Learning to Rank for Arabic Transcriptions Retrieval

Abstract

The amount of spoken documents being shared on the web per minute is increasing dramatically posing a true challenge for any search engine in order to satisfy its customers’ queries. With the ingoing improvement in the speech recognizers accuracy, this research addresses the problem of ranking transcriptions that can be obtained by speech recognizers to enhance search engine ranking results. Depending on the title of the video and some of its meta-data only is not sufficient for some queries that have the information need to get relevant spoken segments within audio files. Feature extraction based on both the meta-data of the spoken documents and the timed spoken content transcription for an Arabic audio dataset for Quran is proposed. The results revealed that applying learning to rank techniques are superior to the baseline unsupervised BM25 scoring. In addition, using transcription-based features proved its effectiveness in terms of both the Normalized Discounted Cumulative Gain (NDCG@10) and Expected Reciprocal Rank (ERR@10).