draft 1995.April.13
Dr. Alan K. Melby
ATA Translation & Computers Committee
[Brigham Young University]
[Translation Research Group]
2129 JKHB,
Provo,
Utah 84602 USA
+1 801 378 4414 / +1 801 378 4649 (fax)
<trg@byu.edu> / 74511,1515 (CompuServe)
When a computer translates an entire document automatically and then presents it to a human, the process is called machine translation. When a human composes a translation, calling on a computer for assistance in specific tasks such as looking up specialized words and expressions in a dictionary, the process is called human translation.
There is a gray area between human and machine translation, in which the computer may retrieve whole sentences of previously translated text and make minor adjustments as needed. However, even in this gray area, each sentence was originally the result of either human translation or machine translation.
A basic requirement for machine translation is that the source text (i.e., the document to be translated) be available in machine-readable form. That is, it must come on diskette or cartridge or tape by modem and end up as a text file on your disk. A fax of the source text is not considered to be in machine-readable form, even if it is in a computer file. A fax in a computer file is only a graphical image of the text, and the computer does not know which dots compose the letter a
or the letter b
. Conversion of a source text on paper or in a graphical image file to machine-readable form using imaged character recognition (ICR) is not usually accurate enough to be used without human editing, and human editing is expensive. Thus, for machine translation to be appropriate, it is usually necessary to obtain the word processing or desktop publishing file from the organization that created the source text.
But this is only one of many requirements.
All translations projects have specifications. The problem is that they are seldom written down. Specifications tell how the source text is to be translated. One specification that is always given is what language to translate into. But that is insufficient. Should the format of the target text (i.e., the translation) be the same as that of the source text or different? Who is the intended audience for the target text? Does the level of language need to be adjusted? In technical translation, perhaps the most important specification is what equivalents to use for technical terms. Are there other target texts with which this translation should be consistent? What is the purpose of the translation? If the purpose is just to get a general idea of the content of the source text, then the specifications would include "indicative translation only." An indicative translation is usually for the benefit of one person rather than for publication and need not be a high-quality translation. Sometimes, an indicative translation will be used to decide whether or not to request a high-quality translation.
In many organizations, indicative translations are done using machine translation and high-quality translations are done using human translation. This fact reveals a basic difference between humans and computers. Humans, with proper study and practice, are good at producing high-quality translations but typically can only translate a few hundred words an hour, depending on the difficulty of the source text. Even with very familiar material, human translators are limited by how fast they can type or dictate their translations. Computers are good at producing low-quality translations very quickly. Some machine translation systems can translate tens of thousands of words an hour. But as they are "trained" by adding to their dictionaries and grammars, they reach a plateau where the quality of the output does not improve. By upgrading to a more powerful computer, the speed of translation improves but not the quality. By upgrading to a "more powerful" human translator, the quality of translation improves but not necessarily the speed. Here we have a classic case of a trade-off. You can have high speed or high quality but not both.
Indicative translation (high speed, low cost, but low quality) represents a new and growing market but does not substantially overlap with the existing market for translation. The existing market, variously estimated at 10,000,000,000 to 20,000,000,000 US dollars world-wide per year, is primarily for high-quality technical translation. If your specifications include low quality then machine translation is for you and you can stop reading right here. If, on the more likely hand, your specifications include high-quality, then it is not obvious that machine translation is appropriate. Here quality would be measured by whether the target text is grammatical, readable, understandable, and usable. Usability can be measured by selecting tasks, such as maintenance operations, which can be accomplished by a source-language reader with the help of the source text and seeing whether those same tasks can be performed by a target-language reader with the help of the target text. Such measurements are notoriously expensive, but a skilled reviewer can accurately predict usability simply by studying the source and target texts.
Translation requesters typically want the terminology in their translated documents to mesh closely with terminology in related documents. For example, a software company will want all revisions of a software manual to use the same terms as the original, to avoid confusing readers. Translation requesters should track all terminology relevant to a given document, and deliver that terminology to the translation provider along with specifications and source text.
Therefore, the specification of which terminology database to use (which is part of the second leg of the translation tripod) and the inclusion the actual terminology database (the third leg) are not factors in the choice of whether to use human or machine translation. Unless, of course, the terminology database does not yet exist, in which case the translation job should be delayed until the terminology database is ready. So far, the major factors have been whether the source text is available in machine-readable form and whether high-quality translation is needed. If the source text is available in machine-readable form and high-quality translation is required, then the deciding factor is the nature of the source text.
Skilled human translators are able to adapt to various kinds of source text. Some translators can even accomodate poorly written source texts and produce translations that exceed the quality of the original. However, current machine translation systems strictly adhere to the principle of "garbage in -- garbage out." Therefore, if high quality translation is needed yet the source text is poorly written, forget about machine translation. There is more. Machine translation systems cannot currently produce high-quality translations of general-language texts even when well written. It is well-known within the field of machine translation that current systems can only produce high-quality translations when the source text is restricted to a narrow domain of knowledge and, furthermore, conforms to some sublanguage. A sublanguage is restricted not just in vocabulary and domain but also in syntax and metaphor. Only certain grammatical constructions are allowed and metaphors must be of the frozen variety rather than dynamic. Naturally occurring sublanguages are rather rare, so the current trend is toward what is called "controlled language." A controlled language is almost an artificial language. Rules of style are set up to reduce ambiguity and to avoid known problems for the machine translation system. This leads to another requirement concerning the nature of the source text: There must be lots of it. It is cheap to set up a machine translation system to produce indicative translation. It is expensive to develop a document production chain that includes high-quality machine translation. Therefore, for such a document chain to be cost-effective, there must be a large quantity of similar text in the same sublanguage going into the same target language or languages.
Now it should be clear why less than ten percent of what is translated is appropriate for machine translation. To qualify, a text must be available in machine readable form and, unless low quality output is acceptable, the same text must be voluminous and be restricted to a single sublanguage.
If a computer (or a human) is only allowed to the word "cut" and the rest of the sentence is covered up [Figure 2: 19k GIF], it is impossible to know which meaning of "cut" is intended [Figure 3: 24k GIF]. This may not matter if everything stays in English, but when the sentence is translated into another language, it is unlikely that the various meanings of "cut" will all be translated the same way. We call this property of languages "asymmetry".
We will illustrate an asymmetry between English and French with the word "bank." The principal translation of the French word banque (a financial institution) is the English word "bank." If banque and "bank" were symmetrical then "bank" would always translate back into French as banque. However, this is not the case. "Bank" can also translate into French as rive, when it refers to the edge of a river [Figure 4: 66k GIF]. Now you may object that this is unfair because the meaning of "bank" was allowed to shift. But a computer does not deal with meaning, it deals with sequences of letters, and both meanings, the financial institution one and the edge of a river one, consist of the same four letters, even though they are different words in French. Thus English and French are asymmetrical.
Early researchers in machine translation (in the late 1940s and early 1950s) were already aware of the problem of asymmetry between languages, but they seriously underestimated the difficulty of overcoming it. They assumed that by giving the computer access to a few words of context on either side of the word in question the computer could figure out which meaning was intended and then translate it properly. By about 1960, some researchers had realized that even if the entire sentence is available, it is still not always obvious how to translate without using knowledge about the real world. A classic sentence that illustrates this difficulty uses the word "pen," which can refer to either a writing instrument or to an enclosure in which a child is placed to play so that it will not crawl off into another room. The ambiguity must be resolved or the word "pen" will probably be translated incorrectly.
This sentence will typically be interpreted by a human as referring to a writing instrument inside a cardboard box [Figure 5: 52k GIF], such as a gift box for a nice fountain pen or gold-plated ballpoint pen, rather than a play pen in a big box. However, look what happens if the sentence is rearranged as follows:
This sentence will typically be interpreted by a human as referring to a normal-size cardboard box inside a child's play pen [Figure 6: 81k GIF] rather than as a tiny box inside a writing instrument. A human uses knowledge about typical and relative sizes of objects in the real world to interpret sentences. For a human, this process is nearly effortless and usually unconscious. For a computer that does not have access to real-world knowledge, this process is impossible.
The situation is also taken into account. Returning to the sentence about the pen in the box, there are texts, such as a description of a family with small children moving their affairs to another apartment, in which a human would interpret the pen as the child's play pen being put into a large box to protect it while it is moved to a new location. And there are texts, such as a spy story about ultra-miniature boxes of top secret information, in which the sentence about the box in the pen would be interpreted as referring to a writing instrument containing a tiny box. The words in these sentences do not change, yet the interpretation changes. Here even real-world knowledge is insufficient. Some sense of the flow of discourse and the current situation are needed.
Let us consider an example of how restriction to a sublanguage within a domain can simplify the translation process. Consider the word "bus." Without restricting the source text in any way, this word could refer to either a large vehicle for transporting people or to a component of a computer that consists of slots into which circuit cards are placed [Figure 7: 52k GIF]. However, if the source text is known to consist of a sublanguage which is concerned uniquely with instructions on how to repair microcomputers, then the word "bus" will almost certainly refer to the slots for circuit cards rather than to the vehicle.
There is a connection between the sublanguage approach and the Artificial Intelligence approach. The Artificial Intelligence approach has until now only been successful on sublanguage texts. Artificial Intelligence has been by far the most successful on sublanguage texts limited to extremely narrow domains known as microworlds. An classic example of a microworld is limited to a certain type of kitchen water faucet and the task of replacing a washer in that faucet. So far as that microworld is concerned, the fridge does not exist, nor the stove, nor any other part of the kitchen, and the other rooms of the house, such as the bedroom cannot be mentioned. It is not surprising that such restrictions are helpful to the computer programmer who is designing a system to process texts in various ways. The really interesting question is whether these restrictions can be overcome. Can artificial intelligence or any other computer-based processing ever work on general language?
To attempt an answer to this fundamental question, we must look at the assumptions behind the computer processing. Many linguistic theories and approaches to Artificial Intelligence are based on what has been called objectivism. Hold onto your hats. The next section is seriously philosophical. I realize this is probably not what you wanted when you started reading. You probably just wanted to know which brand of machine translation software to buy and find assurance that it works. Sorry for the cold shower.
According to objectivism, real-world knowledge and the situation and flow of discourse can be safely ignored until the possible meanings of a sentence have been computed. Then the appropriate meaning is selected.
One problem with objectivism is that it assumes that the world divides itself up exactly one way into categories independently from how humans view the world, and that we then associate words with these pre-existing categories. However, there is not just one way to divide up the world. To take an extremely mundane example, there is not just one way for a butcher to divide up a beef. There are different cuts of meat in various countries. [Figure 8: 14k GIF] shows how beef is butchered in the United States and in Switzerland).
But this question of categorization goes much deeper than what you find in a grocery store. It pervades every aspect of our thinking, and it is dynamic. A few years ago, there was a television commercial in the United States extolling the size of the beef patty in the hamburgers from one chain of fast food restaurants as compared with the size of the patties used in the hamburgers sold by a competing chain of restaurants. In this commercial was the phrase:
This sentence took on a metaphorical meaning of challenging whether some project had produced sufficient visible results, even if it had nothing to do with beef or even with food. General language is full of such dynamic metaphor. We as humans are not usually even aware of minor shifts in meaning, because we are capable of handling them. Indeed, they give spice to language and are necessary to true creativity. However, dynamic metaphor is contrary to the assumptions of objectivism, since metaphorical meanings cannot be computed step-by-step.
All current approaches to processing language, including all commercial machine translation systems, are based on objectivism, whether the designers are aware of this fact or not. Indeed, objectivism has been so entrenched in the thinking of the Western world for hundreds of years that only recently are philosophers becoming aware of objectivism and considering alternatives. Unfortunately for machine translation, no one has yet conceived of a way to program a non-objectivist approach to language on a computer. This is why we can say that machine translation will not deal effectively with general language in the foreseeable future. Humans are able to deal with language without the constraints of objectivism. However, at this point, no one can foresee if or when computers will be able to deal with human language in a non-objectivist way.
Think of a microworld as a tree-house. Suppose two boys are sitting in a tree-house and suddenly are victims of severe amnesia. They could look around the tree-house for a long time and never be aware of an outside world even though there is a beam of light shining in from the outside [Figure 9: 119k GIF]. It is one thing to look at the beam of light, but it is an entirely different experience to look along the beam of light and through the hole so that you can see the outside world. The tree-house can be compared with a microworld. Currently, computers are based on assumptions that make it impossible to look along the beam of light and so they are restricted by the walls of some particular microworld.
Maybe some day, computers will be built that can look along the beam. In the meantime, computers are successful in dealing with human language according to how successfully language can be restricted to a microworld in which language does behave as if objectivism were an adequate portrayal of the nature of the world. The sublanguage used in a microworld must be carefully controlled.
If neither condition holds for you, machine translation is not for you.
Please let us know about your machine translation adventures in the philosophical world of our book or in the brutal real world.