CoAStaL at SemEval-2019 Task 3: Affect Classification in Dialogue using Attentive BiLSTMs

CoAStaL at SemEval-2019 Task 3: Affect Classification in Dialogue using Attentive BiLSTMs Ana Valeria Gonz´alez∗and Victor Petr´en Bach Hansen∗and Joachim Bingel and Anders Søgaard University of Copenhagen, Dept. of Computer Science Copenhagen, Denmark ana|victor.petren|bingel|soegaard@di.ku.dk Abstract This work describes the system presented by the CoAStaL Natural Language Processing group at University of Copenhagen. The main system we present uses the same attention mechanism presented in (Yang et al., 2016). Our overall model architecture is also inspired by their hierarchical classification model and adapted to deal with classification in dialogue by encoding information at the turn level. We use different encodings for each turn to create a more expressive representation of dialogue context which is then fed into our classifier. We also define a custom preprocessing step in order to deal with language commonly used in interactions across many social media out- lets. Our proposed system achieves a micro F1 score of 0.7340 on the test set and shows significant gains in performance compared to a system using dialogue level encoding. 1 Introduction Recognizing emotion is crucial to human-human communication and has for a long time been a goal in human-machine interaction. Although there has been growing interest in emotion detec- tion across many fields (Liscombe et al., 2005; Agrafioti et al., 2012; Craggs and Wood, 2004), much of the work has focused on developing em- pathetic systems using multimodal approaches i.e. speech and gestures as well as text (Hazarika et al., 2018). Approaching emotion detection as a multi- modal problem certainly makes sense, as face-face human communication involves many modalities, however, this fails to consider all the communica- tion that is increasingly happening solely via chat, or written means. Detecting emotion in textual di- alogue without the other modalities, such as work done by Gupta et al., can allow us to improve a number of applications dealing with social media ∗Authors contributed equally interactions, opinion mining, and customer inter- actions, unfortunately, this is a great challenge that has remained largely unexplored. SemEval 2019 Task 3 attempts to encourage research in this di- rection. Given a user utterance and the previous two turns of context, the task consists in classi- fying the user utterance according to one of four emotion classes: happy, sad, angry or other. For a full description of the task see (Chatterjee et al., 2019). In this paper, we describe our turn-level at- tention model used to tackle this task, specifically using the attention mechanism presented in (Yang et al., 2016). Our model encodes turns in a conver- sation separately using an Attentive Bidirectional LSTM encoder. In the model presented in the shared task, the turn encoders do not share param- eters, achieving a micro F1 score of 0.7340. The code for all experiments presented here is avail- able. 1 2 Related Work Due to its many potential applications across many fields, detection of speaker emotional state in spo- ken dialogue systems has been studied extensively. Early studies showed that the use of prosody as well as speaking style …

Content truncated to 500 words