Answer Set Programming (ASP) is a well-known declarative paradigm within the field of knowledge representation and reasoning, extensively used in both academic and industrial settings. The success of ASP in critical decision-making applications makes it crucial to explain its complex logic and reasoning processes to non-experts. Recently, Large Language Models (LLMs), like ChatGPT, have demonstrated remarkable capabilities in explaining complex code segments, particularly in widely-used imperative languages such as Python. Notably, these models are able to interpret ASP rules as well. However, a significant challenge with LLMs is their tendency to “hallucinate”, meaning they can generate misleading or incorrect outputs. This issue is particularly problematic in applications where precision is critical. The paper introduces two novel tools, ASP2CNL and CNL2NL, which can be combined together to translate ASP rules into natural language with the goal of mitigating the errors of LLMs when explaining ASP rules. The effectiveness of these tools has been validated during a review phase involving three external ASP experts, demonstrating a significant reduction in the occurrence of wrong outputs from LLM tools.
On the Translation of ASP Rules to (Controlled) Natural Language Sentences
Dodaro C.;Lo Scudo F.;Maratea M.;Reale K.
2025-01-01
Abstract
Answer Set Programming (ASP) is a well-known declarative paradigm within the field of knowledge representation and reasoning, extensively used in both academic and industrial settings. The success of ASP in critical decision-making applications makes it crucial to explain its complex logic and reasoning processes to non-experts. Recently, Large Language Models (LLMs), like ChatGPT, have demonstrated remarkable capabilities in explaining complex code segments, particularly in widely-used imperative languages such as Python. Notably, these models are able to interpret ASP rules as well. However, a significant challenge with LLMs is their tendency to “hallucinate”, meaning they can generate misleading or incorrect outputs. This issue is particularly problematic in applications where precision is critical. The paper introduces two novel tools, ASP2CNL and CNL2NL, which can be combined together to translate ASP rules into natural language with the goal of mitigating the errors of LLMs when explaining ASP rules. The effectiveness of these tools has been validated during a review phase involving three external ASP experts, demonstrating a significant reduction in the occurrence of wrong outputs from LLM tools.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
