Bernard E. Scott
Computers stand up well to a grand master when it comes to the logic of chess, but they can't match the skills of a 7-year-old when it comes to language.
The reason for the glacial pace of progress in MT (machine translation) over the past four decades can be found in one factor: the intractable ambiguities of natural language. An MT system must peel away at least five layers of ambiguity before it is able to map sentences from one language to another with any degree of accuracy (see figure A). If you understand how MT copes with these difficulties, you will have a clear idea of just how these systems work and why they do not work better.
Step Inside the Beasty
Step inside an MT system and see how it handles the following simple sentence: The heavy-duty truck turned into a driveway. As you follow this sentence through the system, notice how nearly every other word poses a challenge--and an opportunity for error.
When the system looks in the dictionary for the word truck, it immediately encounters ambiguity: The word is encoded in the dictionary as both a noun and a verb. The system's dictionary can tell you only that truck can take the form of two parts of speech. It can't tell you which form it takes in this sentence.
To make that determination, you must move further into the system and view the word in the context of the sentence. At this point, the system still has no idea what the sentence means and sees it only as a syntactic string containing elements that have more than one interpretation.
To operate at the sentence level of the syntactic stage, you must have some kind of grammar--typically stored as a set of rules. One of these rules will determine that, in the given sentence, without violating grammatical rules, truck cannot be anything but a noun. So far, so good-- although it is not always going to be that easy.
Now that you know that heavy-duty truck is a noun phrase, a second layer of ambiguity comes to light. The system still sees your noun phrase purely syntactically, as the string Adj N1 N2. It has no idea, for example, whether the adjective heavy modifies duty or truck. The system has to resolve this ambiguity if it's to get the agreement right. Therefore, you have to go beyond syntax into lexical semantics.
The Next Stage
At this deeper stage, more intelligent rules come into play and use the semantic properties that were retrieved for the words earlier in your sentence during the dictionary lookup stage. These semantic- property codes are designed to resolve ambiguities such as that posed by heavy- duty truck. Now you're going to run into some rough going.
The majority of low-end MT systems don't get into semantics--or they do so only in trivial ways. These systems generally are weak, but even high-end systems will have trouble trying to figure out which noun heavy modifies. The issues are subtle. At this point, most developers will resort to brute force by storing the phrase as a unit in the dictionary.
Slightly more tractable examples of this kind of ambiguity would be old people and children and smart girls and boys. If a smart rule uses a test for semantic symmetry (or lack thereof) among the noun pairs, it could figure out that the adjective old modifies only people and that smart modifies both boys and girls. Clearly, getting a machine to cope with this challenge isn't easy.
Processing at the lexical semantic stage introduces its own kind of confusion--the third layer--having to do with multiple meanings of words. For example, the verb turn into has at least two lexical meanings: One is the sense of motion, and the other is the sense of becoming. To decide which meaning applies in your sentence, you have to move to sentence- level semantics, where the verb turn into can be examined in its semantic context.
A semantic rule associated with the words turn into would know that the meaning of this verb is going to be a function of the verb's direct object. So, in this sentence, the rule has to test only the semantic- property code for driveway to determine the verb's meaning: If driveway were given a semantic-property code signifying a path, the rule would know to select the verb's motional sense. Such a rule would work with Cinderella, too, if her carriage turned into a driveway rather than a pumpkin.
Going for the Gold
A fourth layer of difficulty has to do with ambiguities introduced at the sentence level of the semantic stage. Unfortunately (or fortunately), the sample sentence doesn't illustrate this kind of complexity. But to get the idea, consider the meaning of the preposition for in the following sentences: Check the newspapers for errors. Check the newspapers for dates.
In the first sentence, the preposition for signifies for the presence of, and in the second sentence, it means for information about. As used in this example, in a language like Vietnamese, the preposition for would be expressed differently in each case.
Thus, the system has to determine which case applies if it's to translate the meaning correctly. You can see that the meanings of the word for are a function of the sentence as a whole; you won't find them in any dictionary. Also, notice how the sentence as a whole affects the meaning of the verb check. In the first sentence, check means to examine. In the second, check means to consult.
A fifth layer of ambiguity concerns more technical issues, such as ellipses and anaphora (e.g., antecedents of pronouns). This level of translation sophistication calls for processing at a discourse level--something few systems can do.
But don't worry. MT systems are steadily becoming more robust. The key is semantics. As the semantics of these systems improve, so will the power of the systems to dispel ambiguity and, in the final analysis, to translate.
Figure A: MT systems are composed of one or more levels of parsing sophistication (going from syntactic to semantic). A simulated, advanced, multilevel system such as this one might determine the meaning of each word or phrase in the sentence (i.e., parse the sentence) by posing and/or resolving questions, eliminating ambiguity as the words or phrases move through the various levels in the system.
Bernard E. Scott is the head of his own consulting agency, Parse International (Ledgewood, NJ). Previously, he founded Logos (Dedham, MA) and was the principal architect of the Logos MT system. You can reach him on BIX c/o "editors."
Per comentaris i observacions sobre el servidor, poseu-vos en contacte amb l'Administrador WWW