Unexpected translation of tagged text

[szhitansky (Sergey Zhitansky)]

Hello,

We are using Okapi Tikal tool for extraction of text from HTML files, also we are restoring HTML after translation by Tikal as well. So source for translation have tags and we have strange translation of these lines. Just a few examples:

SENT 1069: <g id="1"> Winaray </g>
PRED 1069: would the Commission be forced to insist on the situation in Western Germany ?
PRED SCORE: -43.64

SENT 1071: <g id="1"> ייִדיש </g>
PRED 1071: would the Commission be forced to insist on the situation in Western Germany ?
PRED SCORE: -43.64

SENT 1073: <g id="1"> Yorùbá </g>
PRED 1073: would the Commission be forced to insist on the situation in Western Germany ?
PRED SCORE: -43.64

Any thoughts how we can fix it? We have same results with -replace_unk and without it.

Thanks!

[guillaumekln (Guillaume Klein)]

Hi,

Does these tags appear in the training data?

More generally the model will behave unexpectedly if the input is far from the training data.

[jean.senellart]

see this other thread too:

If there is any way to keep placeholders as same as source when call NMT

Hi @liluhao1982 - what you get is what you trained your model to do. So if you want that your model preserve your placeholders, you do need to have some special training. What I would suggest is that you keep your placeholders as one single tokens, and do not have them with increasing count - otherwise your model will have to learn for each of them how to deal with them. For instance (or whatever). Then you need to make sure you have enough (possibly artificial) examples in your training corp…

let us know if this works for you!

[szhitansky (Sergey Zhitansky)]

I don’t know all placeholders so I can’t create specified model. Best case if model can return same non-translated word at same position. But as I know how network works - probably it’s not possible in direct way.

Now main questions - how to use Glossary with high priority than network translation and how to save non-translated words on their places without changes.

[guillaumekln (Guillaume Klein)]

how to save non-translated words on their places without changes

It all depends on the training data and what the model has learned.

Generally if a source word is out of vocabulary, it is likely that the network will translate it by an unknown word and then the -replace_unk feature might be able to copy it in the target. But remember that it is not doing a word by word translation so it might just drop it and the copy might also fail because the source word is unknown.

A more consistent approach is to apply the same logic as the named entities on the training data:

Using word features for named entities

Hi, That approach should work. Did you already run some experiment? I think a more robust approach would be to replace all entities by placeholders like __ent_person_lastname, __ent_organization, __ent_numeric_time, etc., use attention to align them with source placeholders and finally replace them back with their original value. However, that obviously means you have a system that produces these placeholders and some preprocessing and post-processing steps to replace them both ways. Anoth…

how to use Glossary with high priority than network translation

Personal thoughts: you could just build a radix tree with all the words you can translate and query it before each neural translation.