A Corpus-Based Approach for the Prediction of Language Impairment in Monolingual English and Spanish-English Bilingual Children

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 46–55, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics A Corpus-Based Approach for the Prediction of Language Impairment in Monolingual English and Spanish-English Bilingual Children Keyur Gabani and Melissa Sherman and Thamar Solorio and Yang Liu Department of Computer Science The University of Texas at Dallas keyur,mesh,tsolorio,yangl@hlt.utdallas.edu Lisa M. Bedore and Elizabeth D. Pe˜na Department of Communication Sciences and Disorders The University of Texas at Austin lbedore,lizp@mail.utexas.edu Abstract In this paper we explore a learning-based ap- proach to the problem of predicting language impairment in children. We analyzed sponta- neous narratives of children and extracted fea- tures measuring different aspects of language including morphology, speech fluency, lan- guage productivity and vocabulary. Then, we evaluated a learning-based approach and com- pared its predictive accuracy against a method based on language models. Empirical re- sults on monolingual English-speaking chil- dren and bilingual Spanish-English speaking children show the learning-based approach is a promising direction for automatic language assessment. 1 Introduction The question of how best to identify children with language disorders is a topic of ongoing debate. One common assessment approach is based on cut- off scores from standardized, norm-referenced lan- guage assessment tasks. Children scoring at the lower end of the distribution, typically more than 1.25 or 1.5 Standard Deviations (SD) below the mean, are identified as having language impair- ment (Tomblin et al., 1997). This cutoff-based approach has several well-documented weaknesses that may result in both over- and under-identification of children as language impaired (Plante and Vance, 1994). Recent studies have suggested considerable overlap between children with language impairment and their typically developing cohorts on many of these tasks (e.g., (Pe˜na et al., 2006b; Spaulding et al., 2006)). In addition, scores and cutoffs on stan- dardized tests depend on the distribution of scores from the particular samples used in normalizing the measure. Thus, the validity of the measure for chil- dren whose demographic and other socioeconomic characteristics are not well represented in the test’s normative sample is a serious concern. Finally, most norm-referenced tests of language ability rely heav- ily on exposure to mainstream language and expe- riences, and have been found to be biased against children from families with low parental education and socioeconomic status, as well as children from different ethnic backgrounds (Campbell et al., 1997; Dollaghan and Campbell, 1998). This paper aims to develop a reliable and auto- matic method for identifying the language status of children. We propose the use of different lexico- syntactic features, typically used in computational linguistics, in combination with features inspired by current assessment practices in the field of lan- guage disorders to train Machine Learning (ML) al- gorithms. The two main contributions of this pa- per are: 1) It is one step towards developing a re- liable and automatic approach for language status prediction in English-speaking children; 2) It pro- vides evidence showing that the same approach can be adapted to predict language status in …

Content truncated to 500 words