NeuroLit Navigator: A Neurosymbolic Approach to Scholarly Article Searches for Systematic Reviews

Why it matters

• Systematic review query development averages 8-15 hours, so the first search iteration is labor-intensive.
• Systematic review publications are growing at 26% per year, increasing demand for scalable and reproducible search workflows.
• General LLM tools struggle with domain-specific vocabulary, controlled terminology, and citation reliability.
• Keyword-based and LLM-only retrieval systems often trade precision for recall (or vice versa) and often lack reproducibility.
• The first iteration of article retrieval strongly influences downstream screening effort and overall systematic review quality.

What we did

• Introduced NeuroLit Navigator, a neurosymbolic system combining domain-specific LLMs with biomedical knowledge graphs (MeSH, UMLS).
• Used named entity recognition, knowledge graph expansion (up to two hops), and ClinicalBERT-based query expansion.
• Applied semantic re-ranking with MPNet embeddings, which achieved 36% relevance, the highest among tested embedding choices.
• Retrieved and presented the top k = 5 articles for structured librarian feedback.
• Demonstrated a 90% reduction in initial search time while supporting controlled vocabulary, interpretability, and reproducibility.

How it works

• User input + sentinel article: extract key biomedical entities via SciSpacy NER.
• Vocabulary extension: expand entities using MeSH/UMLS terms within two hops, filtered by semantic similarity.
• Query expansion: mask and substitute terms using ClinicalBERT to generate related variants.
• Iterative query refinement: start specific and progressively relax constraints until a minimum result set is retrieved.
• Retrieval + re-ranking: query PubMed via Entrez API and re-rank with MPNet embeddings, then return top k = 5 articles for librarian review.

The full neurosymbolic pipeline is illustrated in Figure 1.

Key contributions

• A neurosymbolic SR retrieval pipeline integrating controlled vocabulary and LLM-based expansion.
• A zero-shot query formulation approach that eliminates manual setup for first-iteration searches.
• Empirical comparison of biomedical embedding models showing 36% relevance with MPNet.
• Demonstrated 90% reduction in initial search time in librarian-facing deployment.
• Comparative evaluation showing NeuroLit Navigator uniquely combines relevance, reproducibility, interpretability, and controlled vocabulary support.

As shown in Table 1, the system uniquely combines reproducibility and controlled vocabulary with competitive relevance.