NL-FIIT at SemEval-2019 Task 3: Emotion Detection From Conversational Triplets Using Hierarchical Encoders

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 272–276 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 272 NL-FIIT at SemEval-2019 Task 3: Emotion Detection From Conversational Triplets Using Hierarchical Encoders Michal Farkas, Peter Lacko Slovak University of Technology in Bratislava Faculty of Informatics and Information Technologies Ilkovicova 2, 842 16 Bratislava, Slovakia michal.farkas@stuba.sk, peter.lacko@stuba.sk Abstract In this paper, we present our system submis- sion for the EmoContext, the third task of the SemEval 2019 workshop. Our solution is a hierarchical recurrent neural network with ELMo embeddings and regularization through dropout and Gaussian noise. We have mainly experimented with two main model architec- tures: simple and hierarchical LSTM network. We have also examined ensembling of the models and various variants of an ensemble. We have achieved microF1 score of 0.7481, which is significantly higher than baseline and currently the 19th best submission. 1 Introduction Sentiment analysis has a long and successful his- tory in the context of natural language processing. As with the majority of the problems in this do- main, we have seen a gradual shift towards so- lutions based on neural models. Nowadays, such models can be readily used as a part of a larger so- lution, for example to analyse communication on social networks. Then, perhaps unsurprisingly, the EmoContext task (Chatterjee et al., 2019b) with its format is highly evocative of the social networks. That is, it consists of conversational triplets, where the task is to correctly guess the emotion category of the last conversational turn. Over the years, there was a number of differ- ent sentiment analysis and recognition competi- tions, workshops and shared tasks (Klinger et al., 2018; Rosenthal et al., 2017), however the conver- sational nature of the data is not common. Our system is based on the recurrent neural networks, both simple and hierarchical architec- tures. We have experimented with various tech- niques and hyper-parameters, such as regular- ization, class weights, embeddings, strength of dropout and added noise. Finally, we have im- proved our results by creating an ensemble, with various voting methods and sample reweighing. 2 Approach The first step in any natural language processing system is to preprocess the input data so it can be easily understood by the system. Since prepro- cessing was the major part of our previous work (Pecar et al., 2018), we have decided to prior- itize work on the model instead. Nevertheless, we need to do some kind of preprocessing, hence we used the readily available Ekphrasis (Baziotis et al., 2017) tool. We have experimented with both standard GloVe embeddings (Pennington et al., 2014) and more sophisticated ELMo embeddings (Peters et al., 2018). The improvements contextual em- beddings, like ELMo, can bring are already well known and there is little need for additional com- parisons, however the model that uses standard embeddings is considerably faster and hence more useful for quick experiments. Since training ELMo can be quite time consuming and resource intense, we have opted for pretrained models that …

Content truncated to 500 words