EmotionLines: An Emotion Corpus of Multi-Party Conversations

EmotionLines: An Emotion Corpus of Multi-Party Conversations Sheng-Yeh Chen1∗, Chao-Chun Hsu1∗, Chuan-Chun Kuo1, Ting-Hao (Kenneth) Huang2, Lun-Wei Ku1 Academia Sinica, Taiwan Carnegie Mellon University, USA 1{alan012102, joe32140, william0617, lwku}@iis.sinica.edu.tw 2tinghaoh@cs.cmu.edu Abstract Feeling emotion is a critical characteristic to distinguish people from machines. Among all the multi-modal resources for emotion detection, textual datasets are those containing the least additional information in addition to semantics, and hence are adopted widely for testing the developed systems. However, most of the textual emotional datasets consist of emotion labels of only individual words, sentences or documents, which makes it challenging to discuss the contextual flow of emotions. In this paper, we introduce EmotionLines, the first dataset with emotions labeling on all utterances in each dialogue only based on their textual content. Dialogues in EmotionLines are collected from Friends TV scripts and private Facebook messenger dialogues. Then one of seven emotions, six Ekman’s basic emotions plus the neutral emotion, is labeled on each utterance by 5 Amazon MTurkers. A total of 29,245 utterances from 2,000 dialogues are labeled in EmotionLines. We also provide several strong baselines for emotion detection models on EmotionLines in this paper. Keywords: emotion detection, emotional dialogue dataset 1. Introduction There are two major kinds of dialogue systems: a task- oriented dialogue system and the social (chit-chat) dialogue system. The former focuses on designing a personal assis- tant which can accomplish certain tasks, and for the lat- ter it is important to capture the conversation flow which emphasizes more on the feelings of the speaker. Many re- searchers try to build a “smart” dialogue system by enhanc- ing dialogue breadth (coverage), dialogue depth (complex- ity) or both. Those who want to increase dialogue breadth try to transfer dialogue acts across domains (Chen et al., 2016) to establish multi-domain or even open domain dia- logue system, and those who want to deepen dialogue com- plexity pay their attention to transform a knowledge-based systems to common sense or even empathetic systems that can recognize emotion features, generate emotion-aware responses (Fung et al., 2016), or learn how to plan the dia- logues while users interact via high-level descriptions (Sun et al., 2016). No matter what kind of dialogue system we want to build, a useful and large dialogue dataset is indis- pensable. When building a task-oriented dialogue system, dialogue corpora with dialogue act information is accessible and hence are commonly utilized. However when building a chit-chat conversational bot, though the importance of emo- tion detection has been noticed, pure conversation con- tent such as movie, TV scripts or chat logs without emo- tion labels are more available: no emotion labels on ut- terances can be used for learning. Moreover, when we turn to other datasets with annotated emotion information such as data crawled from Twitter (Mohammad and Bravo- Marquez, 2017), the labeled units (posts or sentences) are independent. As a result, models built with these datasets lack the ability to consider contextual information essential ∗These authors contributed equally to this work. in dialogue systems, not to …

Content truncated to 500 words