Towards Standards-Compliant Assistive Technology Product Specifications via LLMs
Authors:
Chetan Arora,
John Grundy,
Louise Puli,
Natasha Layton
Abstract:
In the rapidly evolving field of assistive technology (AT), ensuring that products meet national and international standards is essential for user safety, efficacy, and accessibility. In this vision paper, we introduce CompliAT, a pioneering framework designed to streamline the compliance process of AT product specifications with these standards through the innovative use of Large Language Models…
▽ More
In the rapidly evolving field of assistive technology (AT), ensuring that products meet national and international standards is essential for user safety, efficacy, and accessibility. In this vision paper, we introduce CompliAT, a pioneering framework designed to streamline the compliance process of AT product specifications with these standards through the innovative use of Large Language Models (LLMs). CompliAT addresses three critical tasks: checking terminology consistency, classifying products according to standards, and tracing key product specifications to standard requirements. We tackle the challenge of terminology consistency to ensure that the language used in product specifications aligns with relevant standards, reducing misunderstandings and non-compliance risks. We propose a novel approach for product classification, leveraging a retrieval-augmented generation model to accurately categorize AT products aligning to international standards, despite the sparse availability of training data. Finally, CompliAT implements a traceability and compliance mechanism from key product specifications to standard requirements, ensuring all aspects of an AT product are thoroughly vetted against the corresponding standards. By semi-automating these processes, CompliAT aims to significantly reduce the time and effort required for AT product standards compliance and uphold quality and safety standards. We outline our planned implementation and evaluation plan for CompliAT.
△ Less
Submitted 3 April, 2024;
originally announced April 2024.
