Walter von Hahn * Machine Translation
Robust processing must not only allow for permanent and useful
work of the system, but must always choose the correct level of
messages (if necessary at all), i.e. recognize the type of difficulty.
- It seems practical to apply a transfer model of translation,
that means not to represent a deep semantic/pragmatic meaning
first and generate from there afterwards. Instead, translation
should be processed on lexical level, if possible, on syntactic
level, if possible etc. The most flexible way seems to be a system
with variable depth, which can decide, on which level a similarity
between the source language and the target language can be established
for which phenomena.
- Which are the translation units to be "read in"
in each component: words (because of a possible shallow transfer),
sentences (because of syntactic analysis), paragraphs (because
of cataphorics) or always the whole text? This question is especially
interesting, when no segmentation is given in the text, e.g. in
continuous speech input.
- technical -
- Complex systems, especially
when they are highly dependent on the theories behind it (linguistic,
communicative, lexical, translation, social,...), must have an
architectural concept about how the components should interact.
A simple sequential architecture does not seem to allow necessary
interaction among distant components. A total communication among
all components will be too expensive for a system with time restrictions
or even real time demands (machine interpretation).
- large scale implementations.
- implementations on standard HW/SW environment.
7. System environment
Modern language technology, especially multilingual systems must
be embedded in the corresponding technical information environment.
In the near future the international data networks will become
an overwhelming information factor. Therefore MT in medium terms
must be designed for network applications, and it has to be integrated
in host systems. Translations are not aims or values as such,
but necessary access and information tools within other activities.
Host systems in this sense range from e-mailing systems over library
access or decision support systems up to Virtual Reality for telepresence
facilities. The awareness of these facts again brings in mind
the necessary international cooperation in data, formalisms, architectures
and strategies.
8. Meeting quality measurements
One of the most difficult problems
in machine translation is the measurement of quality. Most of
the parameters rely on linguistic and communicative measures:
- referential accuracy,
- linguistic correctness,
- pragmatic adequacy,
- cognitive correspondency, and
- cultural correctness.
These parameters, however, cannot be quantified and do not lead
to only one textual solution. Moreover, human estimations about
the quality of translations vary widely.
On the other hand, there are some
technical requirements which can be defined independently of the
linguistic parameters. Such requirement and constraints are derived
from the application class envisaged by designers and developers.
Among them:
- local or temporal availability of the system,
- realistic hardware requirements,
- reliability,
- necessary staff,
- required degree of preparation of the translation material,
- time behaviour,
- post-editing,
- maintenance (duration, intervals, staff),
- portability,
- extendability (costs, structural constraints, staff).
125
