Intervention
de Monsieur Nabil ALI
MACHINE TRANSLATION: A
CONTRASTIVE LINGUISTIC PERSPECTIVE
NABIL ALI
I Background
In 1983, the author was assigned
the task of developing the first bilingual Arabic-English home-computer.
This task was involved with the development of a bilingual operating system
(MSX-based), as well as the establishment of a software development unit
dedicated to Arabic applications. Due to the home-orientation of the project,
the emphasis was mainly on culture-ware and education-ware. This placed
higher priority on the Natural Language Processing and solving the myriad
of problems associated with Arabic computation. At that time, Arabic was
extremely underprivileged in the computation field, suffering the limitations
of a minimal system at a pure character level, and poor printing and display
qualities. Thus, it was necessary to shift to a more developed level, dealing
with larger linguistic units, namely the word, the sentence, and the continuous
text. Seeing as to how English was the most established computation example,
we had to draw on its resources and techniques. Shortly, we were to discover
that these techniques are not suitable for Arabic. This is simply due to
the fact that Arabic, compared to English, is much more complex at almost
all linguistic levels (with phonology as the sole exception). At the character
level, the complexity lays in the cursive shape and concatenation of Arabic
letters, and above all, these letters are characterized by a high degree
or context sensitivity. By this, it is meant that its appropriate shape
is determined by the surrounding letters. Hence, the author had to implement
an automata for the generation of the appropriate shape of letters depending
on their context. At the word level, the morphology of Arabic (the language
subsystem dealing with the structure of words) as it is already known,
is the most sophisticated of all languages world-wide. Therefore, the author
had to develop a morphological processor capable of analyzing any Arabic
word into its morphological primitives, as well as synthesizing the final
form of words out of these primitives. Lastly came the syntactical level,
which no doubt proved to be the most difficult, primarily because Arabic
is usually written without vowels. In essence, written Arabic is a quasi-stenographic
script, and this results in a severe mélange of various ambiguities,
which are unprecedented and absent from any other languages. The morphological
ambiguity, due to the absence of vowels, is intermixed with other types
of ambiguities, mainly those associated with word sense, part – of - speech,
and syntactical structure. For a non-Arabic speaker to appreciate such
a problem, let us assume hypothetically that an English sentence such as
" SOME FIRMS LEND MONEY" is written in the Arabic fashion. The result will
be the following string of consonants: "SM FRMS LND MNY" Each of these
consonantal forms may have a set of alternative vowelized interpretations
(figure:1)
Figure 1:
morphological ambiguity due to drop of vowels
Thus, any syntactical processor
dealing with Arabic text as its input has to primarily disambiguate such
quasi-stenographic script. As a result, an automatic vowelizer became mandatory
as a prerequisite for Arabic computation. To solve this problem, the author
has developed a "shallow" automatic understander of Arabic in order to
disambiguate the unvowelized text, as well as to substitute the missing
vowels. This required the achievement of three main computational linguistic
tasks: (1) the development of an Arabic parser; (2) the development of
a lexical-semantic processor; (3) the development of an automatic generator
of the vowelized text.
Since parsing techniques
developed for English have been proven inadequate for the Arabic language,
both in function and performance, a parsing system based on a multi-level
grammar, was developed and implemented. This system is capable of handling
the previously mentioned intermixed set of ambiguities. The disambiguation
mechanism works incrementally at every level of the grammar. Residual ambiguities
are resolved heuristically, resorting to preferential principles working
on both syntactic and semantic levels.
The availability of this
sophisticated system induced the idea to use it as a generalized model
for other languages. The system was successfully slimmed down to handle
English. For example, the morphological processor was slimmed down to be
used as an English stammer (to extract the stem of inflected word forms),
the shallow understander was slimmed down to an efficient English parser,
and the lexical-semantic disambiguated was tailored to handle types of
ambiguities encountered in English.
In general, the developed
Arabic-English bi-directional system could be characterized as being engineering-oriented.
The author felt that its rather adhoc approach had to be refined theoretically
through more serious investigation in the field of contrastive linguistics.
In this regard, the author has exploited intensively a basic property of
Arabic, namely its non Exotic charactor which places it in an intermediary
position among linguistic extremes, found in other languages.
The paper is intended to
present machine translation from a linguistic divergence. First, multi-linguality
will be overviewed contrastively as a preface to a more specific discussion
focusing on translation. It will conclude, with a brief description of
the contrastive aspects of the bi-directional, Arabic-English translation
system which is currently being developed.
II Multi-linguality: A
Contrastive Perspective.
Language in the present study
is viewed as a system comprised of two main components, grammar and lexicon.
Each of the two components will be considered consecutively.
According to Robert Freidin
(1), the comparative work carried out by nineteenth-century grammarians,
was concerned with establishing an explanatory basis for the relationships
between languages and groups of languages primarily in terms of a common
ancestor. Contemporary comparative grammar, in contrast, is significantly
broader in scope. It is concerned with a theory in grammar that is postulated
to be an innate component of a human mind/brain. In this way, the theory
of grammar is a theory of human language and hence establishes the relationship
among all languages, not just those that happen to be related by historical
accident (for instance, via common ancestry).
One can safely say that the
current advancements in contrastive linguistics are attributed to the adoption
of the generative paradigm. This paradigm has been applied successfully
across the different language subsystems, mainly morphology, syntax, and
semantics. A general theory of generative morphology, both derivational
and inflectional, has been developed (2). It has reached a level of maturity
that has made it possible to be applied cross-linguistically. This led
to different attempts to develop a universal computational morphological
system. These universal approaches deal basically with affixational morphology.
Their performance with regard to diffusional (non-concatenative) morphology
is still questionable. Furthermore, the majority of morphological research
focus on the word form much more than they do on the words’ semantics and
information content. Aronoff has initiated a semantic-based morphology
which views the derivation of words as a generative process from one word
sense to another. This shift toward semantic-based morphology is essential
for contrastive analysis at the word level in general, and for translation,
in particular.
At the syntactic level, the
diversity of languages is based on the "Government Biding" theory developed
by Chomsky. According to GB, a language is not a system of rules, but a
system of specifications for parameters in an invariant system of principles
of Universal Grammar (UG). Linguistic diversity can be explained as variation
in the setting of certain values for a principle of UG. Hence, linguistic
variations would, in part, be reduced to parametric variation for principles
of UG. An example of this parameter is that of the "null-subject" that
differentiates between languages using explicit subjects, such as French
and English, and those that permit the omission of the subject, such as
Arabic and Spanish. While the GB parameterization-based model is highly
theoretically apt, yet due to its general nature, it is too abstract to
be applied as practical computational systems for analysis (parsing) and
synthesis (generation).
As we move towards semantics,
linguistic divergence fades out. Although languages usually exhibit broad
disparity at the morphological and syntactical levels, such linguistic
disparity is greatly diminished at the semantic level, at which various
syntactical forms are converted to their corresponding logical forms. Formal
logic, is basically universal and thus can transcend linguistic boundaries.
At this logical level, the sameness of meaning is explicitly expressed,
this in turn, enables cross-linguistic mapping and transformation. Engineering-oriented
approaches to develop a universal semantic processor which can work multi-lingually
have been developed. They adopt a compositional paradigm which decomposes
meaning using a universal set of semantic primitives, A well-known example
of these processors is the conceptual-dependency model developed by Shank.
Regarding the lexicon, and
according to James Pustejovsky (3), "Computational and theoretical linguistics
have largely treated the lexicon as a static set of word senses tagged
with features for syntactic morphological and semantic information." This
view has undermined the role of the lexicon within the overall language
system. Indeed the pendulum has switched sides to the extent that it has
reached the other extreme of viewing the whole language within the lexicon
itself. The notion that morphology, syntax and semantics are in the lexicons
is gaining acceptance. Under this lexicon-based framework, the major role
of these language subsystems is to express formally (rule-based) the lexical
redundancies and regularities encountered among the different lexical entries.
Currently, the generative paradigm is being introduced in the field of
the lexicon in an attempt to upgrade the art of lexicography to the level
of exact science of what is now termed "lexicology". MIT has launched a
cross-linguistic lexicon project with the main objective to systematize
different aspects of lexical divergence.
A major challenge facing
generative lexicology is that related to metaphors. Since metaphors are
strongly linked to culture, it is difficult to isolate the culture-dependent
subparts from those that are independent. However, Lekov, in his seminal
work, Metaphors We Live By, sheds light on how to tackle the metaphor
problem cross-linguistically. By providing us with many examples of commonly
used metaphors that exist in many different languages, he initiates a new
approach to perceive metaphors at a higher universal level.
III Application to the
Field of Machine Translation
With the above overview of
multi-linguality, we now bring forth our contrastive analysis in the context
of machine translation. This transition could be better visualized if we
note that the changing attitude towards translation studies has always
been determined by transformations in the theory of language. According
to Marie-therese Abdel-Messih (4), one could summarize these transformations
in terms of the following milestones:
Plato’s theory of language:
which assumes the existence of a fundamental prior utterance with a determinant
literal meaning that can be transformed inter-lingually.
-
The rejection of the concept
of universal language that could be reached through an accurate translation.
Jacques Derrida re-interpreted the art of reading in terms of growth of
language whereby translation supplements one language with what the other
lacks.
In the field of machine translation,
"the kinship of languages" has different realizations. We can distinguish
three kinds of multi-lingual systems that can translate from one set of
languages to another. These are the inter-lingua translation model, the
parameterization model, and the
isomorphic grammar model
(figure 2). Here, the former two will be discussed while the latter will
be dealt with separately in the coming section.
Figure 2: Machine translation
models
(a) The Inter-Lingua model:
To recognize the significance of the inter-lingua machine translation model,
one may refer to the much more simplistic transfer scheme used to translate
mono-directionally from a single language to the other. Generally speaking,
this scheme abides by the previously mentioned "sameness" principle, whereby
the generated translated text is defined by the grammar of the source language
and the transfer subsystem which converts it to the corresponding target
language representation. For each language pair, there is a specific transfer
module. This is definitely a tough requirement which makes the transfer
scheme extremely impractical in case of multilingual bi-directional automatic
translators. The inter-lingua came as a solution to this problem. In this
model, a grammar of the source language defines an analysis component which
translates from the source language into an intermediate language known
as the inter-lingua. A grammar of the target language defines a generation
component which translates from the inter-lingua to the target language
(6). By this, instead of using different transfer modules for each language
pair, the inter-ligua works as a multi-lingual transfer module. A major
drawback in the inter-lingual model is that it requires for each member
language specific analysis and synthesis components.
(b) The Parameterization
Model: The Parameterization Model is an enhancement to the inter-lingua
model by overcoming its major drawback just mentioned. The approach to
translation is similar to the traditional approach in that it uses a language
independent representation. Nevertheless, in this approach there are only
one analysis component (multi-lingual parser) and one synthesis which work
multi-lineally. Language-specific information is factored out in terms
of parameter settings and the translation mapping operates uniformly across
all languages. The UNITRAN system developed by Dorr is one implementation
of this model. It uses parameterization in both the syntactic and lexical
distinctions (5). The parameter setting approach is desirable from a number
of different perspectives. First, it allows language-specific knowledge
to be represented independently from the language included in the syntactical
principles and the inter-lingual representation. Second, it appeases the
contrastive aspects. Third, the parameter setting approach allows a machine-translation
system to be easily modified and augmented. As previously mentioned, the
parameterization-based machine translator model is yet too abstract, and
this fact indeed does renders its implementation a difficult task.
IV Isomorphic Grammar
Approach
In developing his bi-directional
Arabic-English machine translation system, the author has adopted the isomorphic
grammar approach whereby the grammar pair of both the source and target
languages are attuned to one another. This differs from the inter-lingual
models in which grammars for different languages can be developed independently.
Isomorphism has been established through different lexical and grammatical
development arrangements. These are:
-
The same grammar model is used
for the language pair. In our system, the Generalized Phrase-Structure
Grammar (GPSG) developed by Gazder has been chosen.
-
Common guidelines for grammar
writing with regard to categorization and subcategorization, as well as
sequence of syntactical constructions and derivations.
-
Same lexicalization principles
which guarantee the compatibility of the detailed parts–of- speech between
the source language and the target language.
-
Identical verbal argumentation
scheme (semantic valence patterns).
-
Same features system to support
the semantic compositional analysis of meaning, and the specification of
morph-syntactical properties.
-
Same notational system for both
grammar and lexicon.
Bidirectionality of the Arabic-English
translation system has rendered its syntactical and lexical transfer components
much more sophisticated than that usually found in mono-directional machine
translation systems-main reason is that both members of the language pair
could act as a source language. Hence, the translation system does not
have the luxury of subletting the linguistic courage, either lexically
or syntactically.
Writing isomorphic grammars
for different languages, particularly those that are members of different
language families such as the case of Arabic and English (Arabic belonging
to the Semitic family and English to the Indo-European) is considered difficult.
However, through the experience of the author in writing grammars, for
both languages, the case has proved less difficult than expected. Isomorphism
emerges naturally, provided that the style of grammar writing abides with
a set of well-established principles. Needless to say, perfect isomorphism
can not be attained, and this is due to genuine differences between languages.
Regarding the Arabic-English language pair, major differences could be
summarized as follows:
-
Word formation: English word
formation is mainly of offixational nature. Arabic combines both offixational
and diffusional methods of word formation.
-
Difference in basic word order:
while English is categorized as SVO (S: subject V: verb O:object), Arabic
is basically VSO. Arabic also allows for nominal sentential patterns initialized
by the subject.
-
Null-subject parameter: where
English uses explicit subject that can be dropped in Arabic.
-
Syntactical Flexibility: English
syntax follows strict word order. Arabic syntax is more flexile.
-
Use of pronouns: Unlike English,
Arabic does not allow for stranded prepositions resumptive pronouns are
used instead to construct a full prepositional phrase.