Using constituency and dependency parse features to identify errorful words in disordered language.
Eric J. Morley, Emily Tucker Prud’hommeaux · 2012
Delayed or disordered language is a characteristic of both autism spectrum disorder (ASD) and specific language impairment (SLI). In this paper, we describe our data set, which consists of transcribed data from a widely used clinical diagnostic instrument (the ADOS) for children with ASD and children with SLI. These transcripts are manually annotated with SALT, an annotation system that applies a descriptive code to errorful words. Here we address a step in automating SALT annotation: identifying the errorful words in sentences that are known to contain an error. We propose a set of baseline features to identify errorful words, and investigate the effectiveness of adding features extracted from dependency and constituency parses. We find that features from both types of parses improve classifier performance above our baseline, both individually and in aggregate.