Re-ranking ASR Outputs for Spoken Sentence Retrieval.
Yeongkil Song, Hyeokju Ahn, Harksoo Kim · 2014
Abstract. In spoken information retrieval, users ’ spoken queries are converted into text queries by using ASR engines. If top-1 results of the ASR engines are incorrect, the errors are propagated to information retrieval systems. If a docu-ment collection is a small set of short texts, the errors will more affect the per-formances of information retrieval systems. To improve the top-1 accuracies of the ASR engines, we propose a post-processing model to rearrange top-n out-puts of ASR engines by using Ranking SVM. To improve the re-ranking per-formances, the proposed model uses various features such as ASR ranking in-formation, morphological information, and domain-specific lexical information. In the experiments, the proposed model showed the higher precision of 4.4% and the higher recall rate of 6.4 % than the baseline model without any post-processing. Based on this experimental result, the proposed model showed that it can be used as a post-processor for improving the performance of a spoken information retrieval system if a document collection is a restricted amount of sentences.