Phrase-Based Query Degradation Modeling for Vocabulary-Independent Ranked Utterance Retrieval

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 182–190, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Phrase-Based Query Degradation Modeling for Vocabulary-Independent Ranked Utterance Retrieval J. Scott Olsson HLT Center of Excellence Johns Hopkins University Baltimore, MD 21211, USA solsson@jhu.edu Douglas W. Oard College of Information Studies University of Maryland College Park, MD 15213, USA oard@umd.edu Abstract This paper introduces a new approach to rank- ing speech utterances by a system’s confi- dence that they contain a spoken word. Multi- ple alternate pronunciations, or degradations, of a query word’s phoneme sequence are hy- pothesized and incorporated into the ranking function. We consider two methods for hy- pothesizing these degradations, the best of which is constructed using factored phrase- based statistical machine translation. We show that this approach is able to significantly im- prove upon a state-of-the-art baseline tech- nique in an evaluation on held-out speech. We evaluate our systems using three differ- ent methods for indexing the speech utter- ances (using phoneme, phoneme multigram, and word recognition), and find that degrada- tion modeling shows particular promise for lo- cating out-of-vocabulary words when the un- derlying indexing system is constructed with standard word-based speech recognition. 1 Introduction Our goal is to find short speech utterances which contain a query word. We accomplish this goal by ranking the set of utterances by our confidence that they contain the query word, a task known as Ranked Utterance Retrieval (RUR). In particular, we are interested in the case when the user’s query word can not be anticipated by a Large Vocabulary Continuous Speech Recognizer’s (LVCSR) decod- ing dictionary, so that the word is said to be Out-Of- Vocabulary (OOV). Rare words tend to be the most informative, but are also most likely to be OOV. When words are OOV, we must use vocabulary-independent tech- niques to locate them. One popular approach is to search for the words in output from a phoneme rec- ognizer (Ng and Zue, 2000), although this suffers from the low accuracy typical of phoneme recogni- tion. We consider two methods for handling this in- accuracy. First, we compare an RUR indexing sys- tem using phonemes with two systems using longer recognition units: words or phoneme multigrams. Second, we consider several methods for handling the recognition inaccuracy in the utterance rank- ing function itself. Our baseline generative model handles errorful recognition by estimating term fre- quencies from smoothed language models trained on phoneme lattices. Our new approach, which we call query degradation, hypothesizes many alternate “pronunciations” for the query word and incorpo- rates them into the ranking function. These degra- dations are translations of the lexical phoneme se- quence into the errorful recognition language, which we hypothesize using a factored phrase-based statis- tical machine translation system. Our speech collection is a set of oral history interviews from the MALACH collection (Byrne et al., 2004), which has previously been used for ad hoc speech retrieval evaluations using one-best word level transcripts (Pecina …

Content truncated to 500 words