Cue Me In: Content-Inducing Approaches to Interactive Story Generation

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 37–43 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 37 FERNet: Fine-grained Extraction and Reasoning Network for Emotion Recognition in Dialogues Yingmei Guo, Zhiyong Wu, Mingxing Xu Department of Computer Science and Technology Beijing National Research Center for Information Science and Technology Tsinghua University, Beijing, China guoym18@mails.tsinghua.edu.cn, zywu@sz.tsinghua.edu.cn xumx@tsinghua.edu.cn Abstract Unlike non-conversation scenes, emotion recognition in dialogues (ERD) poses more complicated challenges due to its interactive nature and intricate contextual information. All present methods model historical utter- ances without considering the content of the target utterance. However, different parts of a historical utterance may contribute differently to emotion inference of different target utter- ances. Therefore we propose Fine-grained Extraction and Reasoning Network (FERNet) to generate target-specific historical utterance representations. The reasoning module effec- tively handles both local and global sequential dependencies to reason over context, and up- dates target utterance representations to more informed vectors. Experiments on two bench- marks show that our method achieves com- petitive performance compared with previous methods. 1 Introduction With the development of human-machine interac- tion (HMI) applications, textual dialogue scenes appear more frequently. These scenes request ef- fective and high-performance emotion recognition systems helping in building empathetic machines (Young et al., 2018). Therefore, emotion recogni- tion in dialogues (ERD) is getting growing atten- tion from both academic and business community. Different from non-conversation scenes, the ERD task poses a more complicated challenge of modeling context-sensitive dependencies. Most of existing approaches adopt Convolution Neural Net- work (CNN) (Krizhevsky et al., 2012), followed by a max-pooling layer to obtain utterance rep- resentations (Kim, 2014; Torres, 2018; Hazarika et al., 2018a,b; Majumder et al., 2019; Ghosal et al., 2019). The process proceeds without the guidance of the target utterance, thus generated historical utterance representations are indistinguishable to- ward different target utterances. Emotion recogni- tion may fail in cases where historical utterances express various emotions toward various targets, which may confuse the emotion recognition of tar- get utterances. As Figure 1 shows, for different tar- get utterances B1 and B2, the model should attend the words “good service” and “bad food” in A1, separately. In a word, it is desired to pay different attention to different words of a certain historical utterance to generate the target-specific historical utterance representation. The restaurant where we had dinner provided [good service] but [bad food]. 😐 What about the food? I can’t agree with you more. 😐 😞 😄 Yeah, I was impressed by the service. 𝐴" 𝐴# 𝐵" 𝐵# Figure 1: A dialogue shows that modeling intricate con- textual information is crucial for emotion recognition. In this paper, we propose Fine-grained Extrac- tion and Reasoning Network (FERNet) to gener- ate target-specific historical utterance representa- tions conditioned on the content of target utter- ances by using the multi-head attention mechanism (Vaswani et al., 2017), extracting more fine-grained, relevant and contributing information for emotion recognition. …

Content truncated to 500 words