Translation Memory improvements

I just released another round of improvements to the translation memory server.

New ranking algorithm

The most important change is the matching algorithm. For the nitty-gritty details, the translation memory was previously using the BM25 algorithm. We now use another algorithm, which is a mix of BM25 with exact phrase matching. I’ve been using it for a few weeks now, and it gives much better results.

I also added the following tweaks:

Aberrations removal

The updated translation memory server removes aberrations. For instance, for a very short source string, you could get very long suggestions:

This is obviously not a good suggestion. The translation memory now reduces the weight to such suggestions and push the suggestion to the bottom of the list.

Pushing 100% matches to the top

The update translation memory server now improves the visibility of 100% Matches strings.

Here’s how a 100% Match was displayed before:

100% Match suggestions are now displayed more prominently and are separated from the other suggestions. You can still view the other suggestions by clicking on the “view xx suggestion” link.

I hope you will find these improvements useful. Thank you for using WebTranslateIt.