Answers are grounded in this paper and cite their sources.
Knowledge-intensive question answering often requires combining facts that are indexed separately, and top-ranked snippets may not form a usable evidence chain. This paper presents a low-cost graph-guided two-hop retrieval method for questions about scientific publications. Authors, papers, venues, and publication years are represented as typed nodes; an author anchor restricts candidate papers, topical similarity ranks them, and a second graph hop returns the requested property together with its evidence path. An evaluation is constructed from 43 individually checked ACL Anthology publication records, yielding 175 nodes, 210 directed relations, and 167 separately indexed atomic facts. We deterministically create 688 two-hop metadata questions using two lexical cue regimes and four question patterns. The graph-guided method achieves 99.42% answer exact match and correct-path recovery; TF-IDF and BM25 over two atomic facts obtain 36.63% and 26.60%, respectively. Increasing a flat TF-IDF budget to four facts yields 55.23%, while a denormalized full-record TF-IDF system obtains 98.55%. A matched-slot BM25 full-record baseline also reaches 100.00%, showing that graph traversal is particularly relevant to fragmented evidence rather than inherently superior to combined records. The entire evaluation runs on a CPU without neural-model training.
No takes yet. Share an insight, caveat, or question.
Campbell et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: