Page 178 - Read Online

Hei et al. J. Mater. Inf. 2026, 6, 15 Page 3 of 22

However, despite extensive research on structured information extraction, effective methodologies for
multi-tuple extraction remain notably lacking. A tuple typically consists of a set of related elements that
represent structured information, often in the form of (subject, predicate, object) or (entity, attribute, value),
although in some cases, tuples may contain more elements. The presence of multiple tuples within text
complicates the accurate assignment of extracted entities to their corresponding tuples. This challenge is
compounded by contextual ambiguity, diverse modes of expression, and the difficulty of obtaining
high-quality annotated datasets. Even though LLMs demonstrate outstanding capabilities in NLP tasks, they
still exhibit limitations in understanding multi-tuple scientific texts and may produce hallucinations.
Moreover, LLMs typically require substantial computational resources for both training and inference, which
may limit their cost-effectiveness and efficiency in practical applications of data mining and knowledge
extraction.

In this study, we explore multi-tuple extraction of mechanical properties from alloys, mainly focusing on
multi-principal element alloys (MPEAs). We propose a domain-specific framework that integrates an entity
extraction model combining MatSciBERT with pointer networks and an entity allocation model employing
inter-entity and intra-entity attention mechanisms. To enable principled evaluation across varying linguistic
densities, we compile and annotate a sentence-level corpus from the MPEA literature, dividing sentences
according to the number of tuples they contain, and assess performance at the tuple level to reflect scientific
completeness rather than isolated spans. To begin, we formalize tuple construction as a two-stage problem,
consisting of entity extraction followed by allocation. Next, we conduct ablation studies to separate the
contributions of the allocation module and its inter- and intra-entity attention, quantifying their roles in
precision, pairing stability, and robustness as tuple density increases. Additionally, we perform error analyses
to characterize residual failure modes, including cross-assignment between properties and conditions,
semantic overlap among entities, and syntax-induced ambiguity, and we relate these errors to specific
architectural choices. Finally, we examine substitutes for the domain encoder and overall paradigm by
replacing MatSciBERT with SciBERT and BERT and by benchmarking against LLMs under controlled
prompts, thereby clarifying the advantages of a domain-specific, scientifically faithful design for precise
multi-tuple alignment.

MATERIALS AND METHODS
Workflow overview
Figure 1 presents the overall workflow of the study. Initially, we constructed a corpus comprising over 200
full-text papers containing the keywords “multi-principal element alloy” and “mechanical properties”. From
this corpus, we extracted sentences describing alloy properties and annotated five distinct types of entities
within these sentences: MATERIAL, PROPERTY, PROPERTY VALUE, CONDITION, and CONDITION
VALUE. This process yielded 255 sentences containing varying numbers of tuples, resulting in a total of 568
tuples stored in a JSON-formatted file. The modeling pipeline proceeds in two stages: the first stage extracts
entity spans, and the second stage assigns entities to complete tuples within each sentence. Detailed
specifications of the annotation protocol, model architecture, and training procedures are provided in the
subsequent subsections.

Literature acquisition
The materials science literature used in this study was obtained from Elsevier with publisher consent. This
work mainly focuses on high-entropy alloys. A large amount of papers relevant to high-entropy alloys were
retrieved, and the top 200 papers were selected and downloaded in HTML/XML format based on their
relevance and timeliness. These papers were then converted into text files to form a corpus for subsequent
processing.