LIA at SemEval-2017 Task 4: An Ensemble of Neural Networks for Sentiment Classification

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 760–765, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics LIA at SemEval-2017 Task 4: An Ensemble of Neural Networks for Sentiment Classification Mickael Rouvier LIA University of Avignon Avignon, France mickael.rouvier@univ-avignon.fr Abstract This paper describes the system developed at LIA for the SemEval-2017 evaluation cam- paign. The goal of Task 4.A was to iden- tify sentiment polarity in tweets. The sys- tem is an ensemble of Deep Neural Network (DNN) models: Convolutional Neural Net- work (CNN) and Recurrent Neural Network Long Short-Term Memory (RNN-LSTM). We initialize the input representation of DNN with different sets of embeddings trained on large datasets. The ensemble of DNNs are com- bined using a score-level fusion approach. The system ranked 2nd at SemEval-2017 and obtained an average recall of 67.6%. 1 Introduction This paper describes the system developed at LIA for the SemEval-2017 sentiment analysis evaluation task 4 (Rosenthal et al., 2017). We have participated in Subtask A: sentiment analysis at the message level in English. It consists in determining the message polarity of each tweet in the test set. The sentiment polarity classification task is set as a three-class problem: positive, nega- tive and neutral. The sentiment analysis task is often modeled as a classification problem which relies on features ex- tracted from the text in order to feed a classifier. Recent work has shown that Deep Neural Networks (DNN) using word representations as input are well suited for sentence classification problems and have been shown to produce state-of-the-art results for sentiment polarity classification (Tang et al., 2014a; Severyn and Moschitti, 2015). Two different types of DNN models are used: Convolutional Neural Network (CNN) and Recurrent Neural Network with Long Short-Term Memory units (RNN-LSTM). Pre- trained word embeddings are used to initialize the word representations, which are then taken as input of a text. Our approach consists in learning classifiers for four types of embeddings, based on the CNN and RNN-LSTM architectures. Each set of word embed- dings models the tweet according to a different point of view. A final fusion step is applied. Our contributions are as follows: • We propose to apply a teacher-student ap- praoch for training the DNN models. • We propose a new way to capture polarity in word embeddings. • The source code of our system, the models trained for the evaluation, as well as the corpus collected for creating word embeddings, are all made available to the community in hope of helping future research 1. The paper is structured as follows. Section 2 presents a quick overview of the system architec- ture, which is then detailed in sections 3 and 4, along with the various word embeddings and other fea- tures used in our system. Results and discussion ap- pear in Section 5. 2 Overview of the approach The system was developed as a two-level architec- ture. Given a tweet, the first level extracts input rep- resentations based various word embeddings. These …

Content truncated to 500 words