Affect-Driven Dialog Generation

Abstract

The majority of current systems for end-toend dialog generation focus on response quality without an explicit control over the affective 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 modeling emotions at a word and sequence level using: (1) a vector representation of the desired emotion, (2) an affect regularizer, which penalizes neutral words, and (3) an affect sampling method, which forces the neural network to generate diverse words that are emotionally relevant. During inference, we use a reranking procedure that aims to extract the most emotionally relevant responses using a humanin-the-loop optimization process. We study the performance of our system in terms of both quantitative (BLEU score and response diversity), and qualitative (emotional appropriateness) measures.

The importance of fillers for text representations of speech transcripts

“…The gradient norm is clipped to 5.0, weight decay is set to 10 −6 , and dropout (LeCun et al, 2015) is set to 0.2. Models have been implemented in PyTorch and trained on a v100 using the same procedure as in (Colombo et al, 2019(Colombo et al, , 2020Witon et al, 2018).…”

Frustration Intensity Prediction in Customer Support Dialog Texts

“…Automated goal-oriented dialog agents have been studied in quite a few works, for example Ham et al, 2020 [18], as have affect-driven free-chat dialog agents e.g. Colombo et al, 2019 [19], and Lubis et al, 2018 [20], focusing on providing affect-sensitive responses, but very few works have investigated dialog agents that attempt to address both concerns simultaneously [21], [13].…”

CDL: Curriculum Dual Learning for Emotion-Controllable Response Generation

“…Recently, a framework called emotional chatting machine (ECM) (Zhou et al, 2018a) was proposed to address the emotion factor in a controlled manner, which focuses on generating a response with a specific emotion (Example 1 in Table 1). In the research field of emotion-controllable response generation, ECM and its successive methods (Colombo et al, 2019; mainly represent the given emotion category as a vector and add it to the decoding steps to influence the procedure of response generation, which would aggravate the safe response problem. For the response generation task, safe response is notorious, as the model tends to produce some generic but meaningless responses, like "Thank you", "I don't know", "Yes", etc.…”