Co-citation linkages are commonly employed to measure implicit relationships between documents. However, traditional co-citation techniques equivalently treat the strengths of co-citation linkages within a single citing document, limiting technique effectiveness. This study explores whether context-aware citation network embeddings can be used to more accurately measure the strength of co-citation linkages within a single citing document. In particular, this study proposes a novel method that incorporates context-aware citation network embeddings into the co-citation approach and empirically evaluates its performance. Experimental results show that the proposed method outperforms a baseline based on traditional co-citation techniques. Thus, the proposed method can discriminate between weak and strong co-citation linkages.
In scientometrics or information retrieval, co-citation linkages are commonly used to measure
implicit relationships between documents. However, traditional co-citation techniques equivalently
treat the strengths of these linkages within a single citing document, presenting a notable
limitation. Although some techniques attempt to distinguish between weak and strong co-citation
linkages using surface structures like paragraphs [
In recent years, techniques that utilize embeddings, such as bidirectional encoder
representations from transformers (BERTs) [
[
Document X (citing document) -------[A].--------- [MASK] --------------------- masked document [MASK].---------- Document B is original cited document --------.---------------[C]--------------- Documents A, C and D ------------[D]. the other cited documents
This study employs this technique to measure the strength of co-citation linkages within a single citing document. The proposed method also predicts suitable documents, where the predicted documents are limited to other documents cited in the same citing document. As shown in Figure 2, the proposed method employs the context of the original citing document B and the cited documents (A, C, and D) to output a list of ranked documents that may be suitable as replacements for masked document B. This ranking also indicates the strength of the co-citation linkage with document B; the strongest linkage is between documents B and D.
Citation context in Doc. X -----[MASK]----------(Originally Doc. B is cited)
The other documents cited by Doc. X
Doc. A Doc. C Doc. D
Masked paper prediction by fine-tuned SciBERT Rank 1st Doc. D 2nd Doc. C 3ird Doc. A
Strong Weak
B B B
Co-citation linkage D between B and D
Co-citation linkage C between B and C
Co-citation linkage
A between B and A
comprises 16,669 citation contexts, each comprising the texts preceding and following the citation
position, the ID of the cited document, the title and abstract of the cited document, and the ID of
the citing document. For each context, the texts were split into sentences using spaCy, which was
loaded along with a biomedical text processing model (the en_core_sci_scibert model) [
document ranking. In this evaluation, a document was considered relevant if it shared one or more
indexing terms with the original cited document. As in a previous study [
The parameters used to fine-tune the SciBERT model were identical to those employed in a
previous study [
position was set to 125. The sum of the token word/document, position, and token-type embeddings was used as the input representation. To fine-tune the model, the experiment used 15,307 citation contexts obtained by excluding the 1,362 contexts used for the ranking tasks from the 16,669 citation contexts.
precision of the top-ranked document (P@1) and the average precision (AP). Moreover, a baseline method based on traditional cocitation techniques was used for comparison. In particular, this baseline method ranks documents randomly because traditional co-citation techniques treat the strength of co-citation linkages within a single citing document as equivalent.
accurately measuring co-citation strength using even a single citing document, the proposed method can enable the creation of co-citation maps with fewer citations. This applicability could be particularly valuable for mapping new knowledge domains where the number of citations is insufficient.