Robust character based tagging with domain lexical features for Chinese spoken language understanding

Word information is useful in natural language understanding. But in Chinese language processing, word information is not given natural. While word-segmentation works well for text in NLU, it deteriorates Chinese SLU because of the flexibility and distortion of spoken utterance plus ASR errors. This paper propose a novel approach, sub-word features, to take use word information and help understanding spoken utterance while retain the robustness of character-wise processing. By means of this approach, we can also effectively use named entity list to improve SLU performance. Experiments show that the sub-word features give an average of 0.7 improvement for ASR, and the usage of named list given an average of 4.7 improvement.