Page 183 - Read Online

Page 8 of 22 Hei et al. J. Mater. Inf. 2026, 6, 15

To encode common order constraints, when two entity types appear in equal numbers within a sentence and
correct alignments typically follow textual order, we add a small diagonal bias λ to the matching matrix
during inference. This favors order-consistent matches without overriding the learned scores; λ is tuned on
the validation set.

Model and training setup
Training and evaluation are conducted on four NVIDIA RTX 3090 GPUs. Hyperparameters follow
established practice for finetuning pre-trained language models for information extraction. For entity
extraction, the model is fine-tuned for up to 50 epochs using the AdamW optimizer, with a learning rate of
1e-5 and a batch size of 16. The learning rate follows standard practice for BERT-based models to avoid
catastrophic forgetting while enabling effective adaptation. For entity allocation, the model is fine-tuned for
up to 50 epochs using the AdamW optimizer, with a learning rate of 1e-5 and a batch size of 8. The reduced
batch size accommodates the computational complexity of the allocation module while maintaining training
stability.

RESULTS AND DISCUSSION
Multi-tuple analyses
We describe several prevalent multi-tuple patterns observed in the aforementioned scenarios, as illustrated in
Figure 2. These patterns include multiple properties of the same material, various property values for the
same material under different conditions, and multiple values of the same property across different materials.
Furthermore, these common patterns can be interwoven within more complex textual contexts, which are
frequently encountered in the literature pertaining to diverse materials. For instance, from the sentence
depicted in Figure 2B, two tuples can be extracted: [“AlNbTiV”, “yield strength”, “1,020 MPa”,
“temperature”, “room temperature”] and [“AlNbTiV”, “yield strength”, “685 MPa”, “temperature”, “800 °C”].
In this context, the extraction of individual entities alone proves inadequate for conveying the scientific
significance embedded within the literature. Instead, complete tuples comprising various types of entities
must be extracted.

Moreover, the presence of multiple tuples conveying scientific meaning within a single sentence is prevalent
in the materials science literature. Our statistical analysis, based on over 100 publications on MPEAs, reveals
that sentences describing alloy properties containing only a single tuple account for merely 24.36%. The
distribution of sentences containing varying numbers of tuples is illustrated in Figure 3. Nevertheless, there
remains a paucity of studies proposing effective and sophisticated solutions to address this issue.

Dataset analyses
Following the protocol detailed in the Methods, the number of sentences and tuples in the test sets and the
entire dataset are presented in Tables 1 and 2 (These datasets are denoted as 1, 2, 3, 4, and Random in the
tables, and this notation is consistent throughout the subsequent sections).

Published materials information extraction datasets differ markedly in domain coverage, task objectives, and
annotation granularity, which complicates size-based comparisons. For instance, the superalloy pipeline
reports 680 text-extracted instances of γ′ solvus temperature. MSMENTIONS includes 595 synthesis
procedures, and SOFCExp covers 45 articles. These datasets target distinct units of annotation, ranging from
named entities and pairwise relations to processes and events. Consequently, comparing corpora based solely
on relation or instance counts can obscure differences in scope and semantic completeness.

Our contribution is designed at the tuple level with scenario-centric annotation. The 568 tuples in our corpus
are drawn from multi-event sentences that explicitly bind at least Material, Property, and Property Value,