CafeTran

> I am pretty sure

I would be quite unsure when guessing user habits. As you see every day, they can be quite different.

> would produce lots of false QA grammar errors

Well, CheckMate, another, but external resource, always renders a reasonable and rather correct number of errors with LT enabled (they are limited to grammar, not spelling). I cannot reproduce this statement.

Out of curiosity, I tried LT and—at least in may target language (Italian)—I found it to be an impressive way to slow down the translation process given the number of false errors it outputs. A junior high school student would do a better work, me think.

Hm. 

> > always renders...

>Well, then you must be right.

This is not the point. It seems there is a lack of experience with CheckMate to see how useful it can be in combination with LT, only using its API and a LT server instance (also a question of privacy and NDAs - using the web interface vs. using a local server instance). Why then isn't this possible with CT?

@mario.cerutti:

Indeed LT can perform differently well in different languages. And: How did you use it? There are several ways to implement it.

tre: Indeed LT can perform differently well in different languages. And: How did you use it? There are several ways to implement it.

I didn't know that there are several ways to implement it; I will have a new look at it tomorrow. I just opened the resource as offered by CT and tried the non-premium version with a well written, rather long Italian sentence, and it "found" three or four places that needed improvement such as article repetition, which the so-so Office proofing tool correctly didn't spot. Therefore, for what I saw today, it's checking mechamism seemed a bit too simplistic to me. But, again, I will be more precise tomorrow. 

LT is pretty good for Polish. It all depends what rules are entered for your target language. At least for now, and for a number of years before now, Polish rules have had a pretty good maintainer (not me :-) )

> I didn't know that there are several ways to implement it

Just to avoid misunderstandings.

• OpenOffice Plug in (local)
  
• Chrome Plug in (online)
  
• OS plug in (online)
  
• API with local server running (local)
  

If I am not mistaken, these things behave differently. And there is not much sense in using the LT spellcheck if Hunspell + user dic is already being used.

Just to get an impression. This is a typical CheckMate output with LT enabled:

I tried again and can only confirm that for Italian LT is almost useless (almost on par with the Word proofing tool). Sometimes it just gives vague (and false) indications on verb tense incompatibilities or article redundancy, areas where LT cannot be trained for obvious reasons, showing that it's not very intelligent and it's of no serious use for professional tasks. 

Therefore, certainly I wouldn't want to use it in an automated QA routine within CTE because of the substantial number of false alarms it would output. Maybe for other languages it's more useful, and I wonder if there are other Italian translators on this forum who can confirm or question my opinion.

As a final proofing stage I prefer to export to a bilingual file and let a TTS software read the text aloud. This is very effective to spot problems that one might otherwise end up skipping when only reading. I wish one day CTE had this capability!

I see. Look at the number of rules (grammar, not spelling):

Spanish: 136

Italian: 149

Polish: 1497

English: 2184

German: 2743

Catalan: 3109

Dutch: 3229

Maybe this gives a rough estimate for the state of the art, though this number of rules might depend on the complexity of a language (and maybe the capacity of the developers to combine them and make them easier). Catalonia might have a number of people hired for this (they do so, I am very sure). Dutch surprises me a bit (Hans, is this you?), as the language is so easy (only one minute to learn how to order a beer and to hitch a girl), but seriously, this input is really tremendous compared to EN. So it should be great in NL, but I never heard an enthusiastic feedback from Tilburg.

>but I never heard an enthusiastic feedback from Tilburg.

Yeah, I'm still totally puzzled by LT. We have a lingo that looks simple but actually is as complicated as the Chinese dialect of Kolobrzeg.

I will reiterate. It all depends on the rule set for your target language and obviously on the skill level on the rule maintainer. I mentioned LT in a FB forum for Polish proofreaders, and they had only words of praise for LT. Too bad it does not work so well for Italian.

Make no mistake, I had no intention to discredit LT at all. I just tried it and gave my opinion for Italian only. Italian has a very complex grammar and a huge number of exceptions, which makes it difficult, if not impossible, to create a reliable and effective proofing tool. 

The fact is, adding another layer of time-consuming check and receiving in turn a considerable number of false positives is not going to be of much help, on the contrary. My own brain, the spell checker and, again—sorry if this might appear a fixation, but it is—a final text-to-speech check, are faster and a way more effective, at least in my experience.

> I had no intention to discredit LT at all. 

No problem. I stumbled upon the numbers below only by curiosity, and indeed, Italian seems to be one of the worst cases (besides Swedish - 46 rules only - while Portuguese and Catalan have > 2000 rules) 

> a final text-to-speech check, are faster and a way more effective, at least in my experience.

From my rudimentary knowledge, this might be difficult to implement, if there is no Java solution (I think we discussed this before). The TTS function in Word (newest version, 365, on Mac) gives - let's say - for German nearly satisfying results. Did you ever test this again in the last months?