Mining heuristic evidence sentences for more interpretable document-level relation extraction

Enhancing cross-evidence reasoning graph for document-level relation extraction

“…In the process of extracting evidence, it is a complex task to establish the entity association among the mixed information of multiple sentences, which usually requires the analysis of cross-sentence relationships, and the relationships between sentences are relatively chaotic, so it is difficult to explicitly determine the relationship between entity pairs between sentences. To utilize higher-quality evidence for mining semantic interactions among entities, we enhance default rules based on Zhu et al (2023) , taking into account the semantic information of complex real-world documents. We employ the center-sentence rule to get more precise collection of evidence.…”

“…We present the statistics of our evidence extraction results on the DocRED development set in Table 1 . In addition, HESRE ( Zhu et al, 2023 ) and DRE-EA-path ( Xu et al, 2022 ) respectively use their own rules to extract evidence sentences. HESRE uses its own heuristic rules to extract evidence, while DRE-EA-PATH uses reinforcement learning to select evidence.…”

“… Huang et al (2021b) proposed three heuristic rules to extract higher-quality evidence, selecting information-rich path sets from the whole document and reducing the noise interference of irrelevant sentences. Based on Huang et al (2021b) , Zhu et al (2023) added heuristic rules combining neighbor sentences and bridge entities on top of the original rules to enhance entity representation. Both Huang et al (2021b) and Zhu et al (2023) have well-designed evidence extraction rules.…”

“…Based on Huang et al (2021b) , Zhu et al (2023) added heuristic rules combining neighbor sentences and bridge entities on top of the original rules to enhance entity representation. Both Huang et al (2021b) and Zhu et al (2023) have well-designed evidence extraction rules.…”