Page 193 - Read Online

Page 18 of 22 Hei et al. J. Mater. Inf. 2026, 6, 15

Figure 5. F1 scores of the proposed complete model and its ablated variants. The proposed model achieves the best performance among all
configurations.

Table 9. Proportion of error causes for each dataset
Dataset 2 3 4
Syntactic complexity 20% 10.5% 28.6%
Annotation sparsity 0 10.5% 9.5%
Semantic overlap 40% 10.5% 47.6%
Entity extraction error 40% 68.4% 14.3%

Syntactic complexity is a recurrent source of error. Sentences with layered grammatical structures, embedded
clauses, long-distance dependencies, or unusual word order increase ambiguity in boundary detection and
argument assignment. For example, in the sentence “Among the developed Ti-xCr-3Sn alloys, Ti-6Cr-3Sn
alloy showed maximum ductility (41% fracture strain), maximum (58%) recovery ratio due to shape memory
effect and maximum pseudoelastic response due to stress induced martensitic transformation and twinning
deformation mode”, the model incorrectly extracted “Ti-xCr-3Sn” as a specific alloy name rather than
recognizing it as a compositional family, which in turn yields an incorrect tuple. As tuples accumulate within
a sentence, the same syntactic devices that improve narrative flow introduce competing local cues, increasing
the incidence of such errors.

Annotation sparsity constitutes a second, more prosaic limitation. Training examples featuring multituple
configurations and rare entity types are less common, reducing the model’s exposure to the very patterns
that dominate highdensity sentences. This limits the learned decision boundaries for complex extraction
cases and reduces generalization to unseen yet systematic configurations.

Semantic overlap represents the most significant challenge when sentences contain several tuples. Shared
lexical items, repeated subjects, or reused contextual frames can lead to conflicts in boundary assignment and
entity pairing. A typical pattern involves multiple properties reported for a single material under distinct
conditions, where the model may struggle to correctly associate each property-value pair with its
corresponding experimental condition.

Entity extraction errors constitute the fourth class and act as upstream error amplifiers. For instance, in the
sentence “The calculated elastic constants C11, elastic constants C12 and elastic constants C44 for the
Ti Zr Hf Nb Ta single crystalline disordered alloy yielded 160.2, 124.4 and 62.4 GPa, respectively”, if
20
20
20
20
20
“elastic constants” is misrecognized due to its rarity in the training data, all three resulting tuples will contain
incorrect property names.