| 4. |
But current systems which reflect the latest thinking in the Mt field are not the only ones actually used. The GAT(Georgetown Automatic Translation) research project began in 1952 and delivered its first fully viable package in 1964. That technologically primitive package, which translates Russian physics texts into what one may generously call "English", is still in use. Westerners who wish to monitor the state of physics research in Russian by reading translated abstracts and texts are willing to tolerate the low quality "English" that the GAT programme produces. Such low quality is less expensive for them than the waiting period and higher cost of carefully prepared human translations of the material. GAT was a direct attack, transfer method, local scope system. Even such a primitive system still finds applications today. |
| 5. |
To understand this, one, must appreciate the sheer enormity of the problem of the translating of extremely large quantities of informative texts and of what Newmark (see section 4.4, paragraph 6) calls vocative texts. One also needs to distinguish between the requirements of information acquisition (as in the case of American scientists needing to know what their Russian counter parts are doing), where high language quality in the output is not called for, and those of information dissemination, as in a company trying to describe its products attractively for foreign markets, where high quality is necessary. MT systems from an older period may be primitive but cheap and thus appropriate for part of the information acquisition market. |
| 6. |
As in the case of human translation, some of the theoretical debates come out of the experience of practical learning on the part of machine translators handling high volume problems and other real life difficulties. When the French set up in 1961 the CETA prohect (Centre d' Etudes pour la Traduction Automatique) at Grenoble, they hoped to learn from the difficulties of Gat. Trying to avoid the problems of a direct strategy, CETA went in for an interlingua method rather than a transfer method. The work of CETA and its successor, GETA (Groupe d'Etudes pour la Traduction Automatique), revealed the specific difficulties of the interlingua method. If the production of the TL output is entirely based on an interlingua representation of the text which suppresses all details of the SL original, then valuable surface lexical clues cannot be used when making sensitive TL word choices in order to produce stylistically acceptable equivalents in the output. Learning from this experience, CETA in its more recent work uses the transfer method and decentralizes the system's processes into many modules with specific tasks. This new work of CETA, according to a survey article by Jonathan Solcum(see references), is handicapped by the decision to write many of its crucial programmes in low level computer languages. Computational techniques evolve rapidly; systems who important software is in low level languages tie themselves too closely to a particular phase of such development and cannot take advantage of later advances in computer science. Thus the field of MT, like any other, encounters new difficulties as it solves old ones. |
| 7. |
Quite early in the MT enterprise, it became clear that no mechanical system would be able to deliver the optimal type of output, FAHQ(Fully Automatic High Quality) translations. Thus, researchers tried to find ways around the problem. The central problem is that natural language is full of ambiguities, which mechanical systems cannot resolve without human intervention. The question is what form this intervention should take, for various cases and in the context of various MT systems. |
| 8. |
One option is for the system to set down various readings in the output. A human post-editor then eliminates the inappropriate ones as part of cleaning up. This option is not a serious departure from the practices of mainstream human translation. Most organized translation outfits make it a point to have senior translators check and revise the work of the subordinates who write the first drafts. That machines should also need post-editing is not surprising. |
| 9. |
The other option is to organize human intervention in such a way that the machine produces clean output that calls for little or no post-editing. One form of such intervention is machine-human interaction, with the machine asking questions when it encounters ambiguities and a human editor giving appropriate a text to make it machine-processable. Or it may be a matter of people giving only particular kinds of text to the system: "one can restrict the grammar and the vocabulary of the input text in such a way that most of the ambiguity is eliminated. This is the sublanguage, or subworld, approach to MT says Nirenburg(see references), as this approach splits the world up into unambiguous subworlds that the system can cope with on a discipline by discipline basis. |
|
