High-quality scientific article embeddings are essential for tasks like document retrieval, citation recommendation, and classification. Traditional citation-based approaches assume citations reflect semantic similarity—an assumption that introduces bias and noise. Recent models like SciNCL and SPECTER2 have attempted to refine citation-based representations but still struggle with noisy citation edges and fail to fully leverage textual information. To address these limitations, we propose a hybrid approach that combines Finding-Citation Graphs (FCG) with contrastive learning. Our method improves triplet selection by filtering out less important citations and incorporating finding similarity relations, leading to better semantic relationship capture. Evaluated on the SciRepEval benchmark, our approach consistently outperforms citation-only baselines, showing the value of text-based semantic structures. While we do not surpass state-of-the-art models in most tasks, our results reveal the limitations of purely citation-based embeddings and suggest paths for improvement through enhanced semantic integration and domain-specific adaptations.

Citations typically mention findings as well as papers. To model this richer notion of citation, we introduce a richer form of citation graph with nodes for both academic papers and their findings: the finding-citation graph (FCG). We also present a new pipeline to construct such a graph, which includes a finding identification module and a citation sentence extraction module. From each paper, it extracts rich basic information, abstract, and structured full text first. The abstract and vital sections, such as the results and discussion, are input into the finding identification module. This module identifies multiple findings from a paper, achieving an 80% accuracy in multiple findings evaluation. The full text is input into the citation sentence extraction module to identify inline citation sentences and citation markers, achieving 97.7% accuracy. Then, the graph is constructed using the outputs from the two modules mentioned above. We used the Europe PMC to build such a graph using the pipeline, resulting in a graph with 14.25 million nodes and 76 million edges.