Chomsky's Syntactic Structures (1957) › Applications in Early NLP

Chomsky's immediate contribution to natural language processing was a formal basis for syntactic analysis. Earlier systems relied mainly on shallow pattern matching or statistical co-occurrence. They might recognize that certain words tended to appear together, but they had no representation of syntactic structure. Researchers could now build systems that analyzed the grammatical relationships between words in a sentence.

SHRDLU, Terry Winograd's famous 1970 system, demonstrated the power of this approach. Operating in a simulated world of colored blocks, SHRDLU could understand and execute complex commands like "Pick up a big red block" or "Find a block which is taller than the one you are holding and put it in the box." It did this by parsing commands into hierarchical syntactic structures, then mapping those structures to semantic representations and finally to actions. The system's linguistic sophistication was remarkable for its time, handling anaphora (like "it" referring back to a previously mentioned object), nested relative clauses, and ambiguous quantification.

Question-answering systems also relied heavily on syntactic parsing. To answer a question like "Who painted the ceiling of the Sistine Chapel?" you need to understand that "Who" is asking about the agent, "painted" is the action, and "the ceiling of the Sistine Chapel" is the object. Shallow keyword matching might find documents mentioning "Sistine Chapel" and "painted," but only syntactic analysis could reliably determine that you're looking for the painter, not, say, the person who commissioned the painting or owns the building. Early systems like BASEBALL (which answered questions about baseball statistics) and LUNAR (which answered questions about moon rocks) used Chomskyan parsing as a core component.

Machine translation initially seemed like an ideal application for transformational grammar. The deep structure represented meaning independent of surface form, so the idea was to parse the source language to deep structure, then generate surface structure in the target language. This was conceptually elegant and aligned with Chomsky's theoretical framework. Unfortunately, it proved practically difficult. Natural languages are messy, ambiguous, and full of idiomatic expressions that don't translate literally. The deep structures required for accurate translation were far more complex than early researchers anticipated, and hand-crafting transformation rules for language pairs was extraordinarily labor-intensive.

Still, the rule-based, syntactic approach dominated machine translation research for decades. Systems like SYSTRAN, which provided commercial translation services, used elaborate hand-crafted rules based on syntactic analysis. While they were eventually superseded by statistical and neural approaches, they represented the state of the art for many years and were the only practical option for large-scale translation. The shift away from rule-based systems didn't happen until the 1990s, when statistical methods demonstrated that data-driven approaches could outperform hand-crafted rules.