Best Practices Guide — Computer-Aided Translation (CAT)
Vocabulary and processes explained
any years ago, before the advent and proliferation of computer-aided translation tools, a potential client called our firm and requested an estimate for a 300 page English to Hebrew translation needed by the end of the afternoon. Upon hearing the quote based upon industry standard rates, and the estimated turn-around time of four to six weeks, he protested, claiming that “All you guys are going to do is dump this into Word and click, ‘Save as Hebrew.’” At the time, this was cause for mirth amongst our project managers. No longer.
The problem is that, while Microsoft Word and other computer-aided translation applications can supply you with quick and easy translations, they are not generally accurate enough for anything but what is commonly referred to as “gisting”, which is the production of “fast and dirty” translations that give one the gist of what a text says, but which are full of grammatical and lexical mistakes, and misleading or ambiguous passages. Typically, such computer-aided translation results can only be decoded by someone who speaks, reads and writes both languages; in short, a translator.
The process of creating an accurate, publication grade of translation always requires some degree of human effort. Typically, unless the text re-uses material that has been translated and stored in computer memory, that human effort is paramount and comprises all but 100% of the intelligence that is brought to bear on a translation project.
Currently, there are three basic types of computer-aided translation tools available to assist the unavoidable human component of translation work.
Terminology Management refers to databases that hold preferred terms specific to an enterprise. Essentially, a terminology database is an intelligent glossary of a company’s preferred translations of specified terms in multiple languages. Typically, such a database can be used both in authoring original texts in one or more languages, and in customizing translation memory (see below) and machine translation implementations.
Translation Memory is perhaps the most utilized of all computer-aided translation tools at present. It is also referred to as machine-assisted human translation (MAHT) because, in all its various forms, it is a desktop tool used by professional translators. As the name implies, these applications “memorize” previously translated material, such that in subsequent translations they “suggest” translations of text strings (phrases, sentences or even entire paragraphs) in the target language to the translator. Translation memory when used appropriately supplies several benefits. It increases speed, accuracy and consistency among multiple linguists, streamlines the efforts of a translation team, and best of all, it saves money.
But it takes work to keep TMs up to date, and alas, if an error is introduced and validated, then the system will suggest the flawed text until it is detected and corrected.
Machine Translation is the most fully automated variety of the computer-aided translation applications. In its most basic form, one simply “dumps” a text into the program, selects a target language and clicks “Translate.” There is no actual machine other than a computer involved, and like all computer-aided translation tools,
MT tools are computer programs linked to databases full of translations created by human translators. MT tools work by either by parsing text segments and applying grammatical rules that are written by human beings (see Babelfish), or by statistical analysis of what the most likely translation is (see Google). Thus, this technology is continually evolving, and there is a significant difference between a highly customized MT implementation used by the automotive industry and others, and the less sophisticated products available to consumers and small businesses. In all instances, unless the output is carefully checked by a trained linguist who is a native speaker of the target language, there will be significant errors.
Generally, contemporary machine translation alone can supply 90% accuracy at best unless used in the context of tightly controlled language applications such as material safety data sheets.
The greater the breadth of the vocabulary and the complexity of style and grammar, the less suitable MT becomes. In the case of most business contexts, machine translation is not suitable unless followed by careful human editing.