[1N3-J-9] Natural language processing, information retrieval: understanding

[1N3-J-9-03] Spoken Language Understanding based on Sentence Segmentation by Language Models

〇Kei Wakabayashi¹, Johane Takeuchi², Makoto Hiramatsu¹, Mikio Nakano² (1. University of Tsukuba, 2. Honda Research Institute Japan Co., Ltd.)

Keywords:Slot filling, Sequence labeling, Language model, Dirichlet process

In this paper, we propose a new approach to solve the slot filling task for spoken language understanding by using a formulation based on the optimum segmentation of an input sentence. This formulation enables us to develop a language modeling-based method that is drastically efficient compared to the existing deep learning approach that formalizes the slot filling as a sequence labeling task. The proposed method trains the language models by a one-pass algorithm and applies a dynamic programming algorithm to find the most likely slot assignment efficiently. We empirically confirmed that the proposed method achieves a competitive accuracy compared to a deep learning method, and even works with drastically less computing resource consumption.