Multi-style adaptive training for robust cross-lingual spoken language understanding

摘要: Given the increasingly available machine translation (MT) services nowadays, one efficient strategy for cross-lingual spoken language understanding (SLU) is to first translate input utterance from second into primary language, and then call SLU system decode semantic knowledge. However, errors introduced in MT process create a condition similar “mismatch” encountered robust speech recognition. Such mismatch makes performance of far acceptable. Motivated by successful solutions developed recognition, we this paper propose multi-style adaptive training method improve robustness tasks. For evaluation, created an English-Chinese bilingual ATIS database, carried out series experiments on that database experimentally assess proposed methods. Experimental results show that, without relying any data significantly improves task while producing no degradation language. This greatly facilitates porting as many languages there are systems human effort. We further study approach another type condition, caused recognition errors, demonstrate its success also.