We use the detection api to decide whether the language needs to be translated. However the 'detect' api returns inaccurate results at times. Eg. 'hi siri', 'hi google' are detected as 'ca' instead of 'en'. This will then be translated to 'there are siri', which is totally unacceptable in our case.

Currently there is no way for user to correct the detection or put more weight/preference on a particular language. In the above example, although 'en' is in the alternative, they all have the same score of 1. We'd like to add some weight on a particular language to control the 'detect' accuracy. For example, we'd like to see 'hi siri' returns 'en' with score 1 but 'ca' with score 0.998 etc.

Another way is to allow user to customize/improve the English dictionary. Eg. we want to add a few terms/phrases in the dictionary so that those words are NOT detected as other language than 'en'.

Typically, verbs and nouns are mixed up, resulting in hilarious, yet totally unusable (business-wise) translations.

It seems that the translation engine lacks any kind grammar knowledge - it doesn't even try to make up sentences of what it translates.

While this technique might play out with germanic<->germanic languages, it certainly does not with Finnish language.. It might be unfair to compare it to Google Translate (which has no problem translating even complex sentences), but that's the consumer standard so ofc people are comparing to it. As a matter of fact, even Microsoft's translation built-in-feature in Twitter does way better job!

Are there any plans on improving this feature?

Below are minor UX improvements to boost user experience (some are nice to have):

1. Ability to choose a workspace as default (when ever you open the portal), this is helpful when you are member of multiple workspaces.

2. Add clipboard icon next to category id for easy copy/paste. At the moment, entire line of category id has anchor link (end up clicking it multiple times during id selection).

3. Send a welcome/invitation mail when someone adds you as a workspace owner.

4. In the projects page, when you Ctrl+click on any project - allow opening it in new window/tab (as it has unique url) for multi tasking.

5. Sticky filters - In project page, when a filter is applied, clicked on one of the project and when you navigate back to the projects page - allow fiter condition is still available until cleared (for large number porjects, it is getting tedious to filter again and choose the projects).

6. Allow user can change the number of items visible in the grid (page size), currently it is 10. When you have large number of projects, it is tedious to browse through multiple pages or redefining the filter each time to scope down.

Is it possible to make some sort of sentence-level translation or confidence score available via the api (similar to the language detection confidence score) along the lines of:

'translations': [{'text': 'Hello, World!', 'to': 'en', 'quality_score': 0.98}]

We want to train our humans to write in easily translatable sentences. We also want to work through our historic library of old webpages and make them easier for machines to translate by making complex or poorly written sentences clearer.

In both cases this probably means sentences that aren't too long, don't contain colloquialisms and have reasonable grammar and structure. Ideally, every time someone posts a new web page we would then check it sentence by sentence and assess how good it is and highlight sentences that are hard to translate by machine.

We can create an API call to translate the entire website but we have no way to assess/audit the quality of the translations and so highlight problem words, sentences or pages for re-writing.

Right now the system supports html and plain text. And you can use document translator that "supports" SRT files. But what it is doing is breaking the file up into pieces and then putting it back together.

This is highly problematic for this very standard use case because in caption files (SRT, WebVVT and others) a sentence can easily be broken across multiple lines and multiple time markers resulting in bad translations because context is missed.

As an example the following snippet will result in a bad translation in Spanish:

00:00:49.800 --> 00:00:55.100
display in the Workspace menu, the order
in which those workspaces appear in the

00:00:55.100 --> 00:01:00.200
list, and whether a separator line
is added between workspace names.

Because "list" that starts the next time marker doesn't have context and as a result it gets dropped from the translation.

Currently there is no way to work around this for several reasons:

1. If you send the entire SRT or WebVVT file the time markers get translated too instead of being left alone.
2. The captions are almost always too long for the 5000 character limit.

I would suggest adding another content type of "captions" and have your translator recognize SRT and WebVVT at the least and allow sending the entire file as text at once.

That way you can read across the time markers and put the sentences together properly for translation without interfering with the time markers in the process.

This would result in vastly better translations for captions which is one of the goto demos of the Microsoft Translation system.