Teaching LLMs to unveil tendentious implicit contents of Italian political communication

This paper investigates the capacity of Large Language Models (LLMs) to interpret and explain non-bona fide true implicit content within the specific context of Italian political communication. While implicit meaning is a fundamental aspect of pragmatic competence, it is frequently exploited in political discourse through presuppositions and implicatures to convey tendentious messages without overt commitment. Building on previous research that highlighted the limitations of general-purpose models in zero-shot settings, we perform instruction-based fine-tuning on two state-ofthe-art open-weight models: Llama3.1 and Qwen2.5. We utilize the IMPAQTS-PID benchmark, a dataset derived from the IMPAQTS corpus, to train these models to generate natural language explanations of manipulative implicit meanings. Our experimental results, validated through manual evaluation by independent annotators, demonstrate that fine-tuning significantly improves model performance compared to previous benchmarks, and that both models perform substantially better at explaining presuppositions than implicatures.