Under-resourced or overloaded? Rethinking working memory deficits in developmental language disorder

• In Event: Poster Session 02
  In Poster Session: PS 02 Topic - Language, Communication

Abstract

Difficulty understanding long or complex sentences is the hallmark of developmental language disorder (DLD). This difficulty is often attributed to limited verbal working memory capacity, where working memory is understood–following Baddeley and Hitch (1974)–as a buffer system to which long-term representations are transferred for in-the-moment processing. Commonly, it is argued that limited verbal working memory capacity makes it difficult for children with DLD to hold successive morphemes in mind and form an integrated and comprehensible sentence representation (e.g. Montgomery, 1995).

This talk presents an alternative account of sentence comprehension deficits in DLD. Our position is motivated by two recent research trends. First is work highlighting the importance of anticipation during sentence processing. Eye tracking and electroencephalography work indicate that from as early as 30 months of age, typically developing children make implicit predictions about the linguistic features they might encounter, supporting processing efficiency and rapid ambiguity resolution. The second trend involves research conceptualizing working memory as activated long-term memory rather than as a functionally distinct buffer. This view is supported not least by neuroimaging work showing equivalent regions of activation during working memory and long-term memory tasks.

Our claim is that among children with DLD, reported auditory processing deficits (e.g. McArthur & Bishop, 2005) affect the quality of long-term language encoding–namely, the mental representations of morpheme sounds–and therefore the efficiency with which long-term memory networks activate through association with features of an unfolding spoken sentence. In essence, children affected by DLD may be activating larger cohorts of long-term language representations during incremental, anticipatory sentence processing than their typically developing peers. As Oberauer (2019) has argued, it is crucial that working memory–understood as activated long-term memory–is gated to allow the efficient deployment of attention to achieve aims such as establishing a comprehensible sentence representation. With attention overwhelmed by the flooding of working memory due to deficient long-term encoding, the child is unlikely to be able to retain considerable sequential information, with the outcome that sentence comprehension fails.

This account is developed through computational simulations using neural networks (LSTMs) trained to predict the next words in sentences from a large corpus of authentic child-directed speech. We present original data showing that implementing a low-level auditory encoding deficit significantly impedes the networks’ ability to predict the next words in a series of naturalistic test sentences (~29% prediction accuracy in DLD networks compared to ~35% prediction accuracy in control networks). Furthermore–and central to the development of our theoretical account–we show that even when the networks make accurate predictions, networks simulating DLD assign significantly lower probability to the next word and produce a predictive distribution characterized by substantially higher entropy than control networks without auditory encoding deficits. This signifies that a greater proportion of the DLD networks’ long-term language memory is activated in response to sentence features, providing proof of concept that the sentence processing and comprehension deficits of children with DLD may stem from working memory being overloaded, rather than under-resourced.

Authors

• Samuel David Jones, Lancaster University
  Presenting Author

• Gert Westermann, Lancaster University
  Non-Presenting Author