Abstract
Research papers reflect scientific advances. Citations are widely used in research publications to support the new findings and show their benefits, while also regulating the information flow to make the contents clearer for the audience. A citation in a research article refers to the information's source, but not the specific text span from that source article. In biomedical research articles, this task is challenging as the same chemical or biological component can be represented in multiple ways in different papers from various domains. This paper suggests a mechanism for linking citing sentences in a publication with cited sentences in referenced sources. The framework presented here pairs the citing sentence with all of the sentences in the reference text, and then tries to retrieve the semantically equivalent pairs. These semantically related sentences from the reference paper are chosen as the cited statements. This effort involves designing a citation linkage framework utilizing sequential and tree-structured siamese deep learning models. This paper also provides a method to create an automatically generated corpus for such a task.
Introduction
In scientific writing, citations serve as the essential links between new findings and established knowledge. However, a citation typically points to an entire document, leaving readers to manually search for the specific method, result, or claim being referenced. This challenge is known as the citation linkage problem.
In the biomedical domain, this task is particularly difficult due to the complex nature of scientific nomenclature—where a single compound like carbon dioxide might be represented as "$CO_2$" or via its chemical structure—and the common practice of paraphrasing. Previous research was limited by a "data desert," as existing human-annotated datasets were too small and imbalanced to effectively train modern deep learning models.
Methodology
Automatic Corpus Creation
To address the scarcity of labeled data, the researchers developed a pipeline to generate a large-scale "silver standard" dataset. By processing over 28,000 research documents from PubMed, they paired citing sentences with potential source sentences. Using a Sent2Vec model and a cosine similarity threshold of 0.57, the framework automatically labeled 74,568 balanced sentence pairs.
Advanced Encoders and the BioCite Ensemble
BioCite treats citation linkage as a semantic similarity task using a Siamese network architecture. The framework evaluates two primary types of encoders to capture the nuances of biomedical language:
- Sequential Encoders: Standard models like Bi-LSTMs and CNNs that process text linearly.
- Tree-Structured Encoders: Dependency and Constituency Tree-Transformers that analyze the grammatical structure of a sentence. These models are better equipped to understand that complex biological terms function as single units, even when paraphrased.
The final BioCite framework is an ensemble that combines word-level dependency and phrase-level constituency information to maximize retrieval precision.
Results and Discussion
The BioCite ensemble established a new state-of-the-art benchmark for citation linkage. When tested against the "Gold Standard" human-annotated dataset, the model achieved a balanced accuracy (BAcc) of 95.17%, significantly outperforming previous methods.
Key insights from the evaluation include:
- Structure Over Sequence: Tree-structured models outperformed sequential ones by better handling the multi-word names of biological components, reducing false positives by nearly 20%.
- Validation of "Silver" Data: The study demonstrated that a large, high-quality automatically generated dataset is superior for training deep models compared to tiny human-curated sets. Expert validation of the automated labeling showed a high Cohen’s $\kappa$ of 0.93–0.95.
- Semantic Gap Resolution: The model showed a sophisticated ability to link related concepts, such as connecting a citing sentence about "acidic environments" to a source sentence discussing "hydrolysis at pH4."
Conclusion
BioCite represents a significant advancement in automated citation linkage for biomedical literature. By leveraging structural sentence analysis and a novel automated data generation strategy, the framework overcomes traditional data limitations. Future work will explore incorporating broader citation contexts and utilizing domain-specific pre-trained models like SciBERT or BioBERT to further enhance precision.
Summary updated