NLP on Spoken Documents Without ASR

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 460–470, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics NLP on Spoken Documents without ASR Mark Dredze, Aren Jansen, Glen Coppersmith, Ken Church Human Language Technology Center of Excellence Center for Language and Speech Processing Johns Hopkins University mdredze,aren,coppersmith,Kenneth.Church@jhu.edu Abstract There is considerable interest in interdis- ciplinary combinations of automatic speech recognition (ASR), machine learning, natu- ral language processing, text classification and information retrieval. Many of these boxes, especially ASR, are often based on consid- erable linguistic resources. We would like to be able to process spoken documents with few (if any) resources. Moreover, connect- ing black boxes in series tends to multiply er- rors, especially when the key terms are out-of- vocabulary (OOV). The proposed alternative applies text processing directly to the speech without a dependency on ASR. The method finds long (∼ 1 sec) repetitions in speech, and clusters them into pseudo-terms (roughly phrases). Document clustering and classi- fication work surprisingly well on pseudo- terms; performance on a Switchboard task ap- proaches a baseline using gold standard man- ual transcriptions. 1 Introduction Can we do IR-like tasks without ASR? Information retrieval (IR) typically makes use of simple features that count terms within/across documents such as term frequency (tf) and inverse document frequency (IDF). Crucially, to compute these features, it is suf- ficient to count repetitions of a term. In particular, for many IR-like tasks, there is no need for an au- tomatic speech recognition (ASR) system to label terms with phonemes and/or words. This paper builds on Jansen et al. (2010), a method for discovering terms with zero resources. This approach identifies long, faithfully repeated patterns in the acoustic signal. These acoustic repe- titions often correspond to terms useful for informa- tion retrieval tasks. Critically, this method does not require a phonetically interpretable acoustic model or knowledge of the target language. By analyzing a large untranscribed corpus of speech, this discovery procedure identifies a vast number of repeated regions that are subsequently grouped using a simple graph-based clustering method. We call the resulting groups pseudo-terms since they typically represent a single word or phrase spoken at multiple points throughout the corpus. Each pseudo-term takes the place of a word or phrase in bag of terms vector space model of a text document, allowing us to apply standard NLP algo- rithms. We show that despite the fully automated and noisy method by which the pseudo-terms are created, we can still successfully apply NLP algo- rithms with performance approaching that achieved with the gold standard manual transcription. Natural language processing tools can play a key role in understanding text document collections. Given a large collection of text, NLP tools can clas- sify documents by category (classification) and or- ganize documents into similar groups for a high level view of the collection (clustering). For exam- ple, given a collection of news articles, these tools can be applied so that the user can quickly see the …

Content truncated to 500 words