Hope is a dangerous thing, but I have it.

【论文阅读】Joint Learning for Emotion Classification and Emotion Cause Detection (EMNLP 2018)

Basic Information

  • Name: Joint Learning for Emotion Classification and Emotion Cause Detection
  • Authors: Ying Chen , Wenjun Hou , Xiyao Cheng , Shoushan Li
  • URL:  http://aclweb.org/anthology/D18-1066
  • Cite:  Chen Y, Hou W, Cheng X, et al. Joint Learning for Emotion Classification and Emotion Cause Detection[C]//Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018: 646-651.

Task & Contributions

The task of this paper is to detect emotions from textual and analysis the cause of emotions.
   The main contributions of this paper is listed as follows:

  1. This paper porposed a joint learning for emotion classification and emotion cause detection. This work is the first attempt to incorporate these two sub-tasks into a unfied framework.
  2. Experimenes conducted on Chineses microblogs show that the proposed method is promising.


The architecture of the proposed method is shown in the figure above. The frame contains two parts: a joint encoder and a linear decoder. The joint encoder part consisits of two sub-encoders: encoder EClass and encoder ECause. Each representation of sub-encoder is a concatenation of a main representation from this part of text and a auxiliary representation from another part of text. Overall, the proposed method used a multi-channel structure. However, two models are trained simultaneously and seperately, since they both have their own loss function.


  • The structure of the method is complicated and elaborately designed, but a complicated network may be hard to unstardand and too specific for a small task.
  • Emotion classification and emotion cause detection are relevant tasks in the field of NLP, and there are also many relevant tasks. If we can make full use of the relevance between them, we may achieve satisfactory results.
  • As for emotion classes only take a small percent of all the emotions, we can ignore them.
  • Dimension reduction can be used to prevent overwheling of the auxiliary information.