Detecting OOV Named-Entities in Conversational Speech

A common cause of errors in spoken language systems is the presence of out-of-vocabulary (OOV) words in the input. Named entities (people, places, organizations, etc.) are a particularly important class of OOVs. In this paper we focus on detecting OOV named entities (NEs) for two-way English/Iraqi speech-tospeech translation. Our approach builds on Maximum Entropy (MaxEnt) classifier trained on a suite of contextual features. These features include: n-gram context, part-of-speech tags (both supervised and unsupervised), and word posterior features computed from the trajectory of the word posteriors within the utterance. Our experimental results show that fusion (both early and late) of these novel word posterior features with rest of the contextual features significantly improves detection accuracy for OOV NEs. However, we also observe that the same features that perform well on OOV NEs can hurt in detecting in-vocabulary NEs. Therefore, the choice of the features should be based on expected occurrence of OOV NEs.