Learning Emotion-enriched Word Representations

Proceedings of the 27th International Conference on Computational Linguistics, pages 950–961 Santa Fe, New Mexico, USA, August 20-26, 2018. 950 Learning Emotion-enriched Word Representations Ameeta Agrawal, Aijun An and Manos Papagelis Department of Electrical Engineering and Computer Science York University, Toronto, Canada {ameeta,aan,papaggel}@eecs.yorku.ca Abstract Most word representation learning methods are based on the distributional hypothesis in linguis- tics, according to which words that are used and occur in the same contexts tend to possess similar meanings. As a consequence, emotionally dissimilar words, such as “happy” and “sad” occurring in similar contexts would purport more similar meaning than emotionally similar words, such as “happy” and “joy”. This complication leads to rather undesirable outcome in predictive tasks that relate to affect (emotional state), such as emotion classification and emotion similarity. In order to address this limitation, we propose a novel method of obtaining emotion-enriched word rep- resentations, which projects emotionally similar words into neighboring spaces and emotionally dissimilar ones far apart. The proposed approach leverages distant supervision to automatically obtain a large training dataset of text documents and two recurrent neural network architectures for learning the emotion-enriched representations. Through extensive evaluation on two tasks, including emotion classification and emotion similarity, we demonstrate that the proposed repre- sentations outperform several competitive general-purpose and affective word representations. 1 Introduction Emotion detection from text is the task of identifying emotions from natural language data such as re- views, blogs, news articles, and so on (Alm et al., 2005; Aman and Szpakowicz, 2007). While numerous taxonomies of emotions have been proposed (Ekman, 1992; Plutchik, 1980; Parrott, 2001), most psy- chologists agree that an emotion is a feeling that characterizes the state of mind such as happiness, sad- ness, anger, among others. The ability to detect emotions in text is critical for a number of applications and services in diverse domains, including market research, customer relations, gaming, and intelligent tutoring systems, to name a few (Mohammad and Turney, 2013). Despite its potentially wide-spread use, the automatic detection of emotions remains a challenging multi-class, sometimes multi-label, classification problem due to a number of reasons, including: (i) different emotion models consist of different number and types of emotion categories; (ii) emotions are not only subjective but also fuzzy, with more than one emotion occurring at the same time. As a result, development of emotion related resources, such as training data, has been limited to a few manually annotated datasets or lexicons, a process that requires much time and effort, and is expensive. To solve the limited training data problem, the recent success of word embeddings has garnered in- creased attention in the design of emotion classification systems (Bravo-Marquez et al., 2016; Pool and Nissim, 2016; Mohammad and Bravo-Marquez, 2017). Word embeddings are distributed word represen- tations (Collobert et al., 2011; Turian et al., 2010), where each word w in the vocabulary V is mapped into a dense, low-dimensional, continuous-valued vector x ∈ Rd, d ≪ |V|. The underlying idea is that words that frequently occur together in same contexts …

Content truncated to 500 words