Page 195 - Read Online
P. 195
Page 20 of 22 Hei et al. J. Mater. Inf. 2026, 6, 15
semantics raise ambiguity. We acknowledge the practical constraint that generating high-quality labels from
raw scientific texts requires substantial time and domain expertise, which can hinder rapid scalability.
Looking ahead, we plan to incrementally expand the corpus in future work while reducing reliance on
extensive manual annotation through transfer learning and data augmentation. A separate study in our
recent work preliminarily shows that transfer across related materials tasks improves performance under
[36]
data scarcity. Active learning could further optimize annotation efforts by prioritizing samples where the
model exhibits high uncertainty, particularly when entity allocation confidence scores fall below predefined
thresholds. This ensures that limited resources are focused on the most informative examples for model
improvement.
The current work focuses on narrative text. We observe that some data are presented in tables and figures,
especially in comparative studies across materials. However, in most cases, authors also describe the semantic
content of figures in the surrounding text, including the variables compared, conditions, and principal
findings. Hence, for the purposes of the present work, text-based extraction is sufficient for most
applications. Nevertheless, there remain cases in which key values exist only in plots or highly structured
tables. Addressing these scenarios requires computer vision techniques for chart and plot understanding,
including axis calibration, legend and marker parsing, and mapping graphical encodings to numerical series
with associated metadata such as units, conditions, and uncertainty. A rigorous treatment demands careful
handling of multi-panel layouts, non-linear axes, overplotting, and figure-specific conventions in materials
science journals.
CONCLUSIONS
In conclusion, to capture structured information in materials science literature, particularly within complex
corpora containing multiple scientific tuples, we propose an approach that combines a MatSciBERT-based
pointer extractor with an allocation module that enforces coherence through inter- and intra-entity
attention. Applied to the mechanical properties of MPEAs, exact match evaluation yields F1 scores of 0.96,
0.95, 0.85, and 0.75 for sentences containing one to four tuples, respectively, and 0.85 on a randomly curated
set, confirming strong and consistent performance as tuple complexity increases. Ablation studies indicate
that the allocation module is the key driver of precision, with inter-entity attention contributing more than
intra-entity attention to tuple correctness. Substituting SciBERT or BERT for MatSciBERT degrades
performance, with the gap widening as tuple density increases. Prompted large models exhibit inconsistent
tuple assembly, leading to lower exact match reliability in structurally complex sentences. These findings
establish the proposed method as a simpler, more effective, and accurate alternative to LLMs in specialized
tasks where precise multi-tuple alignment and scientific correctness are required for downstream modeling.
Future work will emphasize data efficiency and portability through transfer learning and active learning.
DECLARATIONS
Authors’ contributions
Data curation, model construction, and writing - original draft: Hei, M.
Analysis, methodology, and writing - review and editing: Zhang, Z.
Supervision: Liu, Q.; Zhao, X.; Ye, Y.
Data curation and visualization: Pan, Y.; Peng, Y.
Supervision and methodology: Zhang, X.
Supervision and funding acquisition: Bai, S.
All authors have reviewed and approved the manuscript.
Availability of data and materials
The JSON-formatted dataset and all code used in this work are available at https://github.com/HEI-MENGZ
HE/Material_Tuple.
