Pages in topic:   < [1 2 3] >
Lilt: interactive stat MT for translators
Thread poster: Spence Green

Lilt Inc
United States
En<>It Feb 2, 2016

Domenico Trimboli wrote:

Hello Spence,

Your software seems interesting. Any plans for En>It?

Domenico


Yes! We plan to add EnIt in February.


 

FarkasAndras  Identity Verified
Local time: 16:33
English to Hungarian
+ ...
"Interactive MT system" Feb 2, 2016

Does that mean, "you get suggestions from MT while you translate, just like in several major CAT tools"?
To me, an interactive MT system would be one where the translator has some control over what the MT does.


 

Lilt Inc
United States
Re: Interactive MT Feb 2, 2016

FarkasAndras wrote:

Does that mean, "you get suggestions from MT while you translate, just like in several major CAT tools"?
To me, an interactive MT system would be one where the translator has some control over what the MT does.


Yes, that's precisely the definition of an interactive MT system (see link below). Lilt implements both predictive typing---the MT system responds to your partial translation---and model adaptation in which each segment you confirm is added to the training data.

Wikipedia: Interactive MT


 

Siegfried Armbruster  Identity Verified
Germany
Local time: 16:33
English to German
+ ...
In memoriam
What is the difference to Studio 2015 with AutoSuggest 2.0 Feb 2, 2016

Lilt Inc wrote:
Lilt implements both predictive typing---the MT system responds to your partial translation---and model adaptation in which each segment you confirm is added to the training data.


I am using SDL Studio 2015 with AutoSuggest 2.0. Studio allows me to have various TMs that can be ranked, it offers fuzzy matches that need little editing and it offers dynamic AutoSuggest based on my TMs and entries. In addition it can handle numerous file formats and all the data remain on my PC.

I like your idea and your approach, but why should I switch from a proven solution to your system, what is in it for me? And what is in it for you?


 

Michael Beijer  Identity Verified
United Kingdom
Local time: 15:33
Member (2009)
Dutch to English
+ ...
MT vs fuzzy matches Feb 2, 2016

Siegfried Armbruster wrote:

Lilt Inc wrote:
Lilt implements both predictive typing---the MT system responds to your partial translation---and model adaptation in which each segment you confirm is added to the training data.


I am using SDL Studio 2015 with AutoSuggest 2.0. Studio allows me to have various TMs that can be ranked, it offers fuzzy matches that need little editing and it offers dynamic AutoSuggest based on my TMs and entries. In addition it can handle numerous file formats and all the data remain on my PC.

I like your idea and your approach, but why should I switch from a proven solution to your system, what is in it for me? And what is in it for you?


The difference is that Lilt produces machine translated output, whereas Studio merely offers you fuzzy matches. As far as I know, Studio doesn't even offer auto assembly (like CafeTran, DVX, etc.).

[Edited at 2016-02-02 21:37 GMT]


 

Siegfried Armbruster  Identity Verified
Germany
Local time: 16:33
English to German
+ ...
In memoriam
Fuzzy matches? Feb 3, 2016

Michael Beijer wrote:
The difference is that Lilt produces machine translated output, whereas Studio merely offers you fuzzy matches. As far as I know, Studio doesn't even offer auto assembly (like CafeTran, DVX, etc.).

[Edited at 2016-02-02 21:37 GMT]


I am not sure when you used Studio 2015 for the last time, but it does not just offer fuzzy matches.
It offers various MT Plugins (soon including Slate, a MT system that runs on our PCs), AutoSuggest 2.0 that dynamically uses the content of our TMs, RegEx Match AutoSuggest, MyMemory plugin, SDLTmFinVars, Variables Manager etc..

These features help me to improve my productivity dramatically, and as I wrote before, there are other features, such as various file filters, previews etc that are important. In my opinion, a professional production system should offer more than just "predictive typing and model adaptation".

My questions remain, how will Lilt help me to improve my productivity and the quality of my output - and - what is in it for Lilt, what is their business model, how do they want to generate revenue, what is their road map?


 

Michael Beijer  Identity Verified
United Kingdom
Local time: 15:33
Member (2009)
Dutch to English
+ ...
no time to argue Feb 3, 2016

Siegfried Armbruster wrote:

Michael Beijer wrote:
The difference is that Lilt produces machine translated output, whereas Studio merely offers you fuzzy matches. As far as I know, Studio doesn't even offer auto assembly (like CafeTran, DVX, etc.).

[Edited at 2016-02-02 21:37 GMT]


I am not sure when you used Studio 2015 for the last time, but it does not just offer fuzzy matches.
It offers various MT Plugins (soon including Slate, a MT system that runs on our PCs), AutoSuggest 2.0 that dynamically uses the content of our TMs, RegEx Match AutoSuggest, MyMemory plugin, SDLTmFinVars, Variables Manager etc..

These features help me to improve my productivity dramatically, and as I wrote before, there are other features, such as various file filters, previews etc that are important. In my opinion, a professional production system should offer more than just "predictive typing and model adaptation".

My questions remain, how will Lilt help me to improve my productivity and the quality of my output - and - what is in it for Lilt, what is their business model, how do they want to generate revenue, what is their road map?


I don't have time to argue about this, particularly because it is irrelevant to me. Especially since I absolutely detest SDL Studio, and think that it is one of the most clunky, annoying programs I have ever used. But you seem to like it, so I wish you all the best.

For the record, I don't like Lilt either, which is just another stripped-down, online toy, if you ask me. Sorry Spence!


 

Siegfried Armbruster  Identity Verified
Germany
Local time: 16:33
English to German
+ ...
In memoriam
Fully agree Feb 3, 2016

Michael Beijer wrote:
I don't have time to argue about this, particularly because it is irrelevant to me. Especially since I absolutely detest SDL Studio, and think that it is one of the most clunky, annoying programs I have ever used. But you seem to like it, so I wish you all the best.

For the record, I don't like Lilt either, which is just another stripped-down, online toy, if you ask me. Sorry Spence!


I agree, there is no reason to start another discussion on which CAT is better or worse. Yes, I like Studio, I tried other tools, but they did not work for me. My point was not so much to promote Studio, but to find out what Lilt really offers and I am still hoping that we will get some answers.


 

Spence Green  Identity Verified
United States
Local time: 07:33
English to Arabic
+ ...
TOPIC STARTER
Re: your evaluation of Lilt Feb 4, 2016

Michael Beijer wrote:

For the record, I don't like Lilt either, which is just another stripped-down, online toy, if you ask me. Sorry Spence!


If you have specific comments about the interface, I'd be eager for you to share them here.

I'm not sure that you were able to evaluate translation quality. You didn't confirm any segments during your session, and you didn't upload a TMX file. Lilt is an adaptive system. But if you don't supply any data, then it can't adapt. Baseline translation quality is something like what you would get from Google, Microsoft, etc. The system adapts fairly quickly as you use or add data to it.

Lest you worry about privacy, note that I can see your account activity (create account, confirm segment, etc.) but your data is encrypted.

We also don't support EnNl right now. I hope that you'll try it again when we add support for that language pair in a few months.


 

Spence Green  Identity Verified
United States
Local time: 07:33
English to Arabic
+ ...
TOPIC STARTER
Re: productivity, pricing, etc. Feb 4, 2016

Siegfried Armbruster wrote:

I agree, there is no reason to start another discussion on which CAT is better or worse. Yes, I like Studio, I tried other tools, but they did not work for me. My point was not so much to promote Studio, but to find out what Lilt really offers and I am still hoping that we will get some answers.


Siegfried: Thank you for trying Lilt a few days ago. The research upon which Lilt is based aimed to increase translation throughput through both interface interventions and increased machine translation quality. It is well known that there is very high variance among individuals in terms of both translation time and quality. What works for you may not work for others and vice versa.

Since you did mention a preference for a variety of information sources, I would like to point out one phenomenon commonly observed in translation user studies. Translators tend to read what appears on the screen, and if there is a lot information, this can lead to higher cognitive load, for some definition of cognitive load (e.g., pupil dilation, or pause frequency/duration).

Lilt tries to show you the best suggestion in the current context. You can think of translation as a search problem. In search, you usually optimize the relevance of the top result, not so much the fifth or fiftieth.

But I'm an empiricist, so you should measure yourself. This week we added throughput to the interface. You can measure your throughput with 10-second resolution. So time yourself in Trados, and time yourself in Lilt, and see which one is better. I'd be eager to hear the result of that experiment.

As for pricing, Lilt is a venture-backed startup. It's in a free trial period. We'll start charging later in the year.


 

Michael Beijer  Identity Verified
United Kingdom
Local time: 15:33
Member (2009)
Dutch to English
+ ...
my thoughts Feb 4, 2016

Spence Green wrote:

Michael Beijer wrote:

For the record, I don't like Lilt either, which is just another stripped-down, online toy, if you ask me. Sorry Spence!


If you have specific comments about the interface, I'd be eager for you to share them here.

I'm not sure that you were able to evaluate translation quality. You didn't confirm any segments during your session, and you didn't upload a TMX file. Lilt is an adaptive system. But if you don't supply any data, then it can't adapt. Baseline translation quality is something like what you would get from Google, Microsoft, etc. The system adapts fairly quickly as you use or add data to it.

Lest you worry about privacy, note that I can see your account activity (create account, confirm segment, etc.) but your data is encrypted.

We also don't support EnNl right now. I hope that you'll try it again when we add support for that language pair in a few months.


Hi Spence,

Indeed, I never even really tried Lilt. I just logged in, had a look around, and left. My quite harsh criticism was therefore perhaps a bit premature, so let me put my misgivings into context.

I have been translating now for quite a long time, and until only very recently used exclusively grid-based CAT tools. However, fairly recently, and more or less by chance, I rediscovered the joys of translating in MS Word, using not Wordfast Classic or some such, but a little-known CAT tool (which is now open source) called Felix (http://felix-cat.com/ ). As far as I know, Felix is the only CAT tool on the planet that allows you to work inside MS Word, but does not introduce any kind of hidden codes or tags in your document and allows you to do all of your work right inside MS Word, that is, not inside of a little box inside MS Word (Wf Classic, LogiTerm, etc.). This has many advantages, which I am currently trying to summarise in a future blog post. Anyway, to cut a long story short, for any new CAT or MT tool to interest me, it simply has to work inside of MS Word.

I'm almost afraid to say it out loud here in public, but I'm actually a big fan of Google Translate (and a few other MT engines), and currently use them inside MS Word via GT4T (http://gt4t.net/en/ ), which I find very useful as a typing aid and source of ideas. I would therefore love to try out your new system, but as a plug-in/addon inside MS Word.

I was actually one of the first people to pledge money towards another new MT system, this one called Slate Desktop (https://www.indiegogo.com/projects/slate-desktop-the-personal-experience-for-pros#/ ). The developer sent me the release candidate a few days ago, but I haven't had time to install or test it yet. Back when I initially decided to buy Slate Desktop, I was still using CafeTran as my main CAT tool, and so at the time asked the developer of Slate Desktop to build a CafeTran plug-in. Recently, however, I changed my request, and asked him to build an MS Word plug-in instead, but apparently I'm the only translator who requested one. Everyone else requested plug-ins for grid-based CAT tools. Sadly, grid-based CAT tools seem to be much more fashionable these days, but I predict a come-back for MS Word based solutions in the not to distant future ...

OK, gotta get back to work, as I have a deadline to meet!

Michael

PS: I will definitely try out Lilt again once you add NL>EN!
PPS: Do you have any plans to create some kind of MS Word plug-in or Addon?

[This post was dictated using Dragon Professional 14. Please excuse any typos!]

[Edited at 2016-02-04 20:31 GMT]

[Edited at 2016-02-04 20:35 GMT]


 

Spence Green  Identity Verified
United States
Local time: 07:33
English to Arabic
+ ...
TOPIC STARTER
Free ProZ seminar about Lilt Feb 4, 2016

I'll be hosting a free ProZ seminar on February 10th at 17:00 GMT. I'll demonstrate Lilt and discuss the following:

* Lilt's productivity features
* Data transfer to/from other CAT tools
* "The cloud" and data security
* The Pro version of Lilt

Q&A to follow.

...
See more
I'll be hosting a free ProZ seminar on February 10th at 17:00 GMT. I'll demonstrate Lilt and discuss the following:

* Lilt's productivity features
* Data transfer to/from other CAT tools
* "The cloud" and data security
* The Pro version of Lilt

Q&A to follow.

Click here to register
Collapse


 

Siegfried Armbruster  Identity Verified
Germany
Local time: 16:33
English to German
+ ...
In memoriam
I am not backed by venture capital, I am just a translator Feb 5, 2016

Spence Green wrote:
The research upon which Lilt is based aimed to increase translation throughput through both interface interventions and increased machine translation quality. It is well known that there is very high variance among individuals in terms of both translation time and quality. What works for you may not work for others and vice versa.


Increasing translation throughput is a nice objective, but it should not be the only objective. There might be other aspects that are at least as important.


Since you did mention a preference for a variety of information sources,


I am not interested in the “best suggestion” in my specialty field I’m interested in the “correct suggestion”. In medical and in other fields, there exist a lot of regulatory documents that specify which translation must be used. It makes no sense to put these in MT system. Putting them in a simple TM that ranks higher in processing compared to the MT output is faster and results in better quality. This is just one reason why a “plain” SMT system is just not good enough as a production system in my specialty field.


I would like to point out one phenomenon commonly observed in translation user studies. Translators tend to read what appears on the screen, and if there is a lot information, this can lead to higher cognitive load, for some definition of cognitive load (e.g., pupil dilation, or pause frequency/duration).


I have no idea what you want to tell us with this statement. With a modern CAT tool or a Dictation environment, translators can achieve quite high outputs without getting overworked, and when looking at Lilt, it is missing major features that ease our work, so working with Lilt might be a stress factor by itself.


Lilt tries to show you the best suggestion in the current context. You can think of translation as a search problem. In search, you usually optimize the relevance of the top result, not so much the fifth or fiftieth.


What? Translators are not dictionaries and quality translations are no puzzle or a search problem. Specialized translators know their topics and don’t need to search a lot.


But I'm an empiricist, so you should measure yourself. This week we added throughput to the interface. You can measure your throughput with 10-second resolution. So time yourself in Trados, and time yourself in Lilt, and see which one is better. I'd be eager to hear the result of that experiment.


Ok, let us schedule a test, but let us define the task first. How about the following real translation projects:
a) an En-De PIL/SPC translation in DOC format that must comply with the EMA regulations
or
b) a user manual in Framemaker or InDesign format with just a few ISO standards
or
c) a PDF that needs to be translated preserving its layout

I am just a translator who needs to earn money, with real projects, Lilt is in my opinion still just an academic solution that can be used to get some money from investors.

And I forgot the following point:
Ok, you claim that my data are secure and remain my data, but what about the translation engine you generate with my data, is it yours or is it mine? High quality TMs to to train MT systems and high quality MT engines are valuable assets.



[Edited at 2016-02-05 18:13 GMT]


 

Spence Green  Identity Verified
United States
Local time: 07:33
English to Arabic
+ ...
TOPIC STARTER
Many topics! Feb 6, 2016

Siegfried Armbruster wrote:

Increasing translation throughput is a nice objective, but it should not be the only objective. There might be other aspects that are at least as important.


I should have mentioned that I mean increasing throughput while maintaining quality. Our research focused on these two variables: time and quality. If there are other considerations, it would be great if you could list them here.



I am not interested in the “best suggestion” in my specialty field I’m interested in the “correct suggestion”. In medical and in other fields, there exist a lot of regulatory documents that specify which translation must be used. It makes no sense to put these in MT system. Putting them in a simple TM that ranks higher in processing compared to the MT output is faster and results in better quality. This is just one reason why a “plain” SMT system is just not good enough as a production system in my specialty field.


Lilt has an integrated TM, and if there is a TM match, then that match is always shown in lieu of MT. Moreover, Lilt is adaptive, so the MT system learns the lexical and syntactic choices that you make in your field. It's not a general-purpose MT system like Google Translate.

As I said to Michael B. above, I'm not sure that you were able to evaluate translation quality. When you tried Lilt, you only confirmed four segments, and you didn't upload a TMX file. So you didn't see any TM matches, and the MT system couldn't adapt. If that is the extent of your evaluation, then I completely agree that your current CAT setup makes you more productive.


Lilt, it is missing major features that ease our work, so working with Lilt might be a stress factor by itself.


What are the major missing features? Tell me and we'll implement them.


Specialized translators know their topics and don’t need to search a lot.


What I mean by search is any problem setting in which there are multiple competing solutions. Basic arithmetic isn't a search problem because for any computation there is one right answer. But chess is a search problem. Picking out a shirt in the morning is a search problem. Translation is a search problem in that there are often many valid translations licensed by the source, and the translator's job is to select from these alternatives.

Incidentally, the difficulty of routinizing the translation problem is one reason why translators don't really have anything to fear from MT. The problem is too hard.


Ok, let us schedule a test, but let us define the task first. How about the following real translation projects:
a) an En-De PIL/SPC translation in DOC format that must comply with the EMA regulations
or
b) a user manual in Framemaker or InDesign format with just a few ISO standards
or
c) a PDF that needs to be translated preserving its layout


Great! Let's do (a). I'll even pay you for your time. Please send email to info@lilt.com and we'll arrange the test.


I am just a translator who needs to earn money, with real projects, Lilt is in my opinion still just an academic solution that can be used to get some money from investors.


I wish that investors could be fooled like you say, but in my experience they are usually very clever people


And I forgot the following point:
Ok, you claim that my data are secure and remain my data, but what about the translation engine you generate with my data, is it yours or is it mine? High quality TMs to to train MT systems and high quality MT engines are valuable assets.


Our privacy policy and the tutorial in the application very clearly indicate the status of your data. It's private. It isn't shared with other users or third parties. It is encrypted in a commercial datastore. And if you delete your account, the associated parallel data, trained MT models, and TMs all disappear as well.


 

Spence Green  Identity Verified
United States
Local time: 07:33
English to Arabic
+ ...
TOPIC STARTER
Many topics pt.2! Feb 6, 2016


As far as I know, Felix is the only CAT tool on the planet that allows you to work inside MS Word, but does not introduce any kind of hidden codes or tags in your document and allows you to do all of your work right inside MS Word, that is, not inside of a little box inside MS Word (Wf Classic, LogiTerm, etc.). This has many advantages, which I am currently trying to summarise in a future blog post. Anyway, to cut a long story short, for any new CAT or MT tool to interest me, it simply has to work inside of MS Word.


Metatexis is a Word plugin, too. Text editors are cool. Our environment is actually designed after a text editor, well really, after Martin Kay's 1980 conception of a translation tool, which was simply a text editor with some bilingual extensions. The "cards" in our interface will eventually be replaced by a single-pane editor, but that's kind of hard to implement in the browser.

But the benefits of cloud software, especially to you, the translator, are very compelling. Foremost among them are the ability to fix bugs and add features rapidly (in our case, we release code several times per week), and also cross-platform compatibility.


I'm almost afraid to say it out loud here in public, but I'm actually a big fan of Google Translate (and a few other MT engines), and currently use them inside MS Word via GT4T (http://gt4t.net/en/ ), which I find very useful as a typing aid and source of ideas. I would therefore love to try out your new system, but as a plug-in/addon inside MS Word.


Yeah, Google Translate is miraculous. Really. It inspired me to go to grad school and study computational linguistics. But it's a broad-coverage, general-purpose system, hence the motivation for....


I was actually one of the first people to pledge money towards another new MT system, this one called Slate Desktop (https://www.indiegogo.com/projects/slate-desktop-the-personal-experience-for-pros#/ ). The developer sent me the release candidate a few days ago, but I haven't had time to install or test it yet. Back when I initially decided to buy Slate Desktop, I was still using CafeTran as my main CAT tool, and so at the time asked the developer of Slate Desktop to build a CafeTran plug-in.


If you're going to train an MT system locally, then you really should benchmark it against Google Translate. Or Microsoft. Or our system. It would be interesting if you shared the results here.

Thanks for your comments!


 
Pages in topic:   < [1 2 3] >


To report site rules violations or get help, contact a site moderator:


You can also contact site staff by submitting a support request »

Lilt: interactive stat MT for translators

Advanced search






TM-Town
Manage your TMs and Terms ... and boost your translation business

Are you ready for something fresh in the industry? TM-Town is a unique new site for you -- the freelance translator -- to store, manage and share translation memories (TMs) and glossaries...and potentially meet new clients on the basis of your prior work.

More info »
SDL MultiTerm 2019
Guarantee a unified, consistent and high-quality translation with terminology software by the industry leaders.

SDL MultiTerm 2019 allows translators to create one central location to store and manage multilingual terminology, and with SDL MultiTerm Extract 2019 you can automatically create term lists from your existing documentation to save time.

More info »



Forums
  • All of ProZ.com
  • Term search
  • Jobs
  • Forums
  • Multiple search