A similarity score of 0.87 is not proof that two records describe the same person, business, or other entity. It is a cutoff applied to a particular score, produced by a particular method on particular data. Whether it is a useful cutoff depends on the task and on the cost of linking different entities versus missing a genuine match.
Why a high similarity score can still produce a wrong link
Semantic matching is useful for finding records that appear related. But resemblance and identity are different questions. Two records can share broad meaning, common attributes, or similar descriptions while disagreeing on details that distinguish the entity you need to identify.
As an Amazon Associate I earn from qualifying purchases.
For example, two organization records might use nearly identical business descriptions but list different legal identifiers or locations. That hypothetical illustrates why a similarity score should be checked against the fields that matter for identity; it is not a reported incident. The reverse can happen too: genuine matches may score lower when identifiers are incomplete, misspelled, or out of date.
The UK Government’s data-linkage quality guidance notes that errors can occur regardless of linkage method and depend in part on the quality and completeness of identifying data. A semantic score alone does not establish that the evidence separates one entity from another.
What the 0.87 threshold means—and does not mean
A threshold is a decision boundary: pairs on one side may be treated as links and pairs on the other as non-links or candidates for further review. The number has no fixed meaning without knowing how the score was produced, what data it was applied to, and how the resulting decision will be used. In particular, 0.87 is not inherently a probability, confidence level, or industry-standard identity test.
The Government Statistical Service puts the general issue plainly: “In all linkage methods, some choice must generally be made about an evidentiary threshold for classifying record pairs as links or not.” The appropriate balance depends on the requirements of the data and the intended analysis—not on a cutoff considered in isolation.
Rank #2
- 5 THEMED BOOKS & 400+ PUZZLES: Enjoy five spiral-bound books featuring nostalgic themes including Classic TV, the Good Ole Days, American Road Trips, and more. With 400+ puzzles, 10,000+ words to find, answer keys included, and two pencils in every set - you’ll have everything you need to start puzzling.
- EXTRA-LARGE PRINT & EASY TO READ: Large, easy-to-read letters, spacious grids, and clearly printed word lists help reduce eye strain so you can focus on the fun. Designed especially for adults, seniors, and anyone who enjoys brain games and relaxing activities.
- LAY-FLAT SPIRAL BINDING: Unlike ordinary paperback word find books, each book opens completely flat and stays that way. Whether you’re at home, traveling, or relaxing in your favorite chair, every word search puzzle is easy to read, write in, and enjoy.
- SOLUTIONS INCLUDED: Every puzzle includes a clear, easy-to-read answer key in the back of the book, so help is always close at hand. Take your time, challenge yourself, and enjoy every puzzle without frustration.
- GIFT-READY 5-PIECE SET: Thoughtfully packaged and designed, this set makes a memorable gift for birthdays, Mother’s Day, Father’s Day, Christmas, and other special occasions. Proudly published by Bearwood Press, a veteran-owned small business based in the USA!
How false links and missed links differ
Two kinds of error matter, and they do not necessarily have the same consequences:
- False link: records for different entities are joined. This can happen when identifiers are shared or insufficiently distinctive.
- Missed link: records for the same entity are left separate. Recording mistakes, changes over time, or missing and weak identifiers can contribute.
Precision asks what proportion of assigned links are true. Recall asks what proportion of true matches were identified. A stricter threshold may reduce false positives while also excluding valid matches; how much it changes either measure depends on the system, data, and task.
Rank #3
- Large Print Word Search Books for Adults and Seniors: Pack of 4 Deluxe Easy-To-Read Word Find Puzzle Book.
- 4 books filled with stimulating word puzzles -- words cleverly hidden in every puzzle.
- Fascinating themes throughout.
- Cover art may vary. Over 380 pages of word find puzzles total.
- All new puzzles, all new words, new format and layout. Hours of mind-stimulating fun. Set also includes a word search bookmark and black pens.
The downstream use determines which trade-off is acceptable. Broad candidate discovery can tolerate more questionable pairs if people or later steps review them. A sensitive process that merges records automatically may need stronger evidence against false links. There is no context-free best threshold.
What published results say about thresholds
A 2026 study in Frontiers in Artificial Intelligence, “Detecting reconciliation discrepancies in tabular data using transformers,” reports results for its own semantic tabular-reconciliation method. Its evaluations illustrate why metrics must stay attached to the specific task and data rather than being treated as universal guidance.
| Evaluation | Reported result | What it applies to |
|---|---|---|
| Large-scale relationship identification | Experiments covered 185,909 tables; at τ=0.9, precision was 0.958, with F1 scores ranging from 0.77 to 0.87. | The study’s relationship-identification experiments, not an unspecified record-linkage system. Study abstract |
| Representative discrepancy-detection case | At τ=0.7, precision was 0.91, recall was 0.91, and F1 was 0.912. | The paper’s representative discrepancy-detection case, not the large-scale relationship-identification evaluation. Study evaluation |
| Same representative discrepancy-detection case | At τ=0.8, recall was 0.79 and F1 was 0.857. At τ=0.9, precision was 0.958 and recall was 0.676. | In this case, the stricter cutoff improved precision while reducing recall. These figures should not be merged with the separate relationship-identification results. Study evaluation |
These results do not validate 0.87 for another dataset, model, or purpose. The study itself shows that threshold outcomes are tied to the particular evaluation, and its reported cutoffs do not establish a generally correct operating point.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →How to evaluate a linkage threshold for your task
- Define the decision and its consequences. Specify what happens after a pair is linked. Candidate discovery, analyst review, and automatic creation of a merged record have different tolerances for error.
- Test representative labeled pairs. Evaluate pairs from the population and data conditions where the method will be used. Report precision and recall, and inspect examples of false links and missed links; a threshold without this context is not a quality result.
- Use evidence that distinguishes the entity. Keep candidate generation separate from final acceptance when the task calls for stronger proof. Depending on the data, combine semantic similarity with exact or otherwise discriminative fields.
- Retain uncertainty when a hard yes-or-no decision is premature. Preserve borderline links for review where appropriate. The UK guidance recommends retaining less-than-certain links and providing link-level measures so users can tune decisions and conduct sensitivity analysis.
- Check groups formed by chains of links. In transitive matching, a sequence of pairwise links can place records in one cluster even when the endpoints have not been directly compared or sufficiently supported. Review whether the resulting group is justified for your use case.
- Re-evaluate after changes. A cutoff may behave differently if the data, score construction, identifiers, or downstream use changes. Re-test rather than carrying a previously chosen number forward by habit.
What workflow rules can—and cannot—add
Similarity need not be the only evidence in a matching workflow. AWS documentation for its Advanced rule type describes combining exact and fuzzy matching conditions, as well as clustering; it also documents transitive matching as an API-only capability. These are examples of implementation options, not independent evidence that a resulting link is correct.
Best Value
When assessing any workflow, ask what evidence each rule uses, whether identifiers are complete and distinctive in your data, how uncertain pairs are represented, and whether matches are grouped transitively. A tool’s supported features do not replace evaluation against representative examples and the intended decision.
When to revisit a threshold
Reconsider the cutoff when error costs change, records gain or lose identifying fields, source data becomes less complete, the matching method changes, or links begin feeding a different process. The same numeric threshold can produce a different balance of false links and missed links under those conditions. Treat it as a task-specific operating choice, not a permanent property of semantic similarity.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
