Page 180 - Read Online

Hei et al. J. Mater. Inf. 2026, 6, 15 Page 5 of 22

MATERIAL denotes spans referring to specific alloys, typically compositional formulae (e.g.,
Ti–30Nb–1Mo–4Sn) or established abbreviations (e.g., C.P. Ti). PROPERTY denotes spans representing a
material property, such as elastic modulus, tensile strength, or yield strength. Synonymous labels follow the
authors’ original wording within each sentence (e.g., UTS vs. ultimate tensile strength) to preserve textual
fidelity. PROPERTY VALUE denotes numerical values with units corresponding to the property, including
inequality and range expressions (e.g., over 800 MPa and > 40%). CONDITION denotes measurement
conditions that affect the property (e.g., temperature). Its inclusion is essential due to the sensitivity of
certain properties to measurement conditions, and its absence would render the entire tuple introduction
incomplete. CONDITION VALUE denotes the numerical value and unit for a given condition. It is worth
noting that our dataset exhibits varying entity frequencies, particularly for CONDITION and CONDITION
VALUE, which occur less frequently than other entity types. This frequency variation is not attributable to
subjective factors but reflects the common practice in materials science literature, where most experiments
are conducted under standard conditions (room temperature and atmospheric pressure), leading authors to
omit these default parameters. Thus, tuple annotation only allows for the absence of CONDITION and
CONDITION VALUE. Tuples missing information in the first three entity types are not considered gold
labels. Given the stringent requirements and laborious nature of data annotation, a subset of 568 golden
tuples was extracted from 255 sentences within the corpus.

Dataset construction and splits
To evaluate performance as tuple density increases, sentences were categorized according to the number of
gold tuples per sentence: 1, 2, 3, and ≥ 4. Sentences containing more than four tuples were excluded due to
their scarcity. Because categories with 1, 2, or 3 tuples per sentence predominate, 15% of each of these
categories was randomly sampled to form their test subsets. The 4-tuple category is smaller; therefore, to
preserve sufficient training coverage, 10% was sampled for testing. For the Random dataset, 10% was also
sampled following the same principle. The remaining data were split into training and validation sets at a 9:1
ratio after reserving the test portion.

Entity extraction
As shown in Figure 1A, we adopt a pointer-based span extraction method over contextual token
embeddings, retaining the original formulation and notation except where clarification is necessary. Let the
input sentence be t = [t , …, t ]. The encoder produces contextual vectors X for each token :
[17]
n
1
t
X = MatSciBERT( ) (1)
where MatSciBERT is a 12-layer, 768-hidden, 12-head transformer pre-trained on materials science
literature. SciBERT and BERT-base are interchangeable backbones with identical architectures. Replacement
experiments are conducted using the same extraction architecture and hyperparameters. Details about model
selection and comparative outcomes are reported in the Results and Discussion section.

After obtaining the vector representations, a pointer network is used to train two binary classifiers per entity
type to predict the start and end positions. Specifically, the model calculates the probability that each token,
generated by MatSciBERT, corresponds to the start or end of the desired entity. Fully connected layers and
activation functions are used to calculate the probabilities of the head and tail pointers :
[28]

material ( ) = Sigmoid( material · X + b ) (2)

material ( ) = Sigmoid( material · X + b ) (3)