Abstract
Traditional keyword-based scientific search engines often fail to capture the structural and functional relationships between research papers. This study introduces a search system that represents scientific articles through three distinct discourse facets: Objective, Method, and Result. By transforming these facets into high-dimensional vectors and integrating citation context, the system facilitates complex queries, such as identifying papers with similar methodologies applied to different research goals. We demonstrate the effectiveness of this approach using the ACL Anthology corpus, providing tools for faceted ranking and visual research landscape exploration.
Introduction
The exponential growth of scientific literature has made efficient information retrieval increasingly difficult. While traditional search engines excel at keyword matching, they struggle with the structural nuances of scientific discourse. A researcher might seek a specific methodology used in a different field or want to understand the specific reason one paper cites another. Current citation indices show connectivity but lack semantic depth. To address this, we propose a "facet-aware" search system that treats papers as multi-dimensional entities rather than monolithic text blocks.
Methodology
The system employs a multi-stage pipeline to convert raw document data into a searchable, faceted database.
Facet Extraction and Vector Representation
The core of the system is the identification of three primary discourse facets: Objective, Method, and Result. Section titles are classified using National Library of Medicine (NLM) Category Mappings. For ambiguous titles, a fastText classifier trained on labeled abstracts is utilized. Each facet is then mapped to a 100-dimensional vector, resulting in a 300-dimensional representation for each article.
Citation Graph Augmentation
Beyond textual content, the system leverages the context of citations. A classifier determines whether a citation refers to the cited paper's objective, method, or result. Using the LINE (Large-scale Information Network Embedding) algorithm, these citation labels are used to refine the text-based vectors. For instance, if Paper A cites Paper B for its "Method," the corresponding "Method" vectors for both papers are adjusted to be closer in the vector space.
Datasets
The system was developed and evaluated using the ACL Anthology, comprising 20,796 articles and over 300,000 citation links. Initial word vectors were learned via fastText on a combination of English Wikipedia and the ACL corpus. Canonicalization of section titles was supported by 3,032 NLM Category Mapping rules. Training for the citation facet classifier utilized 1,618 contexts from the Citation Function Corpus and CL-SciSumm-2017.
Key Insights and Results
By treating articles as bundles of distinct contributions, the system enables "asymmetric" similarity searches. Users can anchor their search on one paper and rank others based on similarity in a specific facet, such as finding papers with similar results regardless of the methodology used.
The implemented browser-based demonstration offers:
- Faceted Ranking: Ranking results based on specific discourse components (e.g., "Show me papers with a similar Result to this one").
- Color-Coded Citation Graphs: Visualizing the functional nature of links between papers, allowing users to see how ideas have been utilized.
- Facet-Axis Scatter Plots: Using t-SNE to project high-dimensional vectors onto 2D axes (e.g., Method similarity vs. Objective similarity) to identify research clusters.
Conclusion
This faceted approach moves scientific search toward a more functional understanding of the research landscape. While current evaluation is challenging due to the lack of a "gold standard" for faceted similarity, the system provides a powerful tool for non-linear literature review. Future work will explore more granular facets and deeper NLP analysis of section content to further refine vector accuracy.
Summary updated