Emotion Representation Mapping for Automatic Lexicon Construction (Mostly) Performs on Human Level

Proceedings of the 27th International Conference on Computational Linguistics, pages 2892–2904 Santa Fe, New Mexico, USA, August 20-26, 2018. 2892 Emotion Representation Mapping for Automatic Lexicon Construction (Mostly) Performs on Human Level Sven Buechel & Udo Hahn {sven.buechel|udo.hahn}@uni-jena.de Jena University Language & Information Engineering (JULIE) Lab Friedrich-Schiller-Universit¨at Jena, Jena, Germany http://www.julielab.de Abstract Emotion Representation Mapping (ERM) has the goal to convert existing emotion ratings from one representation format into another one, e.g., mapping Valence-Arousal-Dominance annota- tions for words or sentences into Ekman’s Basic Emotions and vice versa. ERM can thus not only be considered as an alternative to Word Emotion Induction (WEI) techniques for automatic emo- tion lexicon construction but may also help mitigate problems that come from the proliferation of emotion representation formats in recent years. We propose a new neural network approach to ERM that not only outperforms the previous state-of-the-art. Equally important, we present a refined evaluation methodology and gather strong evidence that our model yields results which are (almost) as reliable as human annotations, even in cross-lingual settings. Based on these re- sults we generate new emotion ratings for 13 typologically diverse languages and claim that they have near-gold quality, at least. 1 Introduction From its inception, researchers in the field of sentiment analysis aimed at predicting the affective state that is typically associated with a given word based on a list of linguistic features, a problem referred to as word emotion induction (WEI) (Hatzivassiloglou and McKeown, 1997; Turney and Littman, 2003). Early research activities have focused on semantic polarity (the positiveness or negativeness of a feeling) for quite a long time. But more recently this focus on binary representations has been replaced by more expressive emotion representation formats such as Basic Emotions or Valence-Arousal-Dominance. In the meantime, WEI has become an active area of research, regularly featured in shared tasks (Rosenthal et al., 2015; Yu et al., 2016b). Based on these achievements, WEI techniques have become a natural methodological choice for the automatic construction of emotion lexicons (K¨oper and Schulte im Walde, 2016; Shaikh et al., 2016). Yet, only very recently, a radically different approach to automatic emotion lexicon construction has been proposed. Instead of relying on linguistic features (such as similarity with seed words or word em- beddings), the goal of emotion representation mapping (ERM) is to derive new emotional word ratings in one format based on known ratings of the same words in another format (Buechel and Hahn, 2017a). For example, ERM could use empirically gathered ratings for Basic Emotions and convert them into a Valence-Arousal-Dominance representation scheme, with greater precision than currently achievable by WEI algorithms. As a much appreciated side effect, one of the promises of ERM is to make otherwise incompatible resources (lexicons or annotated corpora, as well as tools) compatible, and incomparable systems comparable. Thus, this approach has the potential to mitigate some of the negative effects that arise from not having a community-wide standard for emotion annotation and representation (Calvo and Mac Kim, 2013; Buechel and Hahn, 2018a). …

Content truncated to 500 words