Introduction
The exponential growth of scholarly publications has created a significant hurdle for researchers attempting to stay current with scientific progress. This challenge is particularly acute in interdisciplinary research, where scientists from disparate fields often use different terminologies to describe overlapping concepts. Traditional search engines, such as PubMed or Google Scholar, primarily rely on exact keyword matching or basic semantic similarity. These methods often fail when a researcher lacks the specific domain jargon necessary to find relevant literature.
DiscoverPath is a novel, full-stack retrieval system designed to bridge these terminology gaps. By shifting the search paradigm from "keyword-matching" to "knowledge-exploration," DiscoverPath utilizes a knowledge graph (KG) to visualize connections between complex terminologies and recommend refined queries, enabling researchers to navigate unfamiliar domains effectively.
System Architecture
DiscoverPath is composed of three integrated modules that facilitate the extraction, recommendation, and visualization of biomedical knowledge.
1. Knowledge Graph (KG) Construction
The system processes paper abstracts from the PubMed database to build a structured graph. The construction involves two primary steps:
- Named Entity Recognition (NER): The system employs BERN, a pre-trained neural NER tool, to identify and extract biomedical entities (nodes).
- Relationship Extraction: Using NeuralCoref for part-of-speech (POS) tagging, the system extracts verbs and prepositions to define the relationships (edges) between entities.
In this graph, both biological terminologies and the source articles themselves are treated as nodes, creating a multi-layered network of information.
2. Query Recommendation System
To assist users in narrowing down broad searches, DiscoverPath includes a recommendation engine trained on the article-to-terminology network. The model learns high-dimensional embeddings for both articles and terms. It optimizes a joint likelihood function designed to pull an article closer to its relevant terminology in the embedding space while pushing it away from irrelevant terms. This allows the system to suggest specific, literature-dense concepts based on a user's initial broad query.
3. Interdisciplinarity Visualization (GUI)
The frontend, built with React and Flask, interacts with a Neo4j graph database. To prevent the "information overload" common in large-scale knowledge graphs—often referred to as the "hairball" effect—the interface displays a focused subgraph. Users see only the queried entity and its immediate neighbors, providing a clear and manageable view of the local knowledge landscape.
Case Study: Alzheimer’s Disease
As a system demonstration, the authors evaluated DiscoverPath using a case study focused on Alzheimer’s disease. When a user enters the broad term "Alzheimer," the system does not merely return a list of papers. Instead, it generates a localized graph linking the disease to specific proteins (e.g., ApoE4), biological processes (e.g., Aβ clearance), and potential treatments (e.g., rTMS).
Simultaneously, the recommendation engine suggests narrower refinement queries such as "tau biomarkers," "amyloid beta," and "group of lysosome enhancing compounds." This iterative process allows a researcher—perhaps a chemist or a data scientist—to discover the specific biological jargon necessary to conduct a deep dive into the literature without prior expertise in Alzheimer's pathology.
Discussion and Future Work
DiscoverPath demonstrates the viability of combining NER-driven KG construction with graph databases for real-time scientific discovery. By treating search as an iterative journey rather than a single-step query, the system supports the progressive refinement of research goals.
However, the authors identify several areas for future enhancement:
- Database Expansion: The current implementation is limited to biomedical abstracts; expanding to other scientific disciplines is a priority.
- LLM Integration: Future versions may leverage Large Language Models (LLMs) for more nuanced relationship extraction, moving beyond the limitations of standard POS tagging.
- Explainability: To ensure trust and safety, the authors plan to incorporate features that demonstrate the fairness and transparency of the recommended literature paths.
By helping researchers visually bridge terminology gaps, DiscoverPath serves as a critical tool for fostering interdisciplinary collaboration in an era of information explosion.
Summary updated