In Paranyushkin 2019, weakly-connected community pairs — not the modularity score — are the knowledge-gap signal; the M value itself measures discourse bias

The compressed claim "modularity scores can serve as a knowledge-gap signal" (the form it took in question-gap-detection's source trail) is a downstream elision of a two-role structure in the primary, "InfraNodus: Generating Insight Using Text Network Analysis" (WWW '19) — read directly:

Role 1 — partitioning. "We then apply community detection algorithm [37], [38] based on modularity. This is an iterative algorithm that detects the groups of nodes that are more densely connected together than with the rest of the network." (§2.5; [37]–[38] are Fortunato and the Louvain lineage; the measure itself descends from Newman & Girvan 2004, PRE 69 026113.)

Role 2 — the gap signal. "Graph visualization is also used to identify the structural gaps in the graph … automatically by the software's algorithm (detecting the distinct communities that are not well connected). It has been demonstrated [40], [41] that structural gaps are the parts of a graph that indicate the potential for new ideas." (§2.8)

The refinement direct reading adds: the modularity value M is not unused — the paper bands it into a discourse-structure/bias index (Dispersed M>0.65, Diversified 0.6≥M>0.4, Focused 0.4>M≥0.2, Biased M<0.2). So M measures how clustered the discourse is; the gaps are specific weakly-connected cluster pairs. Two different signals, one measure upstream of both.

No single paper states the compressed claim verbatim (systematic search, 2026-07-01 capture) — Paranyushkin 2019 is the peer-reviewed anchor for the practice. Cluster: moc-knowledge-gap-detection, claim-query-failure-clustering-as-gap-signal, claim-obligatory-attributes-as-gap-signal, claim-grade-gradient-rank-gap-detection.