Zero-shot Domain Adaptation with Inference Relation Paths for Spoken Language Understanding

Zero-shot slot filling methods are proposed to tackle the problems of adapting to new domains with unseen slots. Due to lacking share information across domains based on semantic slot descriptions, the challenge in zero-shot slot filling is handling the unseen slots that are less similar to the training slots semanti-cally. Since people utilizes not only explicit semantic information as estimation cues but also implicit semantic relations, this study attempts to find implicit sematic relation cues between slots and values to tackle unseen slots in zero-shot slot filling. We speculate that the inference paths between slots and values may be one of the implicit semantic relation cues and then investigated an amount of inference paths in the SnipsNLU dataset via the knowledge graph. The inference relation paths (IRPs) are found to implicitly build up semantic relations between slots and their values. Accordingly, we proposed a method to utilize the IRPs for zero-shot slot filling. Experimental results showed that the proposed method outperformed a strong baseline by 3.61 in terms of F1 score and achieved an improvement rate of 40 % for semantically dissimilar unseen slots. These results demonstrate that the proposed method can provide effective share information for handling unseen slots.