Epita at SemEval-2018 Task 1: Sentiment Analysis Using Transfer Learning Approach

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 151–155 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics Epita at SemEval-2018 Task 1: Sentiment Analysis Using Transfer Learning Approach Guillaume Daval-Frerot Abdessalam Bouchekif Graduate school of computer science, EPITA, France firstname.lastname@epita.fr Anatole Moreau Abstract In this paper we present our system for detec- ting valence task. The major issue was to ap- ply a state-of-the-art system despite the small dataset provided : the system would quickly overfit. The main idea of our proposal is to use transfer learning, which allows to avoid lear- ning from scratch. Indeed, we start to train a first model to predict if a tweet is positive, ne- gative or neutral. For this we use an external dataset which is larger and similar to the target dataset. Then, the pre-trained model is re-used as the starting point to train a new model that classifies a tweet into one of the seven various levels of sentiment intensity. Our system, trained using transfer learning, achieves 0.776 and 0.763 respectively for Pearson correlation coefficient and weighted quadratic kappa metrics on the subtask evalua- tion dataset. 1 Introduction The goal of detecting valence task is to clas- sify a given tweet into one of seven classes, cor- responding to various levels of positive and ne- gative sentiment intensity, that best represents the mental state of the tweeter. This can be seen as a multiclass classification problem, in which each tweet must be classified in one of the follo- wing classes : very negative (-3), moderately nega- tive (-2), slightly negative (-1), neutral/mixed (0), slightly positive (1), moderately positive (2) and very positive (3) (Mohammad et al., 2018). Several companies have been interested in cus- tomer opinion for a given product or service. Sen- timent analysis is one approach to automatically detect their emotions from comments posted in so- cial networks. With the recent advances in deep learning, the ability to analyse sentiments has considerably improved. Indeed, many experiments have used state-of-the-art systems to achieve high perfor- mance. For example, (Baziotis et al., 2017) use Bi- directional Long Short-Term Memory (B-LSTM) with attention mechanisms while (Deriu et al., 2016) use Convolutional Neural Networks (CNN). Both systems obtained the best performance at the the 2016 and 2017 SemEval 4-A task respectively. The amount of data is argued to be the main condition to train a reliable deep neural network. However, the dataset provided to build our system is limited. To address this issue, two solutions can be considered. The first solution consists in exten- ding our dataset by either manually labeling new data, which can be very time consuming, or by using over-sampling approaches. The second solu- tion consists in applying a transfer learning, which allows to avoid learning the model from scratch. In this paper, we apply a transfer learning approach, from a model trained on a similar task : we propose to pre-train a model to predict if a tweet is positive, negative or neutral. …

Content truncated to 500 words