Project partners
IMPACT is a trial project in the framework of the European
"User-friendly Information Society" programme. In this regard, CAS
Software, the supplier of the CRM Groupware solution CAS genesisWorld, is
cooperating with The Language Technology Centre (London), a translation and
language technology specialist, as well as with Infoworld Srl (Milan), a
CAS genesisWorld sales partner for Italy, and the software system house New
Emphasis (Greece).
Lessons learnt
The IMPACT project represents a good use of machine translation, in that
the application does not require output of publishable quality, but only
understandable results.
However, the pilot has demonstrated how difficult it is to achieve even
this less demanding standard, and highlighted the extent to which the
quality of the text input to the MT influences the quality (intelligibility,
completeness) of the output. Single words or expressions that are not
translated, or translated incorrectly, can generally be simply added to the
Systran dictionary; this applies particularly to customer-specific
terminology or usage (i.e. terms used by a given customer in a specific
way). On the other hand, some terms will be difficult for Systran (or any
machine translation program) to handle because they are ambiguous, and the
correct translation depends on the context.
Some difficulties cannot be overcome by defining additional rules
Similarly, certain grammatical structures cause difficulties that cannot
be overcome simply by defining additional 'rules'. Many studies have shown
the benefits of enforcing stylistic rules on text to be machine-translated,
either by originators following 'style guidelines', by imposing a degree of
structure on the input, or by having text pre-edited by another person
before entry to the MT system. In the IMPACT scenario, it is unrealistic to
ask end-users to follow strict style guidelines when entering error
notifications, and experience has shown that these messages are of mixed
quality linguistically (particularly where users submit input in a language
other than their own) - hence the reliance on post-editing during the IMPACT
pilot.
Spelling errors in trouble tickets are problem
Another issue was the high number of spelling errors in the trouble
tickets submitted. This was addressed by introducing an external web based
spellchecker to help intercept misspelt or mistyped words at source, but
this had to be invoked manually by the originating user; it did
significantly reduce errors, but could not eliminate them entirely. There is
a limit to how far end-user input can be structured; ideally, data entry
would include selecting various keywords from drop-down lists, but this
involves downloading a lot of data to the web client, which may not be
practical for performance reasons. It might also be possible to use an
Artificial Intelligence (AI) component to help with the assignment, but this
was not in the scope of IMPACT. The messages processed by the IMPACT
prototype were typically short and often elliptical in expression, which
would not be well suited to an AI approach. The quality of outgoing messages (i.e. from the help desk support
engineers) is much easier to control.
Not always new machine translation needed
Where a trouble ticket refers to a
problem that has been reported before, the engineer can simply retrieve an
existing response from the knowledge base; the equivalent text in the user's
language will be found in the translation memory, and no new machine
translation will be needed. For "new" solutions, it should be
quite feasible to train the support engineers to apply consistent style
guidelines and restricted vocabulary in composing their response messages,
to avoid terms and constructions that cannot be dealt with comfortably by
the MT.
|
