Systran MT/TM integration
By Jean Senellart
Director Research and Development,
SYSTRAN
Get the List of 4,500+ Translation Agencies Now!
No Recurring Membership Fees!
 With the longest history
of any MT developer in the world, SYSTRAN, founded
in 1968, has an R&D investment measured in thousands
of person years. From the beginning, SYSTRAN’s development
has been deeply rooted in linguistics and the system
still remains in constant evolution. The translation
engines, linguistic resources and user-interactive
customization tools are in step with the latest computing
standards and new development techniques, as the company
is always working on introducing innovative linguistic
advancements.
SYSTRAN’s goal is to better describe
languages, the gateway to continue improving the quality
of automatic translation.
Today SYSTRAN is the market leading
provider of language translation software products
and solutions for the desktop, enterprise, and Internet
that facilitate communication in 52 commercially
available language pairs and in 20 vertical domains.
Almost 20 additional language pairs have been developed
for specific projects or customers.
SYSTRAN is the choice of leading global
corporations, portals that include AltaVista™, Apple,
Google™, Yahoo!®, and governmental institutions throughout
the world like the US Intelligence Community. Use
of SYSTRAN products and solutions enhances multilingual
communication and increases user productivity and
time-savings for B2E, B2B, and B2C market segments
as they deliver real-time language solutions.
Although there is a wide spectrum
of SYSTRAN products and solutions to choose from,
each allows users to instantly translate any written
text for gisting (understanding the general idea of
what is written, such as quick understanding of foreign
language Web content) and for publishing (near-perfect
translations that also require post-editing, like
user guides, technical support content, and other
common localization projects).
When used for publishing purposes,
professional users often combine MT and TM’s (translation
memories). The simpler integration applies TM’s
first, and MT on “no match” segments as a default
translation. A richer approach uses TM’s and, more
generally user feedback, to supply MT resources.
It is in the context of the second
approach that we provide an overview of how SYSTRAN’s
translation engines work, highlighting the importance
of the linguistic descriptions, how existing TM’s
can be reused to customize the translation engines,
and available tools for managing Translation Memory
and User Terminology within a translation workflow
based on MT.
LINGUISTICS AND THE
IMPORTANCE OF DESCRIBING LANGUAGES
Two key elements distinguishing SYSTRAN’s
MT system from the others are that it’s an incremental
system and is deterministic. The system is designed
to produce incremental translation quality results
between versions which can be easily validated by
users. Additionally, the system produces deterministic
output, meaning the results are consistent and based
on available resources. As a result, users are able
to interact with the system to modify results by customizing
linguistic resources.
It all starts with the three types
of linguistic descriptions provided for each language
pair (source language to target language) implanted
in the system: Analysis, Transfer, and Generation.
The following diagram illustrates
the process for translating a source language sentence
into a target language sentence. The deeper the source
language analysis, the smaller the transfer will be
(and the smaller the effort to build new language
pairs).
 1.
The description of the source language, also referred
to as the analysis is composed of the following analyses:
• Global Document Analysis
considers the input text as one unit and performs
several rounds of analysis that identify important
elements that help describe the source language. These
include language identification at the paragraph level,
the named entities (such as dates and proper nouns)
that define the local terminology, and subject detection
of the document which enables the system to automatically
select preferred meanings by domain.
• Grammatical Analysis
provides the system with the data required to create
the internal representation of each by displaying
the complete linguistic structure of each sentence.
Included are part of speech for each word, the syntagm
to which they belong, the relationship between different
entities, and the function of main elements (verb,
object, subject, and complement). This deep description
builds a hierarchical representation of the dependencies
between the different elements.
The system’s analysis involves several
rounds represented by a sequence, each of which has
several dependencies to other components, including:
• Sentence segmentation – identification
of sentences in the text
• Normalization of the languages
• Morphological analysis
• Grammatical Disambiguation
• Clause identification
• Basic local relationships
• Enumeration analysis
• Predicate/Subject analysis
• Preposition Rattachement
• Semantic Analysis of the different words used
in context allows the system to automatically tag
words with associated semantic features that are
used in a later part of the process.
2. The description of the transfer from the source
language to target language focuses on the transfer
of structures and the transfer of the lexicon. It
is the only description dependent on both the source
and target languages. For instance, the internal analysis
of the previous sentence is transferred (from English
to German) into the structure represented in following
figure.
3. The description of the target language also referred
to as the generation.
These three linguistic descriptions are based on
linguistic resources; monolingual for the analysis
and generation or bilingual for the transfer. There
are two primary types of resources: rules and dictionaries.
Dictionaries describe individual terminological units.
Rules generalize these descriptions while providing
high-level descriptions of linguistic phenomena.
Typically linguistic resources are very large, and
for some language pairs can reach up one million entries.
Most of these resources can be learned (with supervised
semi-automatic extraction tools) from a bilingual
corpus – such as existing translation memories - and
easily therefore adapted to specific domains, which
means the translation is customized.
In comparison, the rules count is smaller but rules
rank much higher in terms of complexity and require
linguistic expertise for creation and maintenance.
ARCHITECTURE
SYSTRAN systems are highly modular and are based on an XML workflow, a mechanism enabling communication
between users and the different modules, and between
the different modules themselves. This mechanism provides
interaction between users and the system’s internal
rules.
The interaction between users and internal rules
is enabled by rich interactive tools embedding MT.
For high productivity, professional users must be
able to understand the translation process, as well
as interact with the rules and resources in order
to fine-tune their translations. An example of a rich
interactive tool is the SYSTRAN Translation Project
Manager, a translation workbench available in select
SYSTRAN products and solutions. All of the features
mentioned below are included in this tool.
• The translation engine produces
rich markup and allows users to view the impact
of their resources, alternative meanings, and indicators
on sentence complexity.
• Build and apply TM’s.
• A step that goes beyond TM’s is the ability to
store and reuse any user-choice resource for other
translations using SYSTRAN’s translation choice
feature.
SYSTRAN PRODUCTS
SYSTRAN’s array of products and solutions for the
desktop, enterprise, and Internet help enterprise
and home users understand foreign language content
in real-time and create multilingual documents. Released
in February of this year, SYSTRAN 6 brings 12 new
language pairs, more than one million new terms created
from aligned data, a dictionary lookup, a comprehensive
environment for post-editing and QA that includes
terminology extraction, flexible TM’s, collaborative
dictionary management, built-in comparators, and other
tools to efficiently optimize translations in a cost-effective
manner.
Noteworthy rich interactive tools and technology
associated to the translation engines follow:
• SYSTRAN Translation Project Manager (STPM) is a
“translation workbench” used to create, manage, and
refine localization projects consisting of hundreds
of files. Using STPM users perform side-by-side comparisons
between original and translated documents and affect
changes to both, as well as add terms to User Dictionaries
and process dictionary updates. In addition, STPM
offers a selection of powerful built-in review tools,
including terminology review, analysis of the original
document, full sentence review, use of alternative
meanings, and others for applying, reviewing, and
building TM’s and other advanced features.
• SYSTRAN Dictionary Manager (SDM)
allows users to create and manage three levels of
linguisti c data types to improve translation quality.
The data types are:
o User Dictionaries (UDs) – user
created bilingual or multilingual glossaries that
are used alongside SYSTRAN’s built-in domain dictionaries.
o Normalization Dictionaries (NDs)
- monolingual resources that can be used to standardize
or correct source text prior to translation, or
to correct target text after translation.
o TM’s
• SDM is based on IntuitiveCoding,
a proprietary SYSTRAN technology that allows users
to massively import and manage entries in User Dictionaries.
• Dictionary Lookup provides additional contextual
information for alternative meanings of selected source
language terms. Users can select a term at any time
across multiple dictionaries covering the SYSTRAN
Main Dictionary, 20 SYSTRAN Domain Dictionaries, User
Dictionaries, and other integrated dictionaries.
Set translation choices and customize terminology
with contextual review tools in SYSTRAN Translation
Project Manager.
CONCLUSION
It is essential to customize the translation engines
for use within a localization workflow. This customization
process is based on a variety of rich user-interactive
tools that leverage existing TM’s and exploit the
post-editing effort to create complementary linguistic
resources. Based on linguistic tools, this customization
process advances the quality of the linguistic descriptions
involved in the translation process and progressively
increases overall translation quality.
Source sentence
Target sentence
Post-analysis representation
Pre-generation representation
Language independent representation
Surface level
Analysis
Transfer
Generation
ClientSide
News Magazine - www.clientsidenews.com
|
