Statistical versus knowledge-based machine translation

over, and what kinds of symbol systems should we create for them? Everytime a new phenomenon is identified as a bottleneck or as problematic, the very actsof describing the phenomenon, defining it,and creating a set of symbols to represent its abstractions are symbolic (in both senses of the word!). The benefits: decreasedlearning time and more powerful rules,hence improved translation output quality. The picture is inverted on the symbolic/ linguistic side. Here the system is designedto use a great deal of knowledge about lexical features, grammatical word classes, and even perhaps semantic knowledge, as in thecase of Pangloss. But this knowledge must Even if the human's work isperfect and complete, thefact that one needs at least120,000 words to cover asignificant portion of alanguage such as English orSpanish means that it takesyears for a group oflexicogrammarians todevelop an adequate MTsystem. be built into the system. Lexicons of words and rules of grammar, acquired by humanlabor, are expensive to compile and slow toaccrue. Where a statistical system can siftthrough thousands of bilingual word corre-spondences an hour, a human cannot build more than a handful of detailed lexicalitems or grammar rules in that time. Even ifthe human's work is perfect and complete,the fact that one needs at least 120,000words to cover a significant portion of alanguage such as English or Spanish means that it takes years for a group oflexicogrammarians to develop an adequateMT system. Generally, in the real world,the oldest systems are still the best.But ARPA had only three years' fundingfor MT. And the ARPA MT program became increasingly ambitious, from initiallycalling for high-quality translations in onlya limited domain (necessitating a small but detailed lexicon), to ultimately requiring thesystems to handle unrestricted newspaper text. Over the four years of the program, ARPA held four formal evaluations, whichused various scales to compare translations produced by research systems, several com-mercial systems, and human experts.Pressure increased on Pangloss, the symbolic/linguistic system, to expand its lexicon and grammar dramatically. The onlyway to respond was to automate: decreasethe amount of information for each lexical item (because this usually requires human analysis), and acquire the lexical items andgrammar patterns by machine. This step immediately introduced statistics-like processing into Pangloss.Until they mature, symbolic systems thusrespond mainly to a drive toward coverage and robustness. Especially in the face ofincreasingly challenging evaluations, symbolic system researchers begin to developgeneral rules to avoid catastrophic failurewhenever the system encounters input forwhich specific rules have not yet been built.Such general rules usually provide not onlythe correct output for any input but a list of possible outputs for a general class of inputs.These outputs, which are correct at a certainlevel of generality, are filtered to select thebest altemative(s). But what filter? When thetask/evaluations prohibit human interven-tion, the filter must be automatic, and thus requires reliability indicators. By the twinmoves of computing reliability numbersand extracting information from resources(semi-) automatically, symbolic systembuilders take their inevitable steps towardstatistics. The benefits: a greatly expanded lexicon and more grammatical coverage, hence translation in larger domains.Once begun, the process of hybridization continued for Candide and Pangloss (Ling-stat and Japangloss, a sibling of Pangloss,were hybrids from the outset).