Frontiers in Psychology (Apr 2014)
Sentence comprehension in aphasia: A noisy channel approach
Abstract
Probabilistic accounts of language understanding assume that comprehension involves determining the probability of an intended message (m) given an input utterance (u) (P(m|u); e.g. Gibson et al, 2013a; Levy et al, 2009). One challenge is that communication occurs within a noisy channel; i.e. the comprehender’s representation of u may have been distorted, e.g., by a typo or by impairment associated with aphasia. Bayes’ rule provides a model of how comprehenders can combine the prior probability of m (P(m)) with the probability that m would have been distorted to u (P(mu)) to calculate the probability of m given u (P(m|u) P(m)P(mu)). This formalism can capture the observation that people with aphasia (PWA) rely more on semantics than syntax during comprehension (e.g., Caramazza & Zurif, 1976): given the high probability that their representation of the input is unreliable, they weigh message likelihood more heavily. Gibson et al. (2013a) showed that unimpaired adults are sensitive to P(m) and P(mu): they more often chose interpretations that increased message plausibility or involved distortions requiring fewer changes, and/or deletions instead of insertions (see Figure 1a for examples). Gibson et al. (2013b) found PWA were also sensitive to both P(m) and P(mu) in an act-out task, but relied more heavily than unimpaired controls on P(m). This shows group-level optimization towards the less noisy (semantic) channel in PWA. The current experiment (8 PWA; 7 age-matched controls) investigated noisy channel optimization at the level of individual PWA. It also included active/passive items with a weaker plausibility manipulation to test whether P(m) is higher for implausible than impossible strings. The task was forced-choice sentence-picture matching (Figure 1b). Experimental sentences crossed active versus passive (A-P) structures with plausibility (Set 1) or impossibility (Set 2), and prepositional-object versus double-object structures (PO-DO: Set 3) with plausibility. Target pictures depicted the observed utterance u; foils depicted a message that could have been distorted to u (Figure 1a-b). Replicating Gibson et al (2013b), both controls and PWA more often chose foils when the possible distortions involved fewer changes (DO-PO compared to A-P: F[1,13]=4.82, p<.05). This is despite passives’ low frequency and common impairment in aphasia (Schwartz, et al., 1980). Furthermore, although both groups more often chose foils when the possible distortion was more plausible, this preference was larger for PWA (plausibility x group interaction : F[1,13]=12.09, p<.01). The strength of the semantic manipulation did not matter: plausibility and possibility manipulations did not differ. Interestingly, there was little evidence that an individual’s reliance on the form of the input vs. the likelihood of a message was predicted by their syntactic vs. semantic abilities. Standardized sentence-comprehension scores (Comprehensive Aphasia Test: Swinburn, et al., 2004) did not predict preference for simpler distortions, nor did conceptual-semantic processing measures (e.g., Kissing and Dancing: Bak & Hodges, 2003; Pyramids and Palm Trees: Howard & Patterson, 1992) predict the size of plausibility effects. Additionally, individual participants showed non-optimization: one PWA with relatively spared syntax (good performance on reversible passives) but impaired semantics (poor conceptual-semantic processing scores) relied almost exclusively on semantics.
Keywords
