15th Annual Conference of the International Speech Communication Association

September 14-18, 2014

Robust Retrieval Models for False Positive Errors in Spoken Documents

Sho Kawasaki, Tomoyosi Akiba

Toyohashi University of Technology, Japan

How to deal with speech recognition errors and out-of-vocabulary (OOV) words, which are referred to as false negative errors, are common challenges in spoken document processing. To deal with them in spoken content retrieval (SCR), the SCR method that incorporated spoken term detection (STD) as the pre-process stage (referred to as STD-SCR) has been proposed. However, the STD-SCR tends to increase false positive errors in compensation for reducing false negative errors. In this work, we propose robust retrieval models for false positive errors by using word co-occurrences. The words that co-occur in a given query are semantically related, so that they are likely to co-occur also in the document to be retrieved. On the other hand, if a word in a given query appears alone in a document, it is more like a false positive. We incorporate this idea into two retrieval models commonly used in the literature, i.e. the vector space model and the query likelihood model. Our experimental result showed our proposed extensions on the retrieval models successfully improved the retrieval performance not only for the STD-SCR but also for the conventional SCR method.

Full Paper

Bibliographic reference.  Kawasaki, Sho / Akiba, Tomoyosi (2014): "Robust retrieval models for false positive errors in spoken documents", In INTERSPEECH-2014, 1757-1761.