?
Less is More: Instructed Text Simplification for People with Aphasia
Background: Patients with aphasia experience significant difficulties in reading and comprehending texts. Nevertheless, text-based activities constitute a substantial part of speech therapy programs for these patients – speech-language pathologists assign tasks such as reading texts aloud, answering comprehension questions, and, in cases of mild impairment, retelling the content. However, the textual material used must be adapted to the patients' linguistic and cognitive abilities, and this presents a challenge for clinicians: on the one hand, the texts should be engaging, relevant to the patient's life experiences, and personally motivating, while on the other hand, adapting them to the patient's capabilities requires considerable time and effort. To address this issue, we have developed a methodology for the automatic simplification of texts for speech therapy in individuals with aphasia.
Aims: The aim of this study is to investigate the effectiveness of using Large Language Models (LLMs) for multilevel text simplification for patients with aphasia.
Methods: A parallel corpus consisting of 134 texts divided into 1010 sentences in Russian was prepared. Using guidelines developed by speech therapists specializing in complex motor aphasia eight Russian-speaking annotators generated three simplified versions for each original sentence, corresponding to three levels of disease severity. Our designed multi-tiered instructions and a dataset comprising 100 original-to-simplified text pairs were used in a few-shot prompting strategy applied to four LLMs, which were then tasked with simplifying the texts.
Outcomes and Results: Models were evaluated using (1) linguistic complexity metrics and (2) BERTScore for semantic similarity. The top LLM's output was clinically validated with 12 complex motor aphasia patients. Patients read six texts (three auto-simplified, three manual), then answered questions or retold stories. Automated texts scored higher in “interestedness” but slightly lower in “easiness” versus manual versions.
Conclusions: Findings indicate that few-shot prompted automated simplification effectively adapts texts for complex motor aphasia patients. Automated versions showed comparable linguistic complexity and high semantic fidelity to originals. While marginally less comprehensible than manual simplifications, patients rated them as more engaging.