Recession Segmentation: Simpler Online Word Segmentation Using Limited Resources

Proceedings of the Fourteenth Conference on Computational Natural Language Learning, pages 88–97, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Recession Segmentation: Simpler Online Word Segmentation Using Limited Resources∗ Constantine Lignos, Charles Yang Dept. of Computer and Information Science, Dept. of Linguistics University of Pennsylvania lignos@cis.upenn.edu, charles.yang@ling.upenn.edu Abstract In this paper we present a cognitively plau- sible approach to word segmentation that segments in an online fashion using only local information and a lexicon of pre- viously segmented words. Unlike popu- lar statistical optimization techniques, the learner uses structural information of the input syllables rather than distributional cues to segment words. We develop a memory model for the learner that like a child learner does not recall previously hy- pothesized words perfectly. The learner at- tains an F-score of 86.69% in ideal condi- tions and 85.05% when word recall is un- reliable and stress in the input is reduced. These results demonstrate the power that a simple learner can have when paired with appropriate structural constraints on its hy- potheses. 1 Introduction The problem of word segmentation presents an important challenge in language acquisition. The child learner must segment a continuous stream of sounds into words without knowing what the in- dividual words are until the stream has been seg- mented. Computational models present an op- portunity to test the potentially innate constraints, structures, and algorithms that a child may be us- ing to guide her acquisition. In this work we de- velop a segmentation model from the constraints suggested by Yang (2004) and evaluate it in ideal- ized conditions and conditions that better approx- imate the environment of a child learner. We seek to determine how these limitations in the learner’s input and memory affect the learner’s performance and to demonstrate that the presented learner is ro- bust even under non-ideal conditions. ∗Portions of this work were adapted from an earlier manuscript, Word Segmentation: Quick But Not Dirty. 2 Related Work Most recent work in word segmentation of child- directed speech has operated within statistical op- timization frameworks, particularly Bayesian ap- proaches (Goldwater et al., 2009; Johnson and Goldwater, 2009). These models have established the state-of-the-art for the task of selecting appro- priate word boundaries from a stream of unstruc- tured phonemes. But while these models deliver excellent performance, it is not clear how they in- form the process of acquisition. Trying to find cognitive insight from these types of models is difficult because of the inherent mis- match in the quality and types of hypotheses they maintain during learning. Children are incremen- tal learners (Brown, 1973), and learners relying on statistical optimization are generally not. child’s competence grows gradually as she hearsA and produces more and more utterances, going through predictable changes to her working gram- mar (Marcus et al., 1992) that statistical optimiza- tion techniques typically do not go through and do not intend to replicate. Statistical models provide excellent information about the features, distributional cues, and priors that can be used in learning, but …

Content truncated to 500 words