Using Weblate Translation Memory and Fuzzy Matching to Cut Redundant Work

The Problem: Repeating the Same Translation

When a localization team works on a large product, the same UI label often appears in many places. Without a way to remember earlier translations, translators may render the same phrase differently, causing inconsistency and extra work.

How Weblate’s Translation Memory Works

Weblate stores each approved source‑target pair in a translation memory (TM) that lives in its PostgreSQL database. When a new source string is opened for translation, the system searches the TM for similar entries and returns a list with a similarity percentage. A 100 % match means the string is identical; a fuzzy match (for example, around 70 % similarity) shows a sentence that shares most words but may need tweaking.

Configuring Fuzzy Matching Thresholds

Administrators can adjust the minimum similarity that triggers a suggestion. In the Admin UI under Translation memory settings, the Minimum Similarity field controls how strict the match must be. A common starting point is 70 % for fuzzy matches and 85 % for high‑quality suggestions. Lower values increase noise; higher values miss useful variants. Other toggles let you enable automatic pre‑translation and decide whether the TM is shared across components or kept per‑component.

Worked Example: Verifying a Fuzzy Match

1. Make sure the TM is enabled for the component you are translating.
2. Translate and approve the string Edit profile.
3. Add a new source string Edit user settings to the same component.
4. Open the translation editor for the new string; the TM panel should list a suggestion based on Edit profile with a similarity score (for example, around 75 %).
5. Accept the suggestion, adjust it to Edit user settings, and save.
6. Add another string Edit account settings and check that the editor now offers a suggestion derived from the previously saved Edit user settings translation.

Limitations and Maintenance

Large TMs can bloat the database and slow queries, especially when many components share a global memory without pruning. Stale or incorrect entries propagate errors, so periodic review via the Admin → Translation memory page is recommended. Fuzzy matching relies on surface similarity; for agglutinative or highly inflected languages a high score may still produce grammatically awkward suggestions, requiring more manual editing.

Actionable Steps for Teams

Enable the TM for each component, set the minimum similarity to 70 % (adjust after observing suggestion quality), export a TMX backup before major changes, and schedule a quarterly cleanup to delete low‑confidence or outdated entries. This keeps the memory useful while limiting storage growth and maintaining translation quality.