A k-Nearest Neighbor Approach towards Multi-level Sequence Labeling

摘要

In this paper we present a new method for spoken language understanding to support a spoken dialogue system handling complex dialogues in the food ordering domain. Using a small amount of authentic food ordering dialogues yields better results than a large amount of synthetic ones. The size of the data makes this approach amenable to cold start projects in the multi-level sequence labeling domain. We used windowed word n-grams, POS tag sequences and pre-trained word embeddings as features. Results show that a heterogeneous feature set with the k-NN learner performs competitively against the state-of-the-art results and achieve an F-score of 60.71.