How exactly does a machine learn to translate?
In general, the learning process of a machine is not too different from that of a human. It is associated with numerous failures and a kind of » teacher » who continuously evaluates and corrects the respective performance.
In the field of machine translation, the learning process basically starts with setting up a machine translation engine. Since MT systems are generally not programmed to determine relevant data independently, various data are imported manually into the «corpus» (the basis of an MT engine). Company-specific translation memories (TM), terminology, glossaries or other multilingual reference documents are particularly suitable for this purpose. Alternatively, prefabricated MT Engines can be customized.
Once the foundation has been laid, training of the MT Engine can begin. For this purpose, a considerable number of test runs with corresponding correction loops are started. During each test run, the quality of the results achieved is determined and the unsatisfactory segments are edited before they are run through another test using specific training materials. This process is repeated until the delivered results have reached a certain quality standard. In addition to the assessment of an experienced posteditior, quality is primarily based on three benchmarks: the Bleu Score, TAR Score and F-Measure.
If the MT Engine’s performance in all three metrics meets the desired threshold, it is ready to be productively integrated into the translation process.
However, the learning process of an MT Engine or NMT does not simply stop after the «training camp» is completed. Just like with a human being, learning goes on and on, because new information enters the system with every imported translation. In this way, a continuous improvement process is set in motion, which primarily affects the quality of the results.