KU-MTL at SemEval-2018 Task 1: Multi-task Identification of Affect in Tweets

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 385–389 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics KU-MTL at SemEval-2018 Task 1: Multi-task Identification of Affect in Tweets Thomas Nyegaard-Signori, Casper Veistrup Helms, Johannes Bjerva, Isabelle Augenstein Department of Computer Science University of Copenhagen {sfq340,wqx727}@alumni.ku.dk, {bjerva,augenstein}@di.ku.dk Abstract We take a multi-task learning approach to the shared Task 1 at SemEval-2018. The general idea concerning the model structure is to use as little external data as possible in order to preserve the task relatedness and reduce com- plexity. We employ multi-task learning with hard parameter sharing to exploit the related- ness between sub-tasks. As a base model, we use a standard recurrent neural network for both the classification and regression subtasks. Our system ranks 32nd out of 48 participants with a Pearson score of 0.557 in the first sub- task, and 20th out of 35 in the fifth subtask with an accuracy score of 0.464. 1 Introduction We consider the task of identifying affect in tweets, as described in Mohammad et al. (2018). Given a tweet, the task is to predict the emo- tions and their corresponding intensities which the tweet portrays. Previous approaches to this task are outlined in Mohammad and Bravo-Marquez (2017). The winning team of the SemEval EmoInt 2017, presented in Goel et al. (2017), tackled a similar task as the regression task presented in this year’s SemEval Task 1. The winning sys- tem utilised an ensemble approach consisting of 5 sub-models and using a weighted average of these models to come up with the final result. This model is utilising most of the different approaches mentioned in the literature and combining them into one and with great success. Our work bears resemblance to the runner up in the SemEval EmoInt 2017, K¨oper et al. (2017), who used a comparatively simple model consist- ing of a CNN-LSTM neural network. The differ- ence between the models presented in this paper and the IMS system is the utilisation of lexicons, and that we take a multi-task learning approach. We focus on two subtasks of this shared task, namely emotion intensity regression and emotion classification (Sub-tasks 1 and 5). These were chosen because of the overlap in tweets but differ- ing truth labels value and types. Furthermore, we only consider the English versions of the subtasks. 2 System Description Our system is a standard RNN, with the exception that we approach the task using multi-task learning via hard parameter sharing (Caruana, 1993). We will now present the details of our implementation. 2.1 Preprocessing of Text and Textual Representation We use embedded word representations as input to our system, initialised to the weights from a large set of pre-trained embeddings which have been trained on Twitter data (∼400 million tweets). These embeddings were obtained from Godin et al. (2015). Words which are out-of-vocabulary are replaced with UNK. Furthermore, user men- tions and numbers get mapped to a place-holder instead of their actual values. Since the model …

Content truncated to 500 words