Hybrid and Generative Models for Material Science Event Extraction, Anran Hao, Jian Su, AI4X 2025 Oral
Abstract:
The vast amount of scientific knowledge stored
in textual format, such as scientific papers, patents,
and technical reports, creates a tremendous opportunity to develop and build information extraction
tools to enable faster discovery, synthesis, and deployment into a wide variety of applications.
In this paper, we look into event extraction, a
sophisticated type of information extraction, which
is needed for knowledge discovery, but the state of
art performances are far from satisfactory. In particular, we focus on event extraction from material science papers which lacks corresponding studies. We extend our hybrid deep learning model
to material science event extraction, which equips
syntactic information embedded in Large language
models (LLMs) through its dependency tree-based
residual connections for effective and efficient reinforcement. Our reinforcement units do not rely
on external linguistic tools or domain-specific annotations, which are expensive to obtain. Experimental results demonstrate that our hybrid model
achieves new state-of-the-art performance on the
SC-CoMIcs benchmark dataset, establishing a highperformance baseline in the field.
We also evaluate the generative approach through
a LLM model. While there remains a performance
gap between the generative approach and our hybrid model, the work leads to the future explorations
on the generative approach with even larger LLMs
as well as the potential integration with the merits
of both generative approaches and the hybrid deterministic model besides incorporating syntactic reinforcement into generative models. Besides, we also
plan to extend our work to other material science
event extraction tasks to support the experiment automation and knowledge discoveries in general, as
well as other science disciplines, such as biology and
climate.
License type:
Attribution 4.0 International (CC BY 4.0)
Funding Info:
This research / project is supported by the National Research Foundation - Campus for Research Excellenceand Technological Enterprise (CREATE) programme for the programme DesCartes
Grant Reference no. : NA