Page 185 - Read Online

Page 10 of 22 Hei et al. J. Mater. Inf. 2026, 6, 15

Figure 3. The proportion of sentences containing different numbers of tuples. The numbers below the pie chart indicate the number of
tuples represented by each colored segment. On the right side of the pie chart are examples of varying numbers of tuples within one
sentence. The proportions, in ascending order, are 24.36%, 33.09%, 23.64%, 11.64%, and 7.27%.

Table 1. Number of sentences and tuples in the test sets
Test dataset 1 2 3 4 Random
Num of sentences 40 38 38 22 23
Num of tuples 40 76 114 88 50

Table 2. Number of sentences and tuples in the entire dataset
Dataset 1 2 3 4 Total

Num of sentences 67 91 65 32 255
Num of tuples 67 182 195 128 568

also accommodates local nesting observed in materials discourse. For instance, “room temperature” contains
a CONDITION span “temperature” and a CONDITION VALUE span “room temperature”. Independent
heads per type allow recovery of both spans without enforcing mutual exclusivity across tags.

To tackle the challenge of multi-tuple extraction, we propose the task of entity allocation to assign extracted
entities of different types to complete tuples and avoid allocation errors caused by entity relationship
confusion. We believe that the key to this task is enabling the model to learn both correct and incorrect tuple
matching patterns contrastively. Hence, in the second stage, we construct an entity matching score matrix to
assess the likelihood of various entities being allocated together, using inter-entity and intra-entity attention
to provide representations informed by both the candidate pair and the surrounding entities of the same
type. Specifically, inter-entity attention yields representations between different types, and intra-entity
attention captures relationships within the same type. This design allows the model to utilize information
from the two entities being matched and from all entities of the corresponding types when performing tuple