{Utrecht University}

Towards New Metaphors
Steven Krauwer

{ELSNET}

Machine Translation: Towards New Metaphors

Steven Krauwer


1. Is Machine Translation possible?

The question whether Machine Translation (MT) is possible, is a
very delicate one for a researcher to answer. If the answer is
'yes' we need evidence that this is indeed the case, i.e. that
there are Machine Translation systems out in the world that can
produce adequate translations, or at least some sort of evidence
that this will be the case in the foreseeable future. If, on the
other hand, the answer is 'no', a large number of academic and
industrial researchers will have to justify the millions they
have spent over the last 40 years, and are still spending, on the
MT enterprise.

Let us first have a closer look at the answer, and then at the question. 1.1. The answer Those who have ever been confronted with the unedited output of an MT system, will easily admit that the result contained errors, ranging from small, stylistic deviations, to an absolute distortion of the meaning of the text in the best case and absolute nonsense in the worst case. Some of the errors may look rather superficial, and easy to repair in a next version of the system (e.g. unknown words), but others might be much more difficult to tackle. Let us take the following two sentences (inspired by an example from Jerry Hobbs) as our starting point: (a) The policemen fired at the demonstrating students because they FEARED violence. (b) The policemen fired at the demonstrating students because they WANTED revolution. From a purely syntactic and lexical point of view the two sentences are almost identical: the underlying syntactic structure is the same, and the only real change is that 'feared' in sentence (a) has been replaced by 'wanted' in sentence (b). But there is a difference in interpretation, which goes further than just changing the verb. In sentence (a) most native speakers would interpret 'they' as to refer to the policemen, whereas in sentence (b) 'they' refers to the demonstrating students. The reason is obvious: we have a common understanding of the roles of policemen and students in society (at least in our part of the world), and there policemen are supposed to protect us against dramatic changes in our way of living, such as revolutions, whereas students have the reputation to be amongst the first ones to adopt a critical attitude, and to advocate changes. No human translator would ever have any difficulty in assigning the correct interpretation and translation to sentences (a) and (b), whereas it is difficult to imagine how one could set up an MT system in such a way that it would make the right choice here. Linguistic knowledge, however deep, will not give any clue, and the sort of knowledge that is needed, is very hard to encode and apply in any systematic way. Some may want to argue that there is no need for the MT system to resolve the anaphoric ambiguity here, as e.g. translation into French would not depend on the choice made here, but it is easy to see that if we replace 'policemen' by 'police women' there is no way to get around the problem. The necessary conclusion is that however far we get in building MT systems, we will never be able to guarantee that we get it right. 1.2. The question Answering the question we asked above, turns out to bring us in a slightly difficult position, so it might be wiser to ask ourselves whether the question itself is a reasonable one to ask. Traditionally, the objective of any MT effort has been to provide a possibly perfect, cheaper and faster imitation of the human translator. Success was measured on the basis of the extent to which this simulation was successful, and as we all know, this success has never been very impressive. As an alternative way of looking at things, we can ask ourselves whether the simulation of the human translator is a meaningful purpose in itself. And the answer is a clear no. Although it may be an interesting intellectual challenge, the ultimate goal of any translation activity is to solve a communication problem, where two parties who want to exchange information happen to speak different languages. From this point of view we can describe the purpose of an MT system as a solution for a communication problem. This allows us to define the notion 'success of a translation system' as the degree to which the intended communication has been made successful. 2. A new metaphor We propose to adopt a new metaphor, instead of the human translator, as the starting point for the construction and the use of MT systems. Our metaphor is "the traveling tourist and his host". We all know from our travels abroad that communication in a different language community poses problems, which can sometimes be solved by one of the parties (e.g. one person decides to speak the other person's language), or sometimes by cooperative action. This leads us to the introduction of two types of tourists (or guests) and hosts: adaptive and unadaptive. The adaptive party is the party that accepts that communication across languages is problematic, and makes a special effort to facilitate communication. The unadaptive party is not prepared to compromise, and insists on using his own language the way he would do at home. We can now make the following grid, based on the two parties' adaptiveness, and investigate what sort of solutions MT or in a broader sense language and speech technology could offer: +-----------------+------------------+ | adaptive host | adaptive host | | adaptive guest | unadaptive guest | +-----------------+------------------+ | unadaptive host | unadaptive host | | adaptive guest | unadaptive guest | +-----------------+------------------+ 3. A war or a series of battles The traveling tourist metaphor breaks down the MT problem into at least four subproblems, the solution of which may require different types of approaches and technologies. It may look less elegant to have four different solutions to four different instantiations of the MT problem, rather than one magic formula that does all the work, but on the other hand one should realize that just like in e.g. transportation (where airplanes are faster than bikes, but more difficult to park in cities), in communication every situation is different, and may have different constraints and success criteria. Therefore, even within the four boxes of the grid one may have to further subdivide the problem into subproblems of a size and complexity that we can manage. The MT problem should not be seen as a single war, aimed at solving the MT problem in one go, but rather as a series of local battles, each of which should serve to reduce, in specific situations, the communication problems caused by the fact that people speak different languages. 3.1. The adaptive host and the adaptive guest If both parties are equally adaptive, the chances to get the communication problem solved, are by far the greatest. A typical solution is for both parties to resort to a third language, known by both. The use of English as a lingua franca in tourism, international transportation and in science is a very well known example. It is important to note that, although neither party may have a perfect command of the foreign language, the communication is normally successful, even if it may be less efficient than it would have been between two native speakers of the language (although even there no 100% guarantee for success can be given). There is still a role for language and speech technology here, such as language learning facilities, or electronic dictionaries and foreign language authoring aids. If none of the parties has a good command of the common language, other facilities may come into play. An excellent example is offered by the Verbmobil project, where a German and a Japanese speaker use their own language to make meeting arrangements via the phone, and where English is used as an interlingua to sort out problems. 3.2. The adaptive guest and the unadaptive host This is a very normal situation for e.g. tourists who speak a minority language, and who are traveling in a country where either the native language belongs to the majority languages, or where most people have received no foreign language education. If the tourist has some managed to acquire some moderate knowledge of the language, he might still be able to communicate successfully, even if some of the details of the communication may escape him. In such a situation a small electronic dictionary might be of help. If the language gap is wider, more sophisticated facilities are needed. An interesting example is the DIPLOMAT system, designed to help UN soldiers in Bosnia to communicate with the local population, which helps asking questions in a foreign language, and translating the answers. For electronic travel obvious (and already existing) facilities are e.g. a translating browser (even if the translations are not perfect, cf the way AltaVista makes use of the SYSTRAN translation system), or an information extractor, which extracts the key information from a foreign language text and presents it in the user's own language. 3.3. The unadaptive guest and the adaptive host. Typically this applies to situations where mass tourism moves people who may have received no foreign language training to places where they can enjoy sun, sea, food and drinks, and who don't want to be confronted with the language barriers. Hosts who want to please their guests have to be adaptive, e.g. by learning their language, or by employing staff with the right native language. From a language and speech technology point of view, obvious facilities to support this are the use of controlled languages in authoring messages from host to guests (as long as the author sticks to the rules, error-free translation can be guaranteed), or by means of multilingual generation from tabular information (e.g. generating weather or avalanche reports in various languages from tables). 3.4. The unadaptive guest and the unadaptive host. This is a fairly extreme situation, and it is clear that ideally one would have to resort to either e human translator or interpreter, or to Fully Automated High Quality Machine Translation -- if only it existed. As a long term research topic, this should certainly be kept high on the agenda, together with the interpreting phone which would ideally allow me to speak my own language over the phone, and to be heard by my Japanese conversation partner as speaking fluent Japanese, in my own voice. For the shorter term the best option seems to be to develop tools that increase the productivity of human translators, or the quality of their output. Typical examples are electronic term banks, on-line dictionaries and thesauri, or translation memories. 4. Concluding remarks First of all it should be clear that MT is not one single simulation problem, but a complex of communication problems, each of which may require a different approach and solution. Each of the subproblems can (and even: should) be further subdivided. The adaptivity grid we have presented here, can also help us in raising new questions, such as "is there a tool for an adaptive guest who wants to express himself in a foreign language unknown to him?". Note that an affirmative answer to such a question will not automatically imply that there will also be a market for the tool. In the spirit of the above, it may not come as a surprise that for the future we would like to advocate the development of domain dependent, goal dependent, and discourse dependent MT facilities. In order to exploit the fact that these systems are constrained, it is important that they should not only be based on what we know about language (the rules, which express what we understand), but also on what we know about how language is used in specific communication contexts (i.e. statistical data which we can measure and extrapolate from).


Steven Krauwer (s.krauwer@uu.nl) Utrecht Institute of Linguistics UiL OTS
Phone +31 30 253 6050 Faculty of Humanities, Utrecht University
[Page last modified: 02-01-2014] Drift 10, 3512 BS Utrecht, Netherlands