Neural Multimodal Belief Tracker with Adaptive Attention for Dialogue Systems

Multimodal dialogue systems are attracting increasing attention with a more natural and informative way for human-computer interaction. As one of its core components, the belief tracker estimates the user's goal at each step of the dialogue and provides a direct way to validate the ability of dialogue understanding. However, existing studies on belief trackers are largely limited to textual modality, which cannot be easily extended to capture the rich semantics in multimodal systems such as those with product images. For example, in fashion domain, the visual appearance of clothes play a crucial role in understanding the user's intention. In this case, the existing belief trackers may fail to generate accurate belief states for a multimodal dialogue system. In this paper, we present the first neural multimodal belief tracker (NMBT) to demonstrate how multimodal evidence can facilitate semantic understanding and dialogue state tracking. Given the multimodal inputs, while applying a textual encoder to represent textual utterances, the model gives special consideration to the semantics revealed in visual modality. It learns concept level fashion semantics by delving deep into image sub-regions and integrating concept probabilities via multiple instance learning. Then in each turn, an adaptive attention mechanism learns to automatically emphasize on different evidence sources of both visual and textual modalities for more accurate dialogue state prediction. We perform extensive evaluation on a multi-turn task-oriented dialogue dataset in fashion domain and the results show that our method achieves superior performance as compared to a wide range of baselines.