Natural Language Understanding
Natural language understanding is the branch of artificial intelligence concerned with getting a machine to extract usable meaning from human language: who did what to whom, what is being asked, which entity a pronoun points to, and what follows from the statement. It is distinct from merely processing text, such as counting words or checking spelling, because the target is an internal representation that supports action or inference. The field has repeatedly redefined what understanding should mean, and its history can be read as three successive answers: hand-written rules, statistical learning, and neural networks trained on vast text.
ELIZA, written by Joseph Weizenbaum at MIT and described in a 1966 paper in Communications of the ACM, showed how little machinery it takes to look like understanding. Its best-known script, DOCTOR, mimicked a Rogerian psychotherapist by matching keywords in the user's input and turning phrases around into questions, with no representation of meaning at all. Weizenbaum was disturbed by how readily people attributed comprehension to it, and the tendency to read depth into fluent output is now called the ELIZA effect. The program remains the standard warning that conversational plausibility is not evidence of comprehension, a warning that applies with new force to modern chatbots.
SHRDLU, built by Terry Winograd at MIT for his doctoral work and reported in 1970-72, was the opposite extreme: deep understanding of a world so small that everything could be modelled. The program conversed in English about a simulated table of blocks, pyramids and boxes, accepting commands such as picking up a block, answering questions about the scene and its own earlier actions, and resolving pronouns by consulting a model of what had just been discussed. Syntax, semantics and reasoning were integrated, and its parser ran procedurally inside a Lisp and Micro-Planner system. Winograd later concluded that the approach did not scale to the open world, since each domain demanded fresh engineering and real language leans on unlimited background knowledge. He turned toward human-computer interaction and critiques of the symbolic program.
Semantic parsing, the task of mapping a sentence into a formal meaning representation such as a logical form or a database query, is the central technical thread connecting the eras. Woods's LUNAR system in the early 1970s answered questions about moon rocks by translating English into formal queries; Zelle and Mooney's work in the mid-1990s learned such mappings from examples of geography questions; and modern text-to-SQL systems do the same with neural models. The difficulty is ambiguity: a sentence usually has many parses and readings, and picking the right one depends on context and world knowledge. The statistical shift in the 1990s, driven by annotated resources like the Penn Treebank, replaced brittle hand rules with probabilities learned from data, and neural word embeddings and pretrained Transformers such as BERT in 2018 later added broad contextual representations.
Winograd schemas are the tests most closely tied to this history, proposed by Hector Levesque as an alternative to the Turing test. A schema is a pair of sentences differing by one word that flips the referent of a pronoun, building on an example from Winograd's 1972 work: the councilmen refused the demonstrators a permit because they feared violence, versus because they advocated violence. Answering requires knowing who typically fears and who advocates, which cannot be done by grammar or word statistics alone. Levesque, Davis and Morgenstern formalized the Winograd Schema Challenge in 2012, and it was run as a contest in 2016, when entrants scored not far above chance. Large language models later reached high scores on the benchmarks, which raised the question of whether the benchmark or the models had changed in meaning.
Understanding is often contrasted with generation: natural language understanding maps language into meaning, while natural language generation maps meaning into language. The two were separate engineering problems for decades, and a dialogue system had one component for each. Modern large language models blur the boundary by treating both as next-token prediction. A frequent misconception is that fluent output proves understanding, which ELIZA already refuted; the opposite error is dismissing all neural capability as mere pattern matching without specifying a test that would distinguish the two. Whether benchmark success reflects comprehension, grounding or shortcut learning remains an open and contested question.
Read natural language understanding as the long attempt to make machines recover what a sentence means rather than what it looks like: ELIZA showed that appearance is cheap, SHRDLU showed that depth is costly and narrow, and the statistical and neural eras trade between them.