Authors: Dr. S. Thivyanathan, Dr. R. Anusha
Abstract: The unprecedented growth of conversational artificial intelligence agents has had a revolutionary impact on human-machine communication, but pragmatic competence—the capacity to understand and produce contextual meaning—is still an open problem for present-day technologies. This research provides a thorough linguistic study of pragmatics in human-AI interaction, which focuses on processing and producing meaning within the contexts of conversational agents' implicatures, presuppositions, speech acts, and common ground. Based on an empirical analysis of 50 transcripts of human-AI conversations, along with experimental work with 36 participants in the comparison of five conversational agents (ChatGPT-4, Google Bard, Microsoft Copilot, Claude 2, and LLaMA 2), the research concludes that although rule-based conversational agents stick to strict literal understanding, transformer models show emergent pragmatic competence through successful interpretation of indirect speech acts in 76% of the cases. However, Gricean implicatures remain difficult (recognized in only 34% of instances) and cross-turn common ground challenging (consistent in only 41% of examples).
DOI: https://doi.org/10.5281/zenodo.19946123
