Table 1 Semi supervised setting for the citation field extraction task

Figure 1: The test reward value of SG-SPEN's outputs trained in the supervised setting and semi-supervised settings We appreciate the reviewers' comments and concerns. SG-SPEN achieves 91.0% and DVN achieves 90.5% token-level accuracy. SG-SPEN with domain-knowledge based citation reward function, which resulted in 90.6% token-level accuracy. PG training with EMA baseline when the model is pre-trained with the labeled data. GE based on Mann & McCallum (2010).