📄 Methods for Matching English Language Addresses

Summary

The paper looks at deciding whether two English address strings point to the same place. Ordinary string similarity struggles here because each token's position and role carry meaning. A one-character edit can be harmless, like an abbreviation, or can change the location entirely, like a different house number. The authors define a match as two addresses referring to the same building, a choice driven by financial-services use cases. They then build a synthetic generator of matching and non-matching pairs from US-style addresses. Matches come from realistic edits such as abbreviation swaps and dropped words. Non-matches include harder cases where property or road numbers are altered. On this data they compare a range of distance measures, including Jaccard, Levenshtein, Jaro-Winkler and TF-IDF, against an ESIM model with GloVe embeddings, with and without character embeddings. The ESIM variants reach about 0.94–0.95 test accuracy and stay balanced between precision and recall. Many distance baselines achieve high precision only at the cost of poor recall. The work matters for entity resolution and record linkage, where addresses are common and there are few English benchmarks.

Key points

• The dataset starts from 10,000 base addresses of the form building, street, city, state:
  • building numbers and street names come from the Faker library;
  • cities come from the top 1000 US cities by population (a plotly dataset), with city–state pairings kept correct.
• Each address gets a prefix of type Floor, Apartment or Name:
  • Apartment uses APT/APARTMENT/SUITE/STE/UNIT;
  • Floor uses FLOOR/LEVEL;
  • Name uses ATTN or C/O plus a Faker name.
• Matching pairs are built through word substitution within equivalence groups (APT/APARTMENT, SUITE/STE, ROAD/RD, STREET/ST, AVE/AVENUE) and through word deletion. Non-matches include tweaked property or road numbers, not just random other addresses.
• In Test 4, averaged over 5 random seeds:
  • ESIM scored 0.93 precision, 0.95 recall and 0.94 accuracy on test;
  • ESIM + Char scored 0.95 precision, 0.94 recall and 0.95 accuracy;
  • test time was about 1 s, and training took roughly 530–830 s.
• Among the distance-based methods, the best accuracy in that table is 0.88. Several of them combine about 0.98 precision with recall near 0.16–0.24.

Interesting ideas

• I1. Matches are produced by meaning-preserving edits, while non-matches are produced by tiny meaning-changing edits such as a different number, which yields hard negatives with controllable difficulty.
• I2. What counts as a match is fixed explicitly at a chosen granularity (building), and the methods are kept independent of that choice so the level can be moved to unit or district.
• I3. The paper shows that two pairs with equal edit distance (an altered house number versus an appended letter) can deserve opposite verdicts, so token role outweighs raw character distance.
• I4. Small curated equivalence groups of abbreviations (ST/STREET, STE/SUITE) do much of the work of turning surface variation into a matchable form.
• I5. Results report precision, recall, accuracy and runtime, with means and standard deviations over seeds, which shows that high-precision baselines can hide very low recall.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Evaluation with negative controls could pair each test sentence with a minimally altered version whose correct answer must change, which exposes readers that only pattern-match surface form.
• I2 (world): Identity of place entities in the world of a text could carry an explicit granularity setting, so "same location" can be answered at building, city or region level.
• I3 (frames): Coreference between place mentions should compare parsed components such as number, street and city slot by slot, rather than falling back to string similarity.
• I4 (tokens): The lemma/normalization layer could map address and title abbreviations to canonical lemmas, so later stages see one form while the original token stays recorded.
• I5 (learn): Evaluation reports could always show precision and recall separately, along with seed variance, so a cautious component that rarely fires is not mistaken for a good one.