.png){kind=icon}
.png){kind=icon}
Return
Contextualised segment-wise citation function classification
.png){kind=icon}
PRE
AI
DOI：10.1007/s11192-023-04778-3.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Much effort has been made in the past decades to citation function classification, but noteworthy issues exist. Annotation difficulty resulted in limited data size, especially for minority classes, and inadequate representativeness of the underlying scientific domains. Concerning algorithmic classification, state-of-the-art deep learning-based methods are flawed by generating a feature vector for the whole citation context (or sentence) and failing to exploit the full realm of citation modelling options. Responding to these issues, this paper studied contextualised citation function classification. Specifically, a large new citation context dataset was created by merging and re-annotating six datasets about computational linguistics. A variety of strong SciBERT-based citation function classification models were proposed, and new states of the art were achieved. Through deeper performance analysis, this study focused on answering several research questions about the effective ways of performing citation function classification. More specifically, the study justified the necessity of modelling in-text citations in context and confirmed the superiority of doing citation function classification at citation (segment) level. A particular emphasis was placed on in-depth per-class performance analysis to understand whether citation function classification is robust enough to suit various popular downstream applications and what further efforts are required to meet such analytic needs. Finally, a naive ensemble classifier was proposed, which greatly improved citation function classification performance.
Keywords:
Citation context analysis
Citation function classification
Deep learning
SciBERT
Ensemble