ClaasColoplastContinentalElectroluxMärklinMANOTTOPriceWaterhouseCoopersSIXTStarbucksZDF

Machine translation

Machine translation (MT) refers to the translation of text from the original language (or source language) into the target language by means of a computer programme. MT is an example of artificial computer intelligence.

Whilst human translation forms part of practical linguistics, MT is researched in the fields of IT and computer linguistics. MT programmes were even being written for the first computers in the forties.

Dream of mankind

Understanding a language, without having learnt it, is one of the oldest dreams of mankind (Tower of Babel, J.Becher’s numeric Interlingua, Babel fish, Pentecost, science-fiction stories). The invention of the computer, combined with the phenomenon of language as a scientific discipline, has for the first time opened up concrete opportunities, to fulfil this dream.

History

Up to this day, military interests have decisively shaped the development of MT. One of the earliest projects was a Russian-English translation programme for the US military. Despite its poor, anecdotal quality the programme enjoyed popularity in the US military, who could for the first time establish an impression of the content of Russian documents without going via third parties (interpreters and translators).

The ALPAC report produced for the Pentagon in 1966 contested the feasibility of MT and with one blow brought research to a practical standstill for almost 20 years. Only in the eighties did electrical companies such as Siemens AG (Metal project) begin research again. At the same time, the Japanese government launched the Fifth Generation Project, where MT from English into Japanese was initially based on Prolog programming language. The close collaboration between universities, electrical companies and government led to the first world-wide commercial MT programmes for PCs and placed Japan in the leading position world-wide in MT research. In the nineties the BMBF’s key project, Verbmobil, was run in Germany, which aimed to interpret spoken dialogue in German, English and Japanese. The Verbmobil system should recognise spontaneous speech, analyse the entry, translate, create a sentence and say it out loud.

The dotcom crash in 2000 and 2001 ruined many small MT companies. In today’s MT software industry there are now only an estimated 10 to 20 companies still active (many programmes are licensed, so as not to give the false impression of a larger variety), meaning that MT is for the most part being developed at universities.

At present, only about 1% of the total revenue on the translation market is due to MT applications.

However, there are several reasons for the increased demand for MT applications:

  • Many texts are now available digitally (therefore easy for the computer to work on).
  • Globalisation calls for the translation of more and more texts into more and more languages (the translation market doubles every four years), whilst the popularity of the translator/interpreter profession is stagnating.
  • Languages, that are difficult to learn and seldom spoken by Western Europeans/Americans, whose speakers for their part can hardly speak Western languages, are becoming more and more important:
    • Commercially important: East-Asian languages Chinese, Korean and Japanese, as well as Thai.
    • Militarily important: Languages in international regions of conflict, primarily those with involvement from US military. In 2003 several US software companies brought out translation programmes for Arabic and Paschtu (language in Afghanistan). Likewise the DARPA held a blind competition in 2003 for an unknown source language.

Translation Methods

All MT systems use bilingual dictionaries and have modules for at least the basic grammar rules. Nevertheless, the individual methods vary considerably.

The most important MT methods/approaches are:

  • Direct MT. The words of the source text are translated word-for-word and in the same order by the dictionary into the target language. Afterwards, word order and inflection are adjusted according to the rules of the target language. This is the oldest and most simple MT method, which formed the basis of the Russian-English system mentioned above.
  • Transfer. The transfer method is the classic MT method in three steps: analysis, transfer and generation. The second step gave the method its name. First the grammatical structure of the source text is analysed, often in a tree structure. A semantic structure is derived from this analysis, depending on the chosen transfer method. Then the structures are transferred into the target language. Finally, sentences are created again from the structures with the grammatical rules, thus generating the target text.
  • Interlingua. The grammatical information from the source text is firstly expressed in a neutral middle language or interlingua. The grammatical information in the target language is created from this interlingua. The Interlingua method is useful for translating complex expressions. For example, you cannot translate the German “Wenn ich arbeiten würde, würde ich mir ein Auto kaufen.” with the Transfer rule würde-->would (“If I would work, I would buy a car.”), because in English if-sentence do not allow would. In Interlingua, the würde information would be recognised as an “unreal conditional” and translated with or without would, depending on the context of the sentence.
  • EBMT (stands for Example-Based Machine Translation). The core application of an EBMT system is a translation memory, where frequently recurring sentences or phrases can be saved to memory with the respective translation. It statistically calculates (via information retrieval methods) how similar all the translation memory entries are to the respective sentence in the source text. The translation is generated from a combination of the most similar sentences translated.
  • SBMT (stands for Statistics-Based Machine Translation). Before the actual translation, a programme analyses the largest and most widely varied corpus of bilingual texts (for example, parliament agendas). This orders words and grammatical forms in the source and target language according to their frequency and proximity to one another, therefore generating a dictionary as well as grammatical rules. Texts are translated on this basis. SBMT has recently become very popular, because it does not require any knowledge of the languages concerned. One subsequent advantage of SBMT is that rules, which are not precisely explained in terms of linguistics, can be theoretically included, by analysing real text extracts. The translation quality is admittedly a lot worse than that produced by existing rule-based systems, partially because statistics-based MT is comparatively new. For example, SBMT is favoured by the Pentagon for languages, which quickly require an MT system, without having the time for compilation of rules by people.
  • HAMT (stands for Human-Aided Machine Translation). Instead of leaving the computer to translate the whole document, the user is asked to translate or avoid ambiguous or complicated constructions (so called controlled language). This can take place in advance, with the user dividing up long sentences into shorter ones for example, or by interaction, for example the programme asks the user to select the intended meaning of a word.

In practice, most systems are a mixture of several methods (often dominated by transfer systems with Interlingua and EBMT elements).

MAHT (Machine-Aided Human Translation), that is computer aided translation where a computer programme supports the human translator by automatically checking terminology (automatic dictionary look-up) and comparing earlier translations (translation memory), does not count as machine translation.

Quality

Results from MT programmes are often unintentionally amusing. The effect is easy to see: simply take any text and enter it into a free translation machine to translate into your mother tongue.

How can you evaluate MT quality?

Instead of the intuitive and non-compelling impression “this translation is abysmal”, MT researchers employ scaled evaluations of translation quality. MT translations are evaluated per sentence; the standardised total of the sentences is the quality of the whole text. In most cases, the evaluation is carried out by a native speaker of the target language and expressed as an index. In Japan, a 5 digit scale with 0-4 points is often used:

  • 4 points: Very comprehensive, almost perfect; no obvious errors.
  • 3 points: One to two incorrect words; otherwise comprehensive.
  • 2 points: Possible to work out, what was originally meant.
  • 1 point: The sentence does not match the intended meaning (if at all). This is often due to a partially or completely incorrect translation of the grammar.
  • 0 points: The sentence makes no sense; looks like an accidental, thrown-together, chaotic arrangement of words.

For long translations, MT researchers also use automatic evaluation algorithms like the BLEU-Score, which is also based on underlying human powers of judgement.

Expectations too high?

Another problem facing MT may be that people simply have too high expectations. As a result, the actual improvements in MT research appear unsatisfactory. One of the conditions for a functional MT is that the source text is intelligible and that humans could also complete a fully detailed translation of the text. How can a computer be expected to understand and translate language that is not understood by another human being? Most linguists assume that the complete understanding of language implies the complete understanding of human intelligence. Some people are also of the opinion that a perfect MT system should simulate the processes of human intellect. As mentioned above, one of the advantages of SBMT is that this problem is dealt with, because in theory as yet unexplained rules can be deduced.

Practical problems

There are also tangible and partly remediable reasons as to why MT quality is often found to be unsatisfactory:

  • “Cheap programmes”. Many people judge the state of MT based of free MT tools, which are available on the internet on Yahoo! or Google for example. These are often pared-down or older versions of otherwise fee-paying (and better) programmes, or just fast (and poor) programmes.
  • Users understand the source language. Particularly in translations between Western languages, the user often understands the source language to a certain degree himself and is therefore more sensitive to discrepancies than someone who is solely reliant on the translation.
  • Language style. Each style of language has its peculiarities, which are not described by linguistics. MT systems tend to be based on written newspaper language. MT systems deliver particularly bad results when translating texts, for which they were not developed, therefore mostly literary texts, spoken language or occasionally technical texts (for example in notorious machine translated instruction manuals from Japan).
  • Lacking multidiscipline. MT is an area of computer linguistics, but most researchers come from one or two original disciplines in this field, either computer science or linguistics, without having sufficient knowledge to be suited to the other respective discipline. Whilst linguists often lack the programming practice, computer scientists often lack the readiness to work with the phenomenon of language. For this reason, a structural language model forms the basis of most MT applications, which does not take into consideration the findings in linguistics in the last 50 years.
  • No exchange between industry and academics. Commercial MT companies often prefer to employ pure programmers, who possess knowledge “on-site”, rather than MT researchers from universities, who give off the impression of being too set on unimportant details.
  • Too small and/or inaccurate dictionary. With the changes in society and science, a language’s vocabulary increases rapidly by the day. Furthermore, many words have multiple meanings (homonyms), which can be differentiated by contextual analysis. Dictionary shortcomings such as in Russian-English are responsible for a surprising percentage of poor translation quality. The larger MT programmes have dictionaries with several million entries and a variety of different meanings. The detailed and error-free compilation of such large dictionaries for MT applications through lexicography is just too great an expenditure for small companies.
  • Lacking Transfer rules. Many grammatical phenomena are strongly distinguishable from language to language, or only present in certain languages. Solving these problems often requires linguistic research; MT companies of course seek to avoid these costs.
  • Computational linguistic problems. MT also encounters many problems, which also occur with other computational linguistic applications, e.g. understanding encyclopaedic knowledge.