In a transfer-based system the process of translation comprises the following processing steps:

Morphological analysis. Word-forms of the source text are analysed using paradigm sets and identified with the dictionary entries.

Syntactic analysis. At this stage using the information from the dic­tionary and paradigmatic data syntactic representation of the source text is formed by the syntactic analyser (called parser). A string of syntactic classes or a syntactic tree of the source sentence is passed over to the transfer module.

Transfer. The transfer module receives the syntactic representation of the source text and, using relevant transfer rules, converts it into an intermediate representation.

Syntactic synthesis. At this stage a final syntactic representation of the target text is formed by combining and matching the transferred structures of the source syntactic representation.

Morphological synthesis. Using the information from the target dictionary and paradigmatic data the target text (translation) is obtained.

The way the syntactic representation is formed, analysed and trans­ferred greatly depends on the grammar used. The most common ways are so-called «templates», applied to the linear string of syntactic classes, dependency grammar (DG) or immediate constituents (IC) grammar.

Transfer-based systems rather often comprise a semantic compo­nent. A network of semantic descriptions and relationships is superim­posed on syntactic structures of the source and target texts. The purpose of the semantic component is to improve the accuracy of translation.

Pivot language-based machine translation is the third basic method. In a way it is similar to the transfer-based; however, there are several critical differences.

As opposed to transfer procedures which are applied mostly at the syntactic level with some corrective semantics, pivot language represen­tation involves all available linguistic information.

Besides, transfer-based translation is intended for a concrete lan­guage pair, whereas pivot language-based systems claim to be universal, i.e. applicable to any language.

Ш* A pivot language is a formal description of morphological, syn­tactic and semantic characteristics of a language unit in the form of one-to-one relationship. Each language unit is related to a specific invariable atom in the pivot language structure and vice versa, each atom of the pivot language structure is invariably connected with the units of various languages.

Ideally a pivot language-based machine translation will comprise the following processing steps:

Morphological, syntactic, and semantic analysis of the source text using information of the source language dictionary and paradigms.

Formation of the pivot language representation of a source text by the pivot language module.

Conversion of the source pivot language representation by the pivot language module into the target text using relevant semantic, syntactic, lexical, and morphological data from the target language dictionary and paradigms.

Usually pivot language formalism has the form of a graphic network or its analytical equivalent. It is, indeed, an extremely complicated system of morphological, syntactic and semantic entities and relations. We can hardly show even a part of it in this Manual because of space limitations, and even more because such a detailed presentation is well beyond the scope of this introductory course.

As a rule, linguistic science accepts only the three basic methods of machine translation mentioned above. Some scholars, however, main­tain that the artificial intelligence (Al)-based method should also be included in the classification despite the fact that it relies on encyclopae­dic rather than linguistic findings⁵².

Hence, to complete the picture it is worthwhile to give a short description of this method as well.

The main component of the AI translation model is its so-called «knowledge base». According to an AI-based translation model the re­sults of linguistic analyses at all language levels are verified against extral-inguistic information contained in the knowledge base.

Here we might want to refer back to our previous discussions. As you may remember, we discussed three factors that help us clarify natu­ral language ambiguities and make translation possible. These are con­text, situation, and background information.

In all three translation modelling methods that we have just dis­cussed, disambiguation is performed only with the help of context. It is rather simple lexical and syntactic context in direct translation model, and much more complicated syntactic and semantic context representa­tions in transfer-based and pivot language models. None of those mod­els, however, makes use of the other two disambiguation tools, i.e. situa­tion and background information.

•* In AI-based translation models disambiguation procedures are radically different and based first and most of all on the analysis of situation and background information (knowledge base), whereas purely linguistic context analysis methods serve only as secondary back-up tools.

The concept of a knowledge base is very similar to that of the subject thesaurus suggested by the communicational theory of translation (see Lectures 3 and 5). Both the knowledge base and subject thesaurus are presumed to contain a specifically arranged hierarchy of the facts of real world with the verbal information playing a subordinate role as labels for the facts and situations.

Another important component of AI-based translation simulation is a decision-making module, comprising a structural hierarchy of logical productions with probability estimations.

The present level of sophistication of AI-based translation modelling is rather ambiguous - on the one hand, results of the development of AI-models intended for translation as such are rather limited, on the other, however, the development of AI models intended for natural language interface, especially for expert systems, is very efficient⁵³.

This discussion of computer-based translation would be incomplete without mention of statistical methods.

•► In purely statistical methods of translation modelling it is pre­sumed that with certain probability each word of the target text may be a translation of each word of the source text

The various statistical models are different as to further probability estimations. For instance, a model suggested by P. Brown and co­workers estimates the probabilities of the word order matching in the source and target texts⁵⁴. Other models estimate the probabilities of word collocations in source and target texts, and so on.

At present statistical method of machine translation modelling is gaining new popularity because of two factors:

1. Virtually unlimited storage capacity and processing capabilities of new generation computers.

2. Availability of large bilingual text corpora in computer-accessible formats.

Besides, since we still know very little about how a translation is per­formed, statistical modelling is another attempt to learn more about translation.

It should be noted, however, that none of the machine translation methods appear in real systems in pure form. Very often it is rather diffi­cult to classify unambiguously whether one or another system is a typical representative of a particular basic translation simulation method. For this reason it is hardly worthwhile to analyse here any particular com­mercial machine translation package.

To the best of our knowledge all available commercial packages use that or another alternative of transfer model and the quality level of their products leaves much to desire. We hope that this Manual helped you to understand why this is the case.

Thus, completing our discussion of machine translation we must conclude that unfortunately available machine translation packages are not capable of providing adequate translation. However, it should be noted that such powerful tool as computer can still be used for transla­tion and this variety of translation is called computer-aided translation

Ш* Unlike machine translation, which by definition is wholly auto­matic, computer-aided translation is a tool to assist human transla­tion

In the process of computer-aided translation a translator is using a machine translation system (usually a direct translation variety) for the search of equivalents both for individual words and small text fragments. At the present stage of machine translation development computer-aided translation seems the most appropriate practical alternative.

Lecture 9. TRANSLATION OF OFFICIAL DOCUMENTS: TEXT FRAMES

A frame is understood as a set language structure with change­able elements

This Lecture:

underscores differences in translating official and literary texts; gives a brief outline of an official document framework; introduces the notions of:

• a frame,

• a slot,

• and a slot filler.

In a way translation of official documents is the opposite of literary translation in what concerns the freedom of translator's actions and choices.

As you may remember literary translation is more art than craft which is accounted for by the nature of literary texts. Translation of a literary text is unique and unprecedented; it cannot be standardised and obeys almost no rules.

Unlike literary texts the texts of official documents are highly stan­dardised: this applies both to the structure of the whole text (macrostructure) and to the arrangement of individual paragraphs and sentences (microstructure).

At the macro level an official document usually consists of a pream­ble, main text body and a finalising (concluding) part. Depending on the type of document the composition and content of its individual parts may slightly vary (see Appendix to this Manual). In modern linguistics standard text structures intended for information presentation at differ­ent language levels are called frames.

The changeable elements within a text frame are called slots. Let us take a preamble to a commercial agreement as an example of a frame.