Data-driven Measurement of Child Language Development with Simple Syntactic Templates

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2151–2160, Dublin, Ireland, August 23-29 2014. Data-driven Measurement of Child Language Development with Simple Syntactic Templates Shannon Lubetich Pomona College Claremont, CA 91711 shannon.lubetich@pomona.edu Kenji Sagae Institute for Creative Technologies University of Southern California Los Angeles, CA 90089 sagae@ict.usc.edu Abstract When assessing child language development, researchers have traditionally had to choose be- tween easily computable metrics focused on superficial aspects of language, and more expressive metrics that are carefully designed to cover specific syntactic structures and require substantial and tedious labor. Recent work has shown that existing expressive metrics for child language development can be automated and produce accurate results. We go a step further and pro- pose that measurement of syntactic development can be performed automatically in a completely data-driven way without the need for definition of language-specific inventories of grammatical structures. As a crucial step in that direction, we show that four simple feature templates are as expressive of language development as a carefully crafted standard inventory of grammatical structures that is commonly used and has been validated empirically. 1 Introduction Although child language has been the focus of much study, our understanding of first language acquisi- tion is still limited. In attempts to measure child language development over time, several metrics have been proposed. The most commonly used metric is Mean Length of Utterance, or MLU (Brown, 1973), which is based on the number of morphemes per utterance. The main appeal of MLU is that it can be easily computed automatically, given machine-readable transcripts. Although MLU values may not be meaningful across languages, the general approach is suitable for analysis within different languages. However, MLU’s ability to track language development from age four has been questioned (Klee and Fitzgerald, 1985; Scarborough, 1990), and its usefulness is still the subject of debate (Rice et al., 2010). Several metrics based on the usage of grammatical structure have been proposed as more sensitive to changes in language over a wider range of ages (Scarborough, 1990; Lee and Canter, 1971; Fletcher and Garman, 1988). These metrics continue to show score increases where MLU plateaus, but their increased expressivity is typically associated with two severe drawbacks. The first is that their use for computa- tion of language development scores involves identification of several specific grammatical structures in child language transcripts, a process that requires linguistic expertise and is both time-consuming and error-prone. This issue has been addressed by recent work that shows that current natural language pro- cessing techniques can be applied to automate the computation of these metrics, removing the bottleneck of manual labor (Sagae et al., 2005; Roark et al., 2007; Sahakian and Snyder, 2012). The second draw- back is that these measures are language-specific, and development of a measure for a specific language requires deep expertise and careful design of an inventory of grammatical structures that researchers be- lieve to be indicative of language development. Going beyond previous work, which addressed the first drawback of …

Content truncated to 500 words