HiGRU: Hierarchical Gated Recurrent Units for Utterance-Level Emotion Recognition

Proceedings of NAACL-HLT 2019, pages 397–406 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 397 HiGRU: Hierarchical Gated Recurrent Units for Utterance-level Emotion Recognition Wenxiang Jiao1, Haiqin Yang2,3, Irwin King1, and Michael R. Lyu1 1 Department of Computer Science and Engineering, The Chinese University of Hong Kong, HKSAR, China 2 Meitu and 3 The Hang Seng University of Hong Kong, HKSAR, China {wxjiao,king,lyu}@cse.cuhk.edu.hk, hqyang@ieee.org Abstract In this paper, we address three challenges in utterance-level emotion recognition in di- alogue systems: (1) the same word can de- liver different emotions in different contexts; (2) some emotions are rarely seen in general dialogues; (3) long-range contextual informa- tion is hard to be effectively captured. We therefore propose a hierarchical Gated Recur- rent Unit (HiGRU) framework with a lower- level GRU to model the word-level inputs and an upper-level GRU to capture the contexts of utterance-level embeddings. Moreover, we promote the framework to two variants, Hi- GRU with individual features fusion (HiGRU- f) and HiGRU with self-attention and features fusion (HiGRU-sf), so that the word/utterance- level individual inputs and the long-range con- textual information can be sufficiently uti- lized. Experiments on three dialogue emo- tion datasets, IEMOCAP, Friends, and Emo- tionPush demonstrate that our proposed Hi- GRU models attain at least 8.7%, 7.5%, 6.0% improvement over the state-of-the-art meth- ods on each dataset, respectively. Particularly, by utilizing only the textual feature in IEMO- CAP, our HiGRU models gain at least 3.8% improvement over the state-of-the-art conver- sational memory network (CMN) with the tri- modal features of text, video, and audio. 1 Introduction Emotion recognition is a significant artificial intel- ligence research topic due to the promising poten- tial of developing empathetic machines for people. Emotion is a universal phenomena across differ- ent cultures and mainly consists of six basic types: anger, disgust, fear, happiness, sadness, and sur- prise (Ekman, 1971, 1992). In this paper, we focus on textual dialogue sys- tems because textual feature dominates the perfor- mance over audio and video features (Poria et al., Role Utterance Emotion Rachel Oh okay, I’ll fix that to. What’s her e- mail address? Neutral Ross Rachel! Anger Rachel All right, I promise. I’ll fix this. I swear. I’ll-I’ll- I’ll-I’ll talk to her. Non-neutral Ross Okay! Anger Rachel Okay. Neutral Nurse This room’s available. Neutral Rachel Okay! Joy Rachel Okay wait! Non-neutral Rachel You listen to me! Anger Figure 1: The word “okay” exhibits different emotions in the American television sitcom, Friends. 2015, 2017). In utterance-level emotion recogni- tion, an utterance (Olson, 1977) is a unit of speech bounded by breathes or pauses and its goal is to tag each utterance in a dialogue with the indicated emotion. In this task, we address three challenges: First, the same word can deliver different emotions in different contexts. For example, in Figure 1, the word “okay” can deliver three different emotions, anger, neutral, and joy, respectively. Strong emo- tions like joy and anger may be indicated by the symbols “!” or “?” along …

Content truncated to 500 words