Affect-Driven Dialog Generation

Proceedings of NAACL-HLT 2019, pages 3734–3743 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3734 Affect-Driven Dialog Generation Pierre Colombo1∗, Wojciech Witon1∗, Ashutosh Modi1, James Kennedy1, Mubbasir Kapadia1,2 1Disney Research, 2Rutgers University {pierre.colombo, wojtek.witon}@disneyresearch.com ashutosh.modi@disneyresearch.com james.kennedy@disneyresearch.com mubbasir.kapadia@rutgers.edu Abstract The majority of current systems for end-to- end dialog generation focus on response qual- ity without an explicit control over the affec- tive content of the responses. In this paper, we present an affect-driven dialog system, which generates emotional responses in a controlled manner using a continuous representation of emotions. The system achieves this by mod- eling emotions at a word and sequence level using: (1) a vector representation of the de- sired emotion, (2) an affect regularizer, which penalizes neutral words, and (3) an affect sam- pling method, which forces the neural network to generate diverse words that are emotion- ally relevant. During inference, we use a re- ranking procedure that aims to extract the most emotionally relevant responses using a human- in-the-loop optimization process. We study the performance of our system in terms of both quantitative (BLEU score and response diver- sity), and qualitative (emotional appropriate- ness) measures. 1 Introduction Recent breakthroughs in deep learning techniques have had an impact on end-to-end conversational systems (Chen et al., 2017). Current research is mainly focused on functional aspects of conver- sational systems: keyword extraction, natural lan- guage understanding, and pertinence of generated responses (Ilievski et al., 2018). Although these aspects are indeed key features for building a com- mercial system, most existing solutions lack social intelligence. Conversational systems could bene- fit from incorporating social intelligence by: (1) avoiding interaction problems that may arise when the system does not understand the user’s request (e.g., inappropriate responses that cause user anger) (Maslowski et al., 2017), and (2) building rapport ∗ Both authors contributed equally to this work. with the user (Strohkorb et al., 2016). Our method makes such conversational systems more social by outputting responses expressing emotion in a controlled manner, without sacrificing grammatical correctness, coherence, or relevance. Existing sequence-to-sequence (seq2seq) archi- tectures, either recurrent- (Sordoni et al., 2015; Ser- ban et al., 2015), attention- (Vaswani et al., 2017) or convolutional neural network (CNN)-based (Fan et al., 2018), do not provide a straightforward way to generate emotionally relevant output in a con- trolled manner. We introduce EMOTIonal CONver- sational System (EMOTICONS), which generates emotion-specific responses. It is based on novel contributions presented in this paper which fall in two main categories: explicit models which allow a controlled emotion-based response generation (e.g., methods based on emotion embeddings, af- fective sampling, and affective re-ranking), and implicit models with no direct control over the de- sired emotion (i.e., affective regularizer). We show that EMOTICONS outperforms both the system proposed by Zhou et al. (2018) (current state of the art for our task) and the vanilla seq2seq in terms of BLEU score (Papineni et al., 2002) (improve- ment up to 7.7%) and response diversity (improve- ment up to 52%). Additionally, we qualitatively evaluate …

Content truncated to 500 words