Characterizing Verb Argument Structure in Aphasia Using Dependency Parsers.

Area of Science:

• Neurolinguistics
• Computational Linguistics
• Speech Pathology

Background:

• Verb argument structure (VAS) is frequently impaired in individuals with post-stroke aphasia.
• Previous research on VAS impairments often relied on constrained experiments or time-intensive manual annotation of narrative speech.
• Automated analysis of VAS in spontaneous speech is needed for efficient clinical application.

Purpose of the Study:

• To develop and validate an automated approach for quantifying VAS use in narrative speech using natural language processing (NLP).
• To apply this automated method to a new dataset of individuals with aphasia and compare their VAS use to controls.
• To assess VAS impairments specifically in agrammatism.

Main Methods:

• Validated an automated VAS quantification method by comparing its output to manually annotated measures from prior research.
• Applied the automated method to narrative discourse samples from 106 participants with aphasia and a control group from AphasiaBank.
• Utilized a dependency parser, a type of NLP tool, for automated annotation of VAS in speech transcriptions.

Main Results:

• The automated VAS analysis revealed that participants with Broca's aphasia and agrammatism used verbs with fewer arguments on average and produced fewer arguments compared to controls.
• Replicated previous findings on VAS impairments in post-stroke aphasia using the automated method.
• Demonstrated significant differences in VAS sensitivity and abilities between participants with aphasia and control participants.

Conclusions:

• Dependency parsers are effective for characterizing VAS use in spontaneous speech with minimal manual effort.
• The developed automated approach is feasible for creating clinically applicable tools for complex language analysis.
• Individuals with aphasia exhibit distinct patterns in their verb argument structure use compared to neurologically healthy individuals.