ParallelDots at SemEval-2019 Task 3: Domain Adaptation with feature embeddings for Contextual Emotion Analysis

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 185–189 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 185 ParallelDots at SemEval-2019 Task 3: Domain Adaptation with feature embeddings for Contextual Emotion Analysis Akansha Jain1 Paralledots, Inc. Ishita Aggarwal1 Paralledots, Inc. 1akansha, ishita, ankit@paralleldots.com Ankit Narayan Singh1 Paralledots, Inc. Abstract This paper describes our proposed system & experiments performed to detect contextual emotion in texts for SemEval 2019 Task 3. We exploit sentiment information, syntactic pat- terns & semantic relatedness to capture diverse aspects of the text. Word level embeddings such as Glove, FastText, Emoji along with sentence level embeddings like Skip-Thought, DeepMoji & Unsupervised Sentiment Neuron were used as input features to our architecture. We democratize the learning using ensembling of models with different parameters to produce the final output. This paper discusses compar- ative analysis of the significance of these em- beddings and our approach for the task. 1 Introduction Emotion Classification is more nuanced version of Sentiment Analysis. While Sentiment Anal- ysis gives you a general idea about user expe- rience by categorizing statements into positive or negative, Emotion Classification extracts spe- cific attributes about each of these 2 categories. Contextual Emotion Classification needs to keep the context of an ongoing conversation to pre- dict their emotional state and therefore comes with its own challenges. Detecting emotions has be- come a crucial part of understanding user gener- ated content and to generate emotion aware re- sponses. This paper describes our approach for SemEval 2019 Task 3: EmoContext. The task is Emotion Classification in the conversational sce- nario. Complete details about the task, evalua- tion and dataset can be found in paper released by organizers (Chatterjee et al., 2019). We use various state-of-the-art Machine Learning models and perform domain adaptation (Pan and Yang, 2010) from their source task to the SemEval Emo- Context task. Our solution uses multiple types of feature embeddings viz Skip-Thought vectors (Kiros et al., 2015), Unsupervised Sentiment Neu- ron (Radford et al., 2017), DeepMoji’s attention and last layer (Felbo et al., 2017) embedding along with Glove (Pennington et al., 2014), Emoji Em- bedding (Eisner et al., 2016) and FastText (Joulin et al., 2016). These feature embeddings are passed to a Deep Learning architecture. We train mul- tiple models with different hyper-parameters. Fi- nally, the results from each models are stacked to- gether in an ensemble (Polikar, 2006). Our main approach for the literature survey was to look for similar research work used in previous SemEval tasks and other published state-of-the-art method- ologies in the same domain. Infact, SemEval 2018 task of finding Affect in Tweet (Mohammad et al., 2018) demonstrates how detection of emo- tion plays an important role in understanding con- tent as well as its creators. It was helpful to learn about how various architectures such as Siamese (Ghosh and Veale, 2018), CNNs (Khosla, 2018) and Deeply connected LSTMs (Wu et al., 2018) can be used to effectively learn emotional context of text. Methodologies such as ensembling …

Content truncated to 500 words