[TMP-068] Knowledge Extraction Through Prompting on Pre-trained Language Models
The use of prompt-based learning to extract temporal and conceptual knowledge from procedural texts for knowledge graph construction
Procedural documents are key sources of temporal procedural knowledge, spanning formats such as administrative processes, service manuals, medical guidelines, and surgical procedures. Extracting this complex, multidimensional knowledge—including temporal dimensions alongside static aspects like resources, tools, and costs—is crucial for tasks like information extraction, validation, and AI system development (e.g., expert surgical assistants).
Knowledge graphs provide a natural structure to represent such multidimensional knowledge, but automating their construction from procedural documents is challenging due to limited annotated data and real-world text repositories. While pre-trained language models show promise for extracting knowledge, their use for constructing knowledge graphs through prompt-based learning strategies remains underexplored.
This microproject investigates using prompt-based, in-context learning to extract conceptual information from procedural texts for conversion into knowledge graphs. A focus will be placed on multi-turn dialog strategies and incorporating prompts with relevant conceptual knowledge or varied examples (including negative ones) to enhance the extraction of tasks and temporal flow relationships. The project aims to advance structured narrative construction using machine learning models, enriching conceptual inputs. Additionally, insights from multi-turn dialog strategies may inform how such models complement traditional domain expert-driven knowledge modeling processes.
Partners
- Fondazione Bruno Kessler, Mauro Dragoni
- University of Verona, Marco Rospocher