Prosodic boundary information helps unsupervised word segmentation

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 953–963, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Prosodic boundary information helps unsupervised word segmentation Bogdan Ludusan, Gabriel Synnaeve and Emmanuel Dupoux Laboratoire de Sciences Cognitives et Psycholinguistique EHESS / ENS / CNRS 29 rue d’Ulm, 75005 Paris, France bogdan.ludusan@ens.fr, gabriel.synnaeve@gmail.com, emmanuel.dupoux@gmail.com Abstract It is well known that prosodic information is used by infants in early language acquisition. In particular, prosodic boundaries have been shown to help infants with sentence and word- level segmentation. In this study, we extend an unsupervised method for word segmen- tation to include information about prosodic boundaries. The boundary information used was either derived from oracle data (hand- annotated), or extracted automatically with a system that employs only acoustic cues for boundary detection. The approach was tested on two different languages, English and Japanese, and the results show that boundary information helps word segmentation in both cases. The performance gain obtained for two typologically distinct languages shows the ro- bustness of prosodic information for word seg- mentation. Furthermore, the improvements are not limited to the use of oracle informa- tion, similar performances being obtained also with automatically extracted boundaries. 1 Introduction Prosodic information is thought to play a fundamen- tal role in early language acquisition, and provide infants with rich structural information about their language (Christophe et al., 1997). In particular, prosody has been claimed to help infants find word boundaries (Christophe and Dupoux, 1996). New- borns are able discriminate between disyllables that contains vs. does not contain a phonological phrase boundary (Christophe et al., 1994; Christophe et al., 2001), showing that they are able to encode the cor- responding prosodic cues. Nine-month olds show evidence of parsing utterances into prosodic units, and show ’surprise’ when a pause is inappropriately inserted inside as opposed to between these units (Jusczyk et al., 1992; Gerken et al., 1994). Ten to 13 month olds show evidence of using prosodic units to parse utterances into words, as they fail to recognize a familiar word if it appears to straddle a prosodic boundary (Gout et al., 2004). Curiously enough, however, prosody is not used very much in unsupervised models of language ac- quisition, and in particular, in models of word seg- mentation. Most such models use text as input, and apply some form of lexical optimization. For in- stance, Brent and Cartwright (1996) used a Min- imal Description Length Principle to optimize the size of the description of a corpus. State of the art systems use hierarchical Bayesian models (Goldwa- ter et al., 2009) which parse a corpus into words or other linguistic units with a bias to reuse pre- viously parsed elements. Adaptor Grammars is a generic framework which enables to formulate such Bayesian models within an overarching architecture based on probabilistic context free grammars (John- son et al., 2007). Such models have been used to study the role of linguistic information such as syl- labic structure …

Content truncated to 500 words