Automating Knowledge Acquisition for Machine Translation
How can we write a computer program to translate an English sentence into Japanese? Anyone who has taken a graduate-level course in Artificial Intelligence knows the answer. First, compute the meaning of the English sentence. That is, convert it into logic or your favorite knowledge representation language. This conversion process will appeal to a dictionary, which maps words (like "canyon") onto concepts (like CANYON), and to a world model that contains facts about reality (like "canyons don't fly"). In this way, an ambiguous sentence like "John saw the Grand Canyon flying to New York" gets the right interpretation. Finally, turn the conceptual structure into Japanese (or whatever), using further grammatical and lexical knowledge bases.
Along the way, there will be many fascinating problems to solve. Like: canyons don't "fly", but do people "fly"? Only in the sense of RIDE-IN-AIRPLANE, with the caveat that the WHEELS of the AIRPLANE must at some point leave the GROUND|otherwise, we're just taxiing. How about "John flew me to New York"? That's another meaning of "fly," involving DRIVE-AIRPLANE as well as RIDE-IN-AIRPLANE. And if "United flew me to New York," I may say that the AIRPLANE that I rode in was driven by an EMPLOYEE of the AIRLINE that OWNS the AIRPLANE. And while we're at it, why don't canyons fly? AIRPLANEs and CANYONs are both inanimate, but a CANYON seems too big to fly, or anyway not aerodynamic enough . . . We seem to be on the right track, but considering the vastness of human language and the intricacies of meaning, we're in for a very long journey.
Meanwhile, in the real world (not the formal model), people are buying shrink-wrapped machine translation (MT) software for fifty dollars. Email programs ship with language translation capacity (optional). Companies use MT to translate manuals and track revisions. MT products help governments to translate web pages and other net-traffic.
What's happening here? Is AI irrelevant? No, but there are many approaches to MT, and not all of them use formal semantic representations. (I'll describe some in this article.) This should come as no surprise, because MT pre-dates AI, as a field. An AI scientist could easily spend two months representing "John saw the Grand Canyon flying to New York," while anybody with a bilingual dictionary can build a general-purpose word-for-word translator in a day. With the right language pair, and no small amount of luck, word-for-word results may be intelligible|"John vi el Grand Canyon volando a New York." That's okay Spanish. But most of the time the translations will be terrible, which is why MT researchers are very busy:
ffl Building high-quality semantics-based MT systems in circumscribed domains, like weather reports [Chandioux and Grimaila, 1996] and heavy equipment manuals [Nyberg and Mitamura, 1992].
ffl Abandoning automatic MT, and building software to assist human translators instead [Isabelle et al., 1993; Dagan and Church, 1994; Macklovitch, 1994].
ffl Developing automatic knowledge acquisition techniques for improving general-purpose MT [Brown et al., 1993b; Yamron et al., 1994; Knight et al., 1995].
There have been exciting recent developments along all these lines. I will concentrate on the third thrust|improving MT quality through automatic knowledge acquisition.