Page 194 - Read Online

Hei et al. J. Mater. Inf. 2026, 6, 15 Page 19 of 22

For dataset 2, semantic overlap and entity extraction each account for 40% of errors, with syntactic
complexity contributing 20%. In dataset 3, entity extraction dominates at 68.4%, while the remaining
categories each contribute 10.5%. In dataset 4, semantic overlap is the primary source of error at 47.6%,
followed by syntactic complexity at 28.6%. Although entity extraction errors constitute the largest overall
proportion, no clear linear relationship is observed between error type distribution and tuple density.
Importantly, these categories often cooccur in dense sentences, where semantic overlap and syntactic
complexity combine to create boundary ambiguities that are difficult even for human annotators. When
focusing specifically on the entity allocation task (excluding upstream entity extraction errors), semantic
overlap and syntactic complexity emerge as the principal bottlenecks, particularly in dataset 4. This
observation aligns with the intuition that allocation is most sensitive to reused constituents and partial token
reuse across tuples, both of which increase with tuple density. In addition, we analyze potential concentrated
vulnerabilities across entity-type combinations and specific sentence structures, with detailed discussion
provided in the Supplementary Materials. Although no such vulnerability is observed for entity-type
combinations, dataset 4 shows a higher error rate associated with increased structural diversity, with errors
concentrated in the most complex template featuring parallel quantification and cross-cutting condition
mentions [Supplementary Figure 1 and Supplementary Table 1].

By examining MPEAs as a case study, our work distinguishes itself from traditional approaches in two key
aspects. First, we prioritize the accuracy of the structured information extracted by the model, rather than
focusing solely on entity extraction performance. Second, instead of using the conventional named entity
recognition and relation extraction (NERRE) framework, we employ an integrated methodology that
combines entity extraction and matching to generate structured information. Isolated entities often lack the
capacity to convey complete scientific meaning; therefore, accurate multidimensional tuples are essential for
representing structured information in the literature. Additionally, we note that it is sufficient to identify
which entities belong to the same tuple without explicitly predicting the relationships between individual
entities. When a sentence contains multiple tuples, our integrated entity extraction and matching approach
proves to be more precise and efficient compared to the NERRE method, which is more susceptible to errors
and incurs unnecessary computational overhead.

LLMs employ a pre-training-plus-prompting approach, where users construct prompts and provide inputs
to guide the model in performing specific tasks. In contrast, our methodology adopts a
pre-training-plus-fine-tuning framework, which involves the deliberate design of deep learning networks
and systematic data annotation. This approach not only enhances model capability but also significantly
improves performance in extracting multiple tuples, as demonstrated in this study. Although LLMs offer a
more user-friendly interface, their outputs are influenced by commercial access limits and weak control over
error modes. Another notable drawback of LLMs is the phenomenon of “hallucination” [34,35] , in which models
generate plausible but misleading information not grounded in factual data, such as misinterpreting a
specific element within an alloy as a CONDITION entity. Although some of these outputs may seem
reasonable, they ultimately compromise the accuracy required for successful entity extraction. Our approach,
on the other hand, as a specialized method designed for this task, offers greater precision compared to the
broader but less accurate capabilities of LLMs and can be easily adapted to other materials domains given
adequate high-quality data for fine-tuning. Nonetheless, this does not imply that LLMs are inherently
inadequate. With appropriate fine-tuning, their extraction performance can improve, and they remain strong
in broader tasks such as language understanding, coding, and reasoning.

Our approach has associated costs. Fine-tuning requires high-quality annotations at the tuple level, typically
on the order of hundreds of verified tuples per domain to stabilize allocation performance. Data scarcity
becomes more pronounced as tuple density increases, where abbreviations, omissions, and overlapping