About Slate (TM)
Thread poster: A.K Janjelo
A.K Janjelo
A.K Janjelo
Lebanon
Arabic to English
+ ...
Mar 15, 2018

Good morning Tom Hoar,

I've been following your posts and contributions on LI for and I've been having these questions in mind for a while and I hope you can provide me with answers.

What is Slate? I mean what does it really do? Is it another CATool?

What about privacy and security of it?

I look forward to reading your answers and hope it initiates a discussion that would benefit everyone.


AK


 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
Slate revisited Mar 15, 2018

A.K Janjelo wrote:

Good morning Tom Hoar,

I've been following your posts and contributions on LI for and I've been having these questions in mind for a while and I hope you can provide me with answers.

What is Slate? I mean what does it really do? Is it another CATool?

What about privacy and security of it?

I look forward to reading your answers and hope it initiates a discussion that would benefit everyone.


AK


Thanks AK. I appreciate you re-asking your question in a forum where others can share.

Our website has lots of information, but maybe too many confusing details. How about this. I'll keep my answers short in this discussion so you and others can follow-up with questions that zero-in on your interests. To the Forum moderator, if I violate the forum rules, please ask me to change and I'll be happy to drop anything that sounds too much like a sales pitch.

Slate is our brand of software applications. They are not CAT tools. They are add-ons to your CAT. We originally launched Slate Desktop Edition 2 years ago. Since then, the company name has changed to Slate Rocks, LLC and we've added Slate Starter Edition, Slate Connect Edition and Slate Trial Edition (and others). Here, I'll describe basic functionality they share. There's a table on our website that shows the different features for each edition.

Slate products install on your computer. There aren't any cloud components. It runs totally isolated from the Internet. So, your work never leaves your computer.

Each product has two "modes." In one mode, you import your translation memories and Slate converts them into a translation engine (or as many different engines as you like). Until you convert TMs into an engine, Slate has no practical functionality.

Once you create your engine, you add the "Slate Desktop connector" to your CAT project, just like you would add Google Translate (or other) connector. You then configure the connector to use one of your Slate engines. From that point on, you use your Slate engine just like you would use any other MT connector.

From a functional standpoint, where's not a whole lot more to it. We've followed the KISS principle. So, the two modes are pretty easy for non-technically oriented translators to use. If you manage your own TM files, you have the skills to make and use Slate engines. There are countless features and benefits that we can talk about at your leisure.

Finally, Slate supports 33 languages in any paired combination. That means you can create engines for 1,056 different language pairs. Yes, Arabic is one of the 33 languages.

Even this "short" description is 7 paragraphs. I'll stop and let you ask some questions.

Tom

[Edited at 2018-03-15 14:52 GMT]


 
Rolf Keller
Rolf Keller
Germany
Local time: 20:28
English to German
Please clarify Mar 16, 2018

Tom Hoar wrote:

Our website has lots of information, but maybe too many confusing details.

Confusing? Yes. Details? Um.

The only details I found about what Slate provides to the translator are these:
"With Slate™, your CAT still offers you high fuzzy matches from your translation memories. When there’s no match, Slate™ steps in to offer suggestions with terminology and style representative of what’s in your translation memories"

This raises some questions. If Slate bases on the TMs you own already, how come that slate is able to offer translations which the CAT tool isn't able to offer? And how does Slate discern and assess style details? In 33 languages and 1,056 language pairs, as the website states? And all this in offline mode, i. e. without Google's or Microsoft's resources (huge databases and a farm of 999 computers with high-level AI software).

Could you please clarify & explain this?


 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
Clarifying (second try) Mar 16, 2018

Rolf Keller wrote:

Confusing? Yes. Details? Um.

  • offline mode? i. e. without huge databases like Google & Microsoft use

  • offline mode? i. e. without a farm of 999 computers like Google & Microsoft use

  • offline mode? i. e. without high-level AI software like Google & Microsoft use

  • In 33 languages and 1,056 language pairs, as the website states?

  • And how does Slate discern and assess style details?

  • how come that slate is able to offer translations which the CAT tool isn't able to offer?




Hi Rolf. I appreciate everyone's feedback to help me do the mental gymnastics to communicate this information. I reordered and listed your questions so each answer build on the previous. Here's an outline with short answers. I'll post a dedicated reply with details for each.

  • Huge Databases: Slate uses huge databases because converts your TMs into huge databases during the machine learning processes.

  • Computers: Slate's hardware requirements are much more substantial than your CAT's spec that supports 1995 hardware. Fortunately, a modern gaming computer for under $700 will serve you well.

  • High-Level AI Software: Slate uses mature, proven open source AI and machine learning software from an open source project called Moses SMT.

  • 33 languages and 1,056 language pairs: Slate ships with the tools necessary to support these languages. You use Slate to convert your TMs with these languages to engines.

  • Discerning style: Slate engines learn from the TMs you use to create the engine and output the style that they learned. That's what machine learning is about.

  • How Can Slate Offer Translations The CAT Can't?: That comes from SMT. Slate works at a sub-segment level. Its database contains every possible 1-, 2-, 3-, 4-, and 5-word bilingual concordance in your TMs. It splices them into translations based on the information in those databases.



 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
Huge Databases Mar 16, 2018

Google & Microsoft use databases of appropriate sizes to serve millions of people per day across unlimited subjects. I shutter to think how big they are considering the stats I'll give about Slate

Slate converts your TMs into two "huge" databases for any given language pair. We recommend starting with no less than ~100,000 segments. After Slate is done with its conversion process, those segments could become 1 billion entries in these databases. Th
... See more
Google & Microsoft use databases of appropriate sizes to serve millions of people per day across unlimited subjects. I shutter to think how big they are considering the stats I'll give about Slate

Slate converts your TMs into two "huge" databases for any given language pair. We recommend starting with no less than ~100,000 segments. After Slate is done with its conversion process, those segments could become 1 billion entries in these databases. This is an appropriate size to serve one translator in the (comparatively) limited life experience of one person. A full-time professional translator can create that number of segments in 3 to 4 years. Because they were created by one translator during his work, the personal TMs are representative of the work at hand and the resulting engine out-performs Google and Microsoft (and others) for new work in this representative field.
Collapse


 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
Computers Mar 16, 2018

Google & Microsoft need a pretty big server farm with many many more than 999 computers. It's large enough to use their databases (which, as above, are large enough to service millions of people per day virtually all subjects), and support the computational demands of millions of people.

Slate serves a much smaller user base -- one translator. Needless to say, one person does not generate enough computing load to require a server farm. We designed Slate t
... See more
Google & Microsoft need a pretty big server farm with many many more than 999 computers. It's large enough to use their databases (which, as above, are large enough to service millions of people per day virtually all subjects), and support the computational demands of millions of people.

Slate serves a much smaller user base -- one translator. Needless to say, one person does not generate enough computing load to require a server farm. We designed Slate to run on modern desktop and notebook computers. Ypu'll find that the hardware requirements of a typical CATs and/or SYSTRAN 8 date back to the early 1990's (not very demanding). Slate's hardware recommendations are more like a modern performance gaming computer. Nonetheless, a new box with Slate hardware specs costs less than $700.
Collapse


 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
High-Level AI Software Mar 16, 2018

Google & Microsoft used advanced AI and machine learning software that relied on statistical machine translation (SMT) technology for 10 years. SDL's Language Weaver SMT technology predates both. KantanMT.com, letsMT.com, languagestudio.com, SAP, and many others used virtually identical SMT technology from an open source project called Moses.

In 2017, all of these companies abandoned SMT technology in favor of new AI/machine learning software called neural-network machine tra
... See more
Google & Microsoft used advanced AI and machine learning software that relied on statistical machine translation (SMT) technology for 10 years. SDL's Language Weaver SMT technology predates both. KantanMT.com, letsMT.com, languagestudio.com, SAP, and many others used virtually identical SMT technology from an open source project called Moses.

In 2017, all of these companies abandoned SMT technology in favor of new AI/machine learning software called neural-network machine translation (NMT) technology. They did not abandon SMT because NMT created better translations. SMT had reached its technical and quality limits within their use case (i.e. serving millions of people). Everyone (including me) agrees that NMT quality started at parity with SMT and it has potential to improve beyond SMT's limits for their use case. That use case is a cloud-based application that requires professional translators to pay a subscription fee and log-in to their services.

Slate uses the same advanced AI and machine learning software that dozens of other companies have used, i.e. the Moses SMT open source toolkit. This software does not require mega databases and huge server farms. It runs quite well on a desktop computer. The difference is, Slate Rocks, LLC, however, is the only company in the world that sees value in a desktop application for translators so they have control of their destinies without the internet and without subscriptions.

There was one small problem. No SMT software anywhere in the world ran on the world's most popular business productivity desktop, i.e. MS Windows. I invested more time, money, blood, sweat and tears than I care to remember to make the MS Windows version a reality. All code updates were contributed to the original open source trunk and anyone in the world can download that source code from the Moses project and compile it for Windows. Yet, no one else has. Why not? They do not see any business opportunity to abandon their cloud-based services and enable translators to be self-sufficient.

Why haven't we abandoned SMT and moved NMT? Because we serve a different use case. Our use case is the translator's desktop which everyone else in the world rejects. For many technical reasons, NMT is not ready for our use case and SMT still delivers better translations in this environment. Yes, Slate has some of the most advanced and proven AI and machine learning software ever. That machine learning learns from your TMs to create huge databases. The AI uses that database to generate candidate translations... All of that runs on your computer because it doesn't have to server millions of other people.
Collapse


 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
33 languages and 1,056 language pairs Mar 16, 2018

Languages are to Slate like colors are to Photoshop. Slate ships with zero ability to translate. If you were to buy SYSTRAN 8, you choose which engine you get for the price. Slate is not like that. You must supply the TMs.

So, when we say Slate supports 33 languages, it means it has the tools to create and uses engines with those language as the source or target language. That makes 1,056 language pairs because EN-FR is a different language pair t
... See more
Languages are to Slate like colors are to Photoshop. Slate ships with zero ability to translate. If you were to buy SYSTRAN 8, you choose which engine you get for the price. Slate is not like that. You must supply the TMs.

So, when we say Slate supports 33 languages, it means it has the tools to create and uses engines with those language as the source or target language. That makes 1,056 language pairs because EN-FR is a different language pair than FR-EN.
Collapse


 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
Discerning Style Mar 16, 2018

The language and translation style that come out of an engine were in the TMs that created the engine. So, an FR-ES translation engine created by Jorge with Jorge's TMs will reflect Jorge's style. If Jorge creates one engine using client-specific TMs, then its output will reflect Jorge's style within the requirements of that client (assuming Jorge was faithful to that client's style guide).

Let's go back to the Photoshop metaphor. A graphic designer can create as many images as they
... See more
The language and translation style that come out of an engine were in the TMs that created the engine. So, an FR-ES translation engine created by Jorge with Jorge's TMs will reflect Jorge's style. If Jorge creates one engine using client-specific TMs, then its output will reflect Jorge's style within the requirements of that client (assuming Jorge was faithful to that client's style guide).

Let's go back to the Photoshop metaphor. A graphic designer can create as many images as they want with Photoshop. Likewise, a translator can create as many engines as you like within the limits of the TMs you have. You can create one engine for Client A and another engine for Client B. You can create another engine for Industry 1 and another for Industry 2. TM management, which should be a skill for all modern translators, is the engine's output. Each engine will discern the style of the TMs that created it.

Of course, it's also true that your engines can create as many translations as you like, i.e. there's no subscription fee or royalty to use the software.
Collapse


 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
How Can Slate Offer Translations The CAT Can't? Mar 16, 2018

The key is how statistical machine translation (SMT) works with that huge database (above). Let's first understand what the CAT can do, then see how Slate does what they can not.

TMs: CATs work with a TM which contains full segments. (Yes, newer CATs have an advanced feature for sub-segment matches, but let's stay with traditional CAT features.) The CAT knows the source sentence (duh!). It searches the TM for similar source segments and suggests the corresponding trans
... See more
The key is how statistical machine translation (SMT) works with that huge database (above). Let's first understand what the CAT can do, then see how Slate does what they can not.

TMs: CATs work with a TM which contains full segments. (Yes, newer CATs have an advanced feature for sub-segment matches, but let's stay with traditional CAT features.) The CAT knows the source sentence (duh!). It searches the TM for similar source segments and suggests the corresponding translation. To oversimplify, if I have a TM with 10 TU's, then it can service a project by offering 10 translations that "fuzzy" match from 10 to 20 (maybe more) source sentences.

CAT w/ TM 1:
source segments in TM: 10
possible source segments: 20-30
possible translated segments: 10
percentage of correct suggestions: 100%

CAT w/ TM 2:
source segments in TM: 100,000
possible source segments: 200,000-300,000
possible translated segments: 100,000
percentage of correct suggestions: 100%

As you see, CAT with TM scales linearly regarding how many source/target pairs it can cover.

Slate: CATs work with the Slate engine which doesn't contain any full segments. The SMT machine learning step (AKA build or training) tears apart each source & target segment pair in the TM into fragments typically no longer that 5 words, and counts those fragments. Finally, it stores the fragments and their corresponding counts in that huge database. So, the segments in your TM easily becomes hundreds of millions, billions or even trillions of entries in the database. Think of the database as a huge store of bilingual concordances, each with no more than 5 consecutive words.

From inside your CAT, SMT splices and dices those fragments to create candidate translations according to the source-target probabilities stored in the database (for those who know a bit more, I'm intentionally oversimplifying). That means the SMT engine can create sentences for every possible combination and permutation of all words in the TM.

Let's go back to that TM with only 10 segments. Let's say each segment has only 6 words. With some repeated words, maybe there are about 450,000 possible permutations of the words. Most are total meaningless garbage. Then, let's assume that out all those permutations, only 1,500 of them are proper sentences in the source and target languages. Based on these assumptions, our table looks very different.

CAT w/ Engine 1 (made from TM 1):
source segments: 10
possible proper source segments: 1,500
possible proper translated segments: 1,500
percentage of correct suggestions: 80% to 90%

The "percentage of correct suggestions" value will never be 100% because SMT technology is not a simple database lookup and therefore it has inherent uncertainties. There are reasonable numbers for an engine from our second TM:

CAT w/ Engine 2 (made from TM 2):
source segments: 100,000
possible proper source segments: 1,500,000
possible proper translated segments: 1,500,000
percentage of correct suggestions: 30% to 40%

CAT w/ Engine 3 (made from TM 3):
source segments: 300,000
possible proper source segments: 4,500,000
possible proper translated segments: 4,500,000
percentage of correct suggestions: 25% to 30%

CAT w/ Engine 4 (made from TM 4, aka Google/Microsoft):
source segments: 10 million
possible proper source segments: 1.5 x 10e9
possible proper translated segments: 1.5 x 10e9
percentage of correct suggestions: 3% to 7%

I made up all of these exact numbers, but they are representative of what my customers experience with their TMs. The first TM, however, is real. I have a training exercise where customers create an engine with only 10 segments and actually observe its performance. In general, customers experiences are very similar to these numbers.

[Edited at 2018-03-18 02:52 GMT]
Collapse


 
Luca Tutino
Luca Tutino  Identity Verified
Italy
Member (2002)
English to Italian
+ ...
A critical question Mar 16, 2018

Thank you, Tom. Slate seems great and I have been keeping it on my radar for a while.

I actually have 2 questions.

Critical: Minimum/Recommended/Ideal hardware requirements. Can you please specify the HW requirements per major component?
- I guess that if building an MT from an average TM on a $700 gaming PC should take several hours, trying to do it on my very old i3 PC, should it work, would require a couple of days. And I should probably forget about
... See more
Thank you, Tom. Slate seems great and I have been keeping it on my radar for a while.

I actually have 2 questions.

Critical: Minimum/Recommended/Ideal hardware requirements. Can you please specify the HW requirements per major component?
- I guess that if building an MT from an average TM on a $700 gaming PC should take several hours, trying to do it on my very old i3 PC, should it work, would require a couple of days. And I should probably forget about my >1MM TUs larger memories - right?
Even if the Trial is free, I would not want to send out my payment info just to download a file that I will not be able to install or use until I will have bought a new machine.

Secondary: Version upgrades. Is there an upgrade option allowing the user to move from a lower version to a higher one without paying it from scratch? If yes how much could it cost?
Collapse


 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
FAQ et. al. Mar 17, 2018

Hi Luca,

Good questions. I should have mentioned our FAQ (https://support.slate.rocks/support/solutions/6000136200). It covers many more questions.

NOTE: Some of these links require that you sign-in to our website (it shares single-signon with the support site) because
... See more
Hi Luca,

Good questions. I should have mentioned our FAQ (https://support.slate.rocks/support/solutions/6000136200). It covers many more questions.

NOTE: Some of these links require that you sign-in to our website (it shares single-signon with the support site) because they're mostly for customers. Requiring sign-in significantly reduces attacks. I often grant access to those who ask as long as I can verify their identity. Everyone here, please feel free to request sign-in credentials.



Critical: Minimum/Recommended/Ideal hardware requirements. Can you please specify the HW requirements per major component?

You'll find the hardware requirement here: What are the system requirements? (https://support.slate.rocks/support/solutions/articles/24000002965)



- I guess that if building an MT from an average TM on a $700 gaming PC should take several hours, trying to do it on my very old i3 PC, should it work, would require a couple of days. And I should probably forget about my >1MM TUs larger memories - right?

Essentially, you're asking what is the effect of hardware spec vs computational performance (I specify because I refer to "linguistic performance" in other contexts). Yes, for the most part, older hardware simply runs slower. However, Slate products only run on 64-bit operating systems. An older PC with 32-bit Windows won't work.

On our Forum (https://support.slate.rocks/support/discussions) you'll find several customers share their experiences with TM size vs computational and linguistic performance. One customer, Igor Goldfarb, bought Slate and first used it with his Core2Duo, 4 GB RAM, 500GB disk system. His first engine finished building an engine from ~80,000 segments in ~26 hours. Once satisfied, he bought a new system (nice one!). That one now builds engines with ~120K segments in about 4 hours.

You might also want to review this article about TM size vs quality “Ideal” data for Slate Desktop (https://support.slate.rocks/support/solutions/articles/6000150332)



Even if the Trial is free, I would not want to send out my payment info just to download a file that I will not be able to install or use until I will have bought a new machine.

Slate™ Trial Edition doesn't have any strings attached. Add it to your shopping cart and checkout with only your email address, i.e. no credit card required. You have 60 days to download the installer. Once installed, you have 30 days to use the system. Please time your download accordingly.

If you want to use the MT connectors as part of your trial, you'll need to buy Slate™ Desktop Edition or Slate™ Starter Edition. They include a 30-day money-back offer. I have been known to extend that for customers with special needs, but never more than 60 days. We use PayPal as our payment gateway (credit cards accepted via their system) and they will not give me a full refund after 60 days.

Note that Slate™ Trial Edition does does not include the MT connectors that enable you to use your engine from your CAT. From most trial users' perspective, they're testing Slate. More accurately, they're testing their translation memories. Slate functions the same regardless of the TMs or customer perspective. Again, back to the Photoshop metaphor. If you click the "green" color, then the fill tool should fill the selected area with green color. Whether or not the graphic designer creates an aesthetically appealing image is the designer's responsibility, balanced with Photoshop's functionality.

The customer's TMs determine Slate's output within the limits of the fundamental technology. That output and the translator's expectations determine if Slate is beneficial. Therefore, I strongly recommend that you try Slate with your current hardware and have a LOT of patience during the build processing. With very old hardware, you might burn 7 or 8 days out of your 30-day free trial just waiting for the build to finish if your TMs are very large.

By the way, my use of the term "very large" differs depending on the audience. Translators might consider 1 million segments "very large." A computational linguist who creates the underlying open source tools considers this "tiny." In his context, "large" is typically in the 10's of millions and "very large" could be hundreds of million and "huge" starts at a billion. Check-out this article: What is "huge?" (https://support.slate.rocks/support/solutions/articles/6000163452)



Secondary: Version upgrades. Is there an upgrade option allowing the user to move from a lower version to a higher one without paying it from scratch? If yes how much could it cost?

Yes, but it is not published on the website's store because I can't figure out how to do it with our e-commerce tool (Easy Digital Downloads). To upgrade from Slate™ Starter Edition to Slate™ Desktop Edition you just send me an email and pay $310 (various methods). That's 10% premium over the delta between the respective $549 and $269 prices for two transactions vice one.
Collapse


 
Luca Tutino
Luca Tutino  Identity Verified
Italy
Member (2002)
English to Italian
+ ...
Thank you Mar 19, 2018

Thank you Tom. I am certainly going to give it a try as soon as I find a moment.

 
Tom Hoar (X)
Tom Hoar (X)
United States
Local time: 15:28
English
Thank you... Mar 22, 2018

Thanks, Luca and the others who have contacted me privately about a trial.

As you work through the trial, please make sure you review articles on the support site and also this one on our blog:

https://slate.rocks/evaluating-slate-desktop/

Working with Slate suggestions is different than working with TM suggestions. With TMs you have a source fuzzy match score that filters the poor
... See more
Thanks, Luca and the others who have contacted me privately about a trial.

As you work through the trial, please make sure you review articles on the support site and also this one on our blog:

https://slate.rocks/evaluating-slate-desktop/

Working with Slate suggestions is different than working with TM suggestions. With TMs you have a source fuzzy match score that filters the poor candidates. With Slate, that crutch disappears and Slate presents everything. Some translators report feeling overloaded. This article suggests strategies for managing that overload.
Collapse


 


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


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

About Slate (TM)







Trados Studio 2022 Freelance
The leading translation software used by over 270,000 translators.

Designed with your feedback in mind, Trados Studio 2022 delivers an unrivalled, powerful desktop and cloud solution, empowering you to work in the most efficient and cost-effective way.

More info »
Wordfast Pro
Translation Memory Software for Any Platform

Exclusive discount for ProZ.com users! Save over 13% when purchasing Wordfast Pro through ProZ.com. Wordfast is the world's #1 provider of platform-independent Translation Memory software. Consistently ranked the most user-friendly and highest value

Buy now! »