysis task is often modeled as a classification problem which relies on features ex- 1http://www.sensei-conversation.eu/ tracted from the text in order to feed a classifier. Recent work has shown that Convolutional Neural Networks (CNN) using word representations as in- put are well suited for sentence classification prob- lems (Kim, 2014) and have been shown to produce state-of-the-art results for sentiment polarity classi- fication (Tang et al., 2014a; Severyn and Moschitti, 2015). Pre-trained word embeddings are used to initialize the word representations, which are then taken as input of a text CNN. Our approach consists in learning polarity clas- sifiers for three types of embeddings, based on the same CNN architecture. Each set of word embed- ding models the tweet according to a different point of view: lexical, part-of-speech and sentiment. A final fusion step is applied, based on concatenating the hidden layers of the CNNs and training a deep neural network for the fusion. Our contributions are as follows: • We extend the deep CNN architecture proposed in (Poria et al., 2015) and introduce lexical in- formation similar to (Ebert et al., 2015). • We introduce polarity embeddings, tweet rep- resentations extracted from the hidden layer of CNNs with different word embeddings as in- put. • We fuse polarity embeddings by concatenat- ing them and feeding them to a neural network trained on the final task. • The source code of our system, the models trained …

Content truncated to 500 words