Simulating Early-Termination Search for Verbose Spoken Queries

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1270–1280, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Simulating Early-Termination Search for Verbose Spoken Queries Jerome White IBM Research Bangalore, KA India jerome.white@in.ibm.com Douglas W. Oard University of Maryland College Park, MD USA oard@umd.edu Nitendra Rajput IBM Research New Delhi, India rnitendra@in.ibm.com Marion Zalk University of Melbourne Melbourne, VIC Australia m.zalk@student.unimelb.edu.au Abstract Building search engines that can respond to spoken queries with spoken content requires that the system not just be able to find useful responses, but also that it know when it has heard enough about what the user wants to be able to do so. This paper describes a simula- tion study with queries spoken by non-native speakers that suggests that indicates that find- ing relevant content is often possible within a half minute, and that combining features based on automatically recognized words with features designed for automated prediction of query difficulty can serve as a useful basis for predicting when that useful content has been found. 1 Introduction Much of the early work on what has come to be called “speech retrieval” has focused on the use of text queries to rank segments that are automatically extracted from spoken content. While such an ap- proach can be useful in a desktop environment, half of the world’s Internet users can access the global information network only using a voice-only mobile phone. This raises two challenges: 1) in such set- tings, both the query and the content must be spo- ken, and 2) the language being spoken will often be one for which we lack accurate speech recognition. The Web has taught us that the “ten blue links” paradigm can be a useful response to short queries. That works because typed queries are often fairly precise, and tabular responses are easily skimmed. However, spoken queries, and in particular open- domain spoken queries for unrestricted spoken con- tent, pose new challenges that call for new thinking about interaction design. This paper explores the po- tential of a recently proposed alternative, in which the spoken queries are long, and only one response can be played at a time by the system. This ap- proach, which has been called Query by Babbling, requires that the user ramble on about what they are looking for, that the system be able to estimate when it has found a good response, and that the user be able to continue the search interaction by bab- bling on if the first response does not fully meet their needs (Oard, 2012). One might question whether users actually will “babble” for extended periods about their informa- tion need. There are two reasons to believe that some users might. First, we are particularly inter- ested in ultimately serving users who search for in- formation in languages for which we do not have us- able speech recognition systems. Speech-to-speech matching in such cases will be challenging, and we would not expect short queries to …

Content truncated to 500 words