Distributed Representations of Emotion Categories in Emotion Space

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 2364–2375 August 1–6, 2021. ©2021 Association for Computational Linguistics 2364 Distributed Representations of Emotion Categories in Emotion Space Xiangyu Wang1,2 and Chengqing Zong1,2,3∗ 1National Laboratory of Pattern Recognition, Institute of Automation, CAS 2School of Artificial Intelligence, University of Chinese Academy of Sciences 3CAS Center for Excellence in Brain Science and Intelligence Technology {xiangyu.wang, cqzong}@nlpr.ia.ac.cn Abstract Emotion category is usually divided into dif- ferent ones by human beings, but it is in- deed difficult to clearly distinguish and define the boundaries between different emotion cat- egories. The existing studies working on emo- tion detection usually focus on how to improve the performance of model prediction, in which emotions are represented with one-hot vectors. However, emotion relations are ignored in one- hot representations. In this article, we first propose a general framework to learn the dis- tributed representations for emotion categories in emotion space from a given emotion classifi- cation dataset. Furthermore, based on the soft labels predicted by the pre-trained neural net- work model, we derive a simple and effective algorithm. Experiments have validated that the proposed representations in emotion space can express emotion relations much better than word vectors in semantic space. 1 Introduction In the past decades, a lot of tasks have been pro- posed in the field of text emotion analysis. The most primary one among them is emotion classifica- tion task (Alm et al., 2005). Based on emotion clas- sification task, many new tasks have been proposed from different considerations. Lee et al. (2010) pro- posed the task of emotion cause extraction, which aims at predicting the reason of a given emotion in a document. Based on the emotion cause ex- traction task, Xia and Ding (2019) introduced the emotion-cause pair extraction task for the purpose of extracting the potential pairs of emotions and corresponding causes in a document. Jiang et al. (2011) proposed a target-dependent emotion recog- nition task, which aims at predicting the sentiment with the given query. To express the intensity of ∗ Corresponding author. a specific emotion in text, Mohammad and Bravo- Marquez (2017) proposed the emotion intensity detection task. However, all the above tasks treat emotions as independent ones and represent emo- tions with one-hot vectors, which definitely ignore the underlying emotion relations. Based on existing emotion detection tasks, many efforts have been made to achieve better perfor- mance (Danisman and Alpkocak, 2008; Xia et al., 2011; Kim, 2014; Xia et al., 2015; Li et al., 2018; Zong et al., 2019) and many datasets have been introduced to train and evaluate the correspond- ing models (Ghazi et al., 2015; Mohammad et al., 2018; Liu et al., 2019). The vast majority of ex- isting emotion annotation work assumes that the emotions are orthogonal to each other and repre- sent the emotion categories with one-hot vectors (Mohammad, 2012; Gui et al., 2016; Klinger et al., 2018). Actually, the boundaries as well …

Content truncated to 500 words