Towards Sentiment-Aware Multi-Modal Dialogue Policy Learning

Creation of task-oriented dialog/virtual agent (VA) capable of managing complex domain-specific user queries pertaining to multiple intents is difficult since the agent must deal with several subtasks simultaneously. Most end-to-end dialogue systems, however, only provide user semantics as inputs from texts into the learning process and neglect other useful user behaviour and information from other modalities such as images. This stresses the benefit of incorporating multi-modal inputs for eliciting user preference in the task. Also, sentiment of the user plays a significant role in achieving maximum user/customer satisfaction during the conversation. Thus, it is also important to incorporate users' sentiments during policy learning, especially when serving user's composite goals. For the creation of multi-modal VA aided with sentiment for conversations encompassing multi-intents, this paper introduces a new dataset, named Vis-SentiVA: Visual and Sentiment aided VA created from open-accessed conversational dataset. We present a hierarchical reinforcement learning (HRL) typically options-based VA to learn policies for serving multi-intent dialogues. Multi-modal information (texts and images) extraction to identify user's preference is incorporated in the learning framework. A combination of task-based and sentiment-based rewards is integrated in the hierarchical value functions for the VA to be user adaptive. Empirically, we show that all these aspects induced together in the learning framework play a vital role in acquiring higher dialogue task success and increased user contentment in the process of creating composite-natured VAs. This is the first effort in integrating sentiment-aware rewards in the multi-modal HRL framework. The paper highlights that it is indeed essential to include other modes of information extraction such as images and behavioural cues of the user such as sentiment to secure greater user contentment. This also helps in improving success of composite-natured VAs serving task-oriented dialogues.