Dynamic Context Extraction for Citation Classification

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 539–549 November 20–23, 2022. ©2022 Association for Computational Linguistics 539 Dynamic Context Extraction for Citation Classification Suchetha N. Kunnath, David Pride, Petr Knoth Knowledge Media Institute (KMi) The Open University Milton Keynes UK {snk56, david.pride, petr.knoth}@open.ac.uk Abstract We investigate the effect of varying citation context window sizes on model performance in citation intent classification. Prior studies have been limited to the application of fixed-size contiguous citation contexts or the use of manually curated citation contexts. We introduce a new automated unsupervised approach for the selection of a dynamic-size and potentially non-contiguous citation context, which utilises the transformer-based document representations and embedding similarities. Our experiments show that the addition of non-contiguous citing sentences improves performance beyond previous results. Evalu- ating on the (1) domain-specific (ACL-ARC) and (2) the multi-disciplinary (SDP-ACT) dataset demonstrates that the inclusion of additional context beyond the citing sentence significantly improves the citation classifi- cation model’s performance, irrespective of the dataset’s domain. We release the datasets and the source code used for the experiments at: https://github.com/oacore/ dynamic_citation_context 1 Introduction Understanding citation types has served a wide range of applications, including research evalua- tion (Jurgens et al., 2018), article summary genera- tion (Nanba et al., 2000) and information retrieval (Valenzuela et al., 2015) to name a few. Classifying citation types according to their purpose or intent can make use of a variety of features, the most es- sential of which is the contextual textual fragment (context window) surrounding the citation marker within the citing article (Abu-Jbara et al., 2013; Jha et al., 2017). This information, also known as citation context, articulates how a cited work is presented in a research paper. Several citation type taxonomies of widely varying granularity have been used for citation type classification in the past (Kunnath et al., 2021). The taxonomy originally in- troduced by Jurgens et al. has been used across the two largest annotated datasets for citation typing, ACT (Pride et al., 2019) and ACL-ARC (Jurgens et al., 2018) and is shown in Appendix A. Although evidence indicates that the size of the citation context window matters, there is not yet a consensus about its optimal size. While some researchers argue that multi-sentence context win- dows only add noise, thus confining their focus to the citing sentence alone (Dong and Schäfer, 2011; Cohan et al., 2019), others emphasise the need to incorporate longer citation context to avoid infor- mation loss (Abu-Jbara et al., 2013; Jha et al., 2017; Lauscher et al., 2021). Most citation intent classification methods rely on a fixed-size contiguous citation context window (most typically one sentence) (Abu-Jbara et al., 2013; Hernandez-Alvarez et al., 2017; Nielsen et al., 2019), or a defined number of characters (Jurgens et al., 2018). Significant variation in con- textual lengths however for each citation makes considering fixed context window size less desir- able (Kunnath et al., …

Content truncated to 500 words