出版社：The Japanese Society for Artificial Intelligence
摘要：In language understanding for dialog systems, slot filling is a fundamental task usually formulated as a sequence labeling problem and solved using discriminative models such as conditional random fields and neural networks. One of the weak points of the discriminative approach is in the robustness against incomplete annotations, which are often generated in practice when we attempt to build large-scale training data. For making the slot filling algorithm more robust against the incompleteness of annotation, this paper leverages an overlooked property of slot filling tasks: Non-slot parts of utterance follow a specific pattern depending on the user’s intent. To reflect this idea, we propose a nonparametric Bayesian model that induces the grammatical role of the non-slot parts using a segmentation-based formulation of slot filling tasks. The proposed method can naturally deal with training data that includes incomplete annotations as a partially supervised grammar induction problem. The experimental result demonstrates that the proposed method estimates the slot information more accurately in a situation that the training data includes incomplete annotations in comparison to the BiLSTM-CRF and HMM.We also show that the proposed model has an advantage in the interpretability of the result of training and prediction by visualizing the parameters and the estimated labeled segmentations with a state transition diagram.
关键词：dialog system;slot filling;incomplete annotation;generative model;nonparametric Bayesian model
