Training Lilt

Why does Lilt not always learn specific translations immediately?

Lilt's unique machine translation system combines large statistical models trained on millions of sentence pairs based on publicly available background data with your personal translation memories and confirmed translated segments in real time. Sometimes, the evidence of a specific translation coming from the background data outweighs your personal preference, and the system will continue to use the background translation. However, over time your preferred translations should be picked up. As an example, consider the English term bank. The background data contains thousands of translations into banco in Spanish, but if you are translating a document on river erosion with the river bank vanishing over time, the system might not immediately be able to learn to prefer your choice of orilla (river bank in English) over banco. It may take multiple segments where you have to correct banco into orilla before the system is able to learn your translation as the one to be suggested in future segments of that document.

Alternatively, you can upload a termbase with the corresponding translations, or add a new termbase from scratch to the memory that is associated with your project. Go to Memories, click on the "Edit" button of the targeted memory (here English > Spanish), and click on "Add a new termbase". You can then edit the empty termbase (terminology.csv) and add the desired translations (e.g. bank and orilla).

Why is there a permanent spinning wheel? I can't confirm my segment!

In very rare cases, translation requests might get stuck (e.g. if there are a lot of users accessing the system). This is indicated by a spinning wheel below the segment number which blocks confirmation of your segment. You can use the filter to only show unconfirmed segments (Show Unconfirmed on the top right of the editor). It is advised to confirm all unconfirmed segments prior to the one with the spinning wheel. This will usually resolve the issue, and you can then confirm the current segment.

Still need help? Get in touch!
Last updated on 19th April 2018