Feature
|
Current
Typical MT System |
Anusaaraka |
| Goal
|
To
produce
Natural translation.
In case of failures produce 'rough'
translation.
(what is 'rough' is not well defined.)
|
To
provide access
to the Source Language
|
| Unit
of input |
Currently,
independent sentences
|
a
complete XML document
|
| System
Components
|
Morph
analysers, POS taggers, Parsers, Sense disambiguation modules,
Generator |
Same
as in MT
+
Anusaaraka
User Interface
|
| Sequence
of Operations |
outputs
are cascaded.
so errors too get cascaded.
|
The
basic tasks are processed independently.
Price paid: Duplication of effort |
Transparency
|
working
is
not transparent to
the end-user |
working
is
transparent to
the end-user |
| Access
|
User
has access
only to the final output |
User
has access to the
output at each level |
Guidelines
for Linguists
|
No
specific guidelines
|
First
write an algorithm
for
'Human beings' and not necessarily for 'computers'!
|
| Principle
|
Ad-hoc
|
"Information
Dynamics" |
Approaches
|
--
EBMT,
-- Rule based,
-- Statistical
-- Hybrid
|
ECLECTIC
Choose the best of each of these approaches.
Use best of the available resources under GPL.
|
CONESEQUENCES
|
- Later
modules are affected by the errors of the previous modules
- Rough
is not well defined. Hence users may get mislead.
- User
can not participate in the development process
- Linguists
end up in reinventing the wheel again and again.
|
- Parallel
processing ensure that different modules do not interfere.
- Well
defined 'Roughness'. Theoretically no chances of user getting
mislead.
- User
can participate in the development activity
- Linguist
prepares data only once.
|
