When the Gazetteer Goes Blank, the Records Still Know Where the Place Is

TL;DR for operators

When a historical place name is missing from a gazetteer, the usual workflow treats it as a lookup failure. This paper shows another option: use multiple records that mention the same place from already known coordinates, convert descriptions such as “1 km east of X” into constraints on where X must be, and aggregate those clues.

Fernando et al.¹ test this idea on 365 gazetteered New Zealand point places whose coordinates are deliberately withheld. A probabilistic model achieves the strongest fine-scale accuracy: 1.43 km median error and 36.0% within 1 km, while returning an estimate for every target. A deterministic hard-constraint approach fails to produce any estimate for 108 targets. GPT-5.1 also returns estimates for all targets and achieves lower mean error, but its median and 1 km accuracy are worse than the probabilistic model.

For archives, biodiversity platforms, gazetteer providers, and geocoding systems, the practical design is therefore not simply “rules or LLM.” The paper supports a pipeline in which flexible language interpretation and explicit probabilistic localisation perform different jobs. It does not yet establish the same accuracy for genuinely missing places, sparse evidence, non-point features, or regions outside New Zealand.

A missing database entry is not the same as missing evidence

Historical data systems frequently encounter names that no longer appear in authoritative geographic databases. The affected record may describe a farm station, trig point, school, recreation site, infrastructure facility, or other locally meaningful feature. Searching another gazetteer can help, but eventually the lookup chain ends.

The Allan Herbarium records examined here contain another kind of information. The same unknown place can appear repeatedly in specimen locality descriptions, while each specimen already has its own collection coordinates. One record might say that a specimen was collected south of the place; another might place the collection point northeast of it; a third might give a distance.

That changes the computational problem. The system is no longer asking only, “Which database entry matches this name?” It can ask, “Which location is most consistent with the distributed spatial evidence attached to this name?”

The paper first establishes that this is not a purely synthetic problem. After screening against the New Zealand Geographic Board Gazetteer, GeoNames, and OpenStreetMap, the authors retain 520 candidate names. Manual inspection of a random 50 confirms 30 as genuinely absent from those gazetteers. Extrapolating that 60% confirmation rate implies roughly 312 true non-gazetteered mentions, about 2.3% of the 13,746 processed locality descriptions.

That estimate is descriptive rather than a complete census, but it shows that unresolved names occur often enough to justify a recovery workflow.

Multiple records turn language into spatial constraints

The mechanism is easiest to see by reading a locality description backward.

Suppose a specimen was collected at a known coordinate described as “1 km east of Kauangaroa.” If the specimen lies one kilometre east of Kauangaroa, then Kauangaroa should lie roughly one kilometre west of the specimen coordinate. The paper calls this inversion of the spatial relation.

A single clue usually leaves considerable uncertainty. Repeated mentions reduce it. Several specimen coordinates surrounding the same target can jointly narrow the plausible location.

The deterministic approach represents each clue as a feasible geometric region and intersects those regions into an Approximate Location Region. This is interpretable, but exact intersection gives every clue veto power. When archival descriptions are approximate or inconsistent, the intersection can disappear.

That failure is visible in the main benchmark. Of 365 pseudo-NGP targets, the deterministic model returns no estimate for 108.

The probabilistic model replaces those hard boundaries with likelihoods. A proximity relation, for example, progressively penalises candidate locations as they move farther from an anchor:

Directional and stated-distance relations receive their own uncertainty terms. Instead of requiring all descriptions to agree exactly, the model searches for the location with the strongest aggregate support.

This is not merely a different mathematical implementation. It changes how the system treats contradictory evidence: disagreement reduces confidence rather than automatically eliminating the candidate location.

Soft constraints win where fine-scale precision matters

The headline comparison needs more than one metric because the methods fail differently.

If the operational objective is typical fine-scale localisation, the probabilistic model is strongest: it has the lowest median error and the highest share of estimates within 1 km. The GPT variants have lower mean error, which indicates that their error distribution contains fewer or less severe large misses.

Those are different advantages. The result does not support saying either that GPT-5.1 dominates explicit modelling or that the LLM experiment failed.

The relation-stratified results make the distinction sharper. With proximity-only evidence, the probabilistic model reaches a 1.09 km median error. With direction plus distance, its median falls to 0.74 km. But with direction alone, its median error rises to 5.25 km because bearing constrains an axis without telling the model how far along that axis the target lies.

GPT-5.1 behaves differently. The authors observe it extracting relations directly from raw locality text, reversing spatial relations, combining distance and direction, performing rough geographic calculations, and sometimes giving greater weight to mutually consistent clues. For direction-only targets, the LLM-based variant records a 1.41 km median error, substantially better than the probabilistic model’s 5.25 km.

The evidence therefore favours capability separation rather than a universal ranking.

LLM interpretation and probabilistic localisation solve different problems

An operational georeferencing pipeline has at least two difficult decisions.

The first is linguistic: what does an irregular historical description actually say? This is where the GPT-5.1 experiment is attractive. It operates on raw descriptions rather than requiring every expression to fit a predefined relation vocabulary.

The second is statistical: after extracting several noisy clues, where should the system place the coordinate? On the benchmark’s principal fine-scale measures, explicit probabilistic aggregation handles this decision better and does so reproducibly under fixed parameters.

For a biodiversity archive, the resulting architecture could use an LLM to interpret heterogeneous locality text, normalise uncommon spatial expressions, or propose structured relations. A probabilistic layer could then combine those observations with known anchors and produce the coordinate estimate. Human review would remain appropriate where source coordinates conflict, evidence forms multiple spatial clusters, or the inferred location would materially affect downstream analysis.

That division of labour is a Cognaptus inference from the comparative results; the paper itself tests separate deterministic, probabilistic, and GPT-based approaches rather than reporting a deployed hybrid system.

The benchmark is controlled, but easier than the full archival problem

The quantitative evaluation uses pseudo-NGPs. These 365 targets are known NZGB point places whose coordinates are hidden during inference and then restored as ground truth for scoring. This is necessary for measuring error, but it means the experiment does not directly measure accuracy on the confirmed non-gazetteered places discovered earlier.

The targets also need at least three locality descriptions. Places mentioned once or twice—the cases most likely to be poorly constrained—are outside the benchmark. The study is limited to New Zealand, assumes point geometries, and depends on the accuracy and spatial coherence of specimen coordinates.

Two extreme errors, involving Fog Peak and Rockdale, appear to reflect disagreement between specimen evidence and NZGB ground truth itself. That is a useful warning for production systems: when multiple independent methods fail on the same record, investigating reference-data quality may be more productive than tuning another model.

Treat unresolved place names as evidence problems before discarding them

The paper’s strongest contribution is not a new fallback geocoder. It shows that records surrounding an unresolved name can contain enough spatial redundancy to reconstruct a plausible location even when no gazetteer supplies a candidate.

For organisations managing historical, ecological, heritage, or archival data, that expands the decision set. A missing lookup can trigger evidence aggregation rather than immediate rejection or manual search.

The comparative results also put a useful limit on enthusiasm for general-purpose LLM reasoning. GPT-5.1 handles raw spatial language flexibly and avoids some severe errors, but explicit probabilistic modelling remains more precise on the benchmark’s central fine-scale measures. Where both language heterogeneity and geographic precision matter, those capabilities are better treated as components with different responsibilities.

Cognaptus: Automate the Present, Incubate the Future.