A computational semantics approach to aphasic sentence comprehension

Abstract

Background: Patients with brain damage often exhibit difficulty understanding sentences, with certain grammatical constructions posing more of a problem than others. Differences between sentences that are “hard” or “easy” to process have been characterised in terms of modern syntactic theory, leading to a number of insightful proposals regarding the nature of sentence comprehension problems in aphasia. However, little attention has been devoted to the semantic aspects of aphasic sentence comprehension. Aims: The primary aim of this research is to validate the use of a computational semantic approach for modelling aphasic sentence comprehension. Methods & Procedures: The model presented here is an extension of natural language processing software designed by Blackburn and Bos (2006). The original program parses a natural language expression by means of a simple definite clause grammar, assigning a semantic representation to each node in the parse tree. The final result of a successful parse is a sentence from first‐order logic that describes the meaning of the natural language sentence. The program was made relevant for the study of aphasia by extending the grammar to parse 14 sentences that present variable degrees of difficulty to aphasic patients. In addition, each constituent in the grammar was endowed with an integrity feature that contained an integer with a maximum value of 100. This number constituted the percent chance that the node would be successfully realised and was reduced in proportion to the node's height in the syntactic tree. The constant of proportionality was then manipulated to simulate various degrees of aphasia severity. Qualities of the model's performance were compared to qualities of aphasic patient performance on four key sentences. The model's quantitative performance on 12 sentences was compared to that of 46 patients with left hemisphere lesions. Outcomes & Results: There was a significant correlation between the performance of the model and that of the patients (r = .85, p < .001). Sentence length was not significantly correlated with patient performance (r = −.52, ns). The probabilistic output of the model resembles variability in performance by aphasic patients. Conclusions: A computational semantics approach to aphasic sentence comprehension may provide a means for explaining continuous variation in degrees of aphasia severity as well as qualitative patterns of agrammatic comprehension.