Exploring Abductive Reasoning in Language Models for Data-to-Text Generation

Abductive reasoning remains underexplored in language models despite its everyday human use, effectiveness in handling incomplete information, and use in automated planning. We present a data-to-text generation pipeline that prompts language models with abductive tasks to investigate its applica-bility. We show its utility in content selection, though generating a discourse plan for selected content presents challenges for non-fine-tuned language models. The three-stage pipeline allows for the deployment of more suitable models for different stages (reasoning and realization). This work highlights the potential of symbolic reasoning approaches in enhancing language models.