Pragmatic inference and mapping for conversational implicature

Conversational implicatures arise when a speaker’s message is implicit rather than explicitly stated. Conversational implicature inference is a task to categorize the implicit message of an utterance into one of the given categories, which is challenging due to the subtle relationship between literal and implicated meanings. Previous work has explored Chain-of-Thought (CoT) prompting to guide large language models (LLMs) in capturing conversational implicatures, achieving notable advancements. However, the performance of earlier work is somewhat limited due to two main challenges. First, it is difficult to establish a clear and formal sequence of thoughts that would help LLMs understand conversational implicatures. Second, CoT is susceptible to misremembering original questions when addressing multiple sub-problems. An error in any subtask can negatively impact the final inference results. To tackle these challenges, we begin with literal meanings and engage in pragmatic inference that leads to implicated meanings. The framework is called Pragmatic Inference and Mapping (PIM), which starts with literal meanings, engages in pragmatic inference to draw intermediate conclusions, and ultimately reaches implicated situations. We conduct thorough experiments using various LLMs as base models on two publicly available datasets: LUDWIG and SwordsmanImp. Experimental results indicate that PIM enhances LLMs’ performances in conversational implicature inference tasks on Chinese and English datasets. Ablation and case studies show that pragmatic inference effectively guides LLMs in understanding conversational implicatures. We also explore the potential effects of decoding strategies and instructional prompts on LLMs, finding that they have minimal impact on the performance of PIM.