RE: [ISSUE-75] - Domain - 2.a.

I find it a reasonable practice to define what is not in scope as a part of any specification, though agree that clear statements of in scope features are crucial.

I am curious about how a multi-engine selection/validation process works.  Christian, you mentioned both TM services as well as MT engines.  I can see value to be able to call from a set of services depending on domain with fallback based on result quality scores.  And you state that ITS 2.0 might be a single service scoped spec.

Yves, you believe that there is support for more than one MT engine as currently spec'd.  My interest in the white spaces between the two comments are when layering n-services of differing processing types, e.g., fuzzy matching TM services versus statistical MT engine results and how that plays out.  It seems very ambitious to me to provide scope for this, and yet having a system that is capable of providing the kinds of metadata needed to enable it would be a pretty powerful in terms of the potential to provide hi-fi results.

Maybe my comments are far out of scope, but the thread here caught my attention.  If this the case, I am happy to discuss it more offline, perhaps in Rome over a coffee.

Jan

________________________________________
From: Yves Savourel [ysavourel@enlaso.com]
Sent: Tuesday, January 29, 2013 7:55 AM
To: public-multilingualweb-lt-comments@w3.org
Subject: RE: [ISSUE-75] - Domain - 2.a.

Hi Christian, all,

I’m always a bit uncomfortable with stating what a mechanism is NOT doing in a specification. It seems we should be able to define what it does do and that should be sufficient.

I would also argue that the scenario “try MT engine A, then MT engine B” can work perfectly well with what we have today. The specification provides domainMapping for some basic mappings that allow for example to point multiple keywords to a more common unique 'domain' label.

For example you have a mapping as this: domainMapping="'sports law' Legal, 'property law' Legal"
and two MT engines: they each have a user-defined table that provide additional re-direction (they are even possibly pair specific: one maps 'Legal' to 'LEGAL_EN_PT' and the other maps 'Legal' to '5242e0762354527_legal'.

Using domainMapping for more than simple grouping is bound to have quick limitations:

a) what if you add a third MT engine? You have to edit every single rules document to add the new mapping?

b) how do you map to engine that are defined per pair?

IMO the mapping to the values used to slect the MT engine belongs to the process side, not the input.

cheers,
-yves



From: Lieske, Christian [mailto:christian.lieske@sap.com]
Sent: Tuesday, January 29, 2013 8:11 AM
To: public-multilingualweb-lt-comments@w3.org
Subject: [ISSUE-75] - Domain - 2.a.

Hi,

One of my comments related to “domain” (see http://lists.w3.org/Archives/Public/public-multilingualweb-lt-comments/2013Jan/0022.html) was the following:

2.a. Domain "systems" may not be harmonized across a processing chain. A Translation Memory component may for example work with different domains than a Machine Translation system that is part of the same processing chain. Since ITS 2.0 "domain" currently does not allow to capture the information "This is for component X" these scenarios cannot be addressed.

During the face-to-face in Prague, we achieved the following status (see http://www.w3.org/2013/01/23-mlw-lt-minutes.html#item09): a note should explain that “domain” (and possibly other data categories) do not accommodate what could be called multi-engine scenario.

Here is my suggestion for a note …

The focus of ITS 2.0, and some of the usage scenarios addressed in ITS 2.0 showcases (see http://www.w3.org/International/multilingualweb/lt/wiki/Use_cases_-_high_level_summary#ITS_2.0_Metadata:_Work-In-Context_Showcase) is on “single engine” environments. Example: the Machine Translation (MT) usage scenarios do not work along the lines of process chains such as “try MT engine A, then MT engine B”. Accordingly, ITS 2.0 has few provisions to support this kind of “multi-engine” environments which for example require domain-related information such as “try MT engine A with domain ‘financials’, then try MT engine B with domain ‘healthcare’”.
Cheers,
Christian

Received on Tuesday, 29 January 2013 16:17:11 UTC