EMTC: Multilabel Corpus in Movie Domain for Emotion Analysis in Conversational Text

EMTC: Multilabel Corpus in Movie Domain for Emotion Analysis in Conversational Text Phan Duc-Anh and Yuji Matsumoto Computational Linguistics Laboratory, Graduate school of Information Science Nara Institute of Science and Technology, Japan {phan.duc anh.oq3, matsu}@is.naist.jp Abstract It is proved that in text-based communication such as sms, messengers applications, misinterpretation of partner’s emotions are pretty common. In order to tackle this problem, we propose a new multilabel corpus named Emotional Movie Transcript Corpus (EMTC). Unlike most of the existing emotion corpora that are collected from Twitters and use hashtags labels, our corpus includes conversations from movie with more than 2.1 millions utterances which are partly annotated by ourselves and independent annotators. To our intuition, conversations from movies are closer to real-life settings and emotionally richer. We believe that a corpus like EMTC will greatly benefit the development and evaluation of emotion analysis systems and improve their ability to express and interpret emotions in text-based communication. Keywords: emotion corpus, movie transcript, text-based communication, emotion analysis, multilabel 1. Introduction In the recent years, we experience rapid development of online communication with the help of mediated devices such as smart phones, tablets, computer. People use talk- over-internet, video conferencing features for both daily and business tasks. Text based methods like emails or text messengers are still very convenient and indispensable to us because of its unique advantages: they do not require inter- mediate responses and can be used for the sake of record- keeping. However, it is already proved that in any online communication methods, users experience more difficul- ties in interpreting and conveying emotions than face-to- face communication due to the limitation in communica- tion modality. Furthermore, text-based methods are where difficulties are encountered the most (Kruger et al., 2005; Arimoto and Okanoya, 2016). Therefore, we targeted the effort to build an emotion analysis system focusing on text data. The starting point is to develop an emotional corpus that has conversational texts and is as close to real-life com- munication as possible. Existing emotional text corpora are often collected from micro-blog platforms using multiclass scheme - one emo- tion per example (Liew et al., 2016). Most of them are au- tomatically annotated by extracting hashtags rather than by human judgements (Dini and Bittar, 2016; Li et al., 2016). While the text data from micro-blog platforms like Twitters are very convenient and easy to collect, the fact that they are limited in the number of characters (140 for a tweet) differs themselves from daily conversation text and there- fore, have limited use in a real-life settings. On the other hand, multiclass scheme has its own limitation. One input is only associated to one emotion. However, it is observed in some research (Liew et al., 2016) that multilabel scheme with no limitation in the number of emotions per example is a better and more natural way of annotating emotion labels. We describe our efforts to construct and annotate partly the Emotional Movie Transcript Corpus (EMTC). Most of the corpus are unsupervised data. We annotated by …

Content truncated to 500 words