Learning Diphone‐Based Segmentation

Abstract

This paper reconsiders the diphone-based word segmentation model of Cairns, Shillcock, Chater, and Levy (1997) and Hockema (2006), previously thought to be unlearnable. A statistically principled learning model is developed using Bayes' theorem and reasonable assumptions about infants' implicit knowledge. The ability to recover phrase-medial word boundaries is tested using phonetic corpora derived from spontaneous interactions with children and adults. The (unsupervised and semisupervised) learning models are shown to exhibit several crucial properties. First, only a small amount of language exposure is required to achieve the model's ceiling performance, equivalent to between 1 day and 1 month of caregiver input. Second, the models are robust to variation, both in the free parameter and the input representation. Finally, both the learning and baseline models exhibit undersegmentation, argued to have significant ramifications for speech processing as a whole.

Diphones-fr: A French database of diphone positional frequency

“…Moreover, such "diphone-based segmentation" has proved valid and learnable in a computational Bayesian model recently developed by Daland and Pierrehumbert (2011). They assessed whether the model could recover word boundaries based on the identity of the surrounding diphone.…”

Comparing Models of Phonotactics for Word Segmentation

“…It complements that of recent work investigating the use of phoneme-level statistical regularities for segmentation (Adriaans & Kager, 2010;Daland & Pierrehumbert, 2011). Our work differs from these latter approaches, however, in comparing several phonotactic models, including ones relying on the syllable-based transitional probability statistics investigated in infant research.…”

“…However, when the same cues are used in the context of a simple, generative probability model with improved unsupervised parameter estimation, the syllablebased models substantially outperform the phoneme-based models. Indeed, the syllable-based transitional probability phonotactic model achieves a word token segmentation f-score of nearly 80%, which is the highest reported performance among purely phonotactically-based segmentation models (Adriaans & Kager, 2010;Daland & Pierrehumbert, 2011). Indeed, this performance compares favorably with state-of-theart segmentation models that involve learning of higher level regularities, such as the lexicon and collocations (Brent, 1999;Venkataraman, 2001;Johnson, 2008a;Goldwater et al, 2009;, and demonstrates that good segmentation performance can be achieved by exploiting simple syllable-level phonotactic cues.…”

“…A number of phonotactically-based segmentation models, focusing on the discovery of word boundaries based on phoneme-level statistics, have achieved more promising results (Adriaans & Kager, 2010;Daland & Pierrehumbert, 2011;see also Brent, 1999). For example, Brent (1999) showed that a local minimum strategy relying on phoneme bigrams correctly extracts about 50% of word tokens in English child-directed speech.…”

“…This work has shown that infants utilize these low-level statistical regularities to segment speech during the second half of the first year of life before they have developed extensive vocabularies that could provide top-down lexical information to guide segmentation. Developmental research indicates that on average infants know fewer than 100 word types during this period (Dale & Fenson, 1996;Daland & Pierrehumbert, 2011).…”