nlp
A million parsed sentences
It is 1995, and the parsing wars have a new referee: a corpus of a million words of newspaper text, every sentence hand-annotated with its syntactic structure. The rule-based parsers, each linguist-decades in the making, cover their author's intuitions and shatter on real newswire. Build the statistical parser instead: learn a probabilistic grammar from the treebank, resolve the ambiguity that makes "I saw the man with the telescope" two sentences, and be measured — precision and recall against held-out annotation, not elegance of formalism. Then say what the treebank's conventions have quietly decided about the science. Get it wrong and parsing stays an unfalsifiable craft; get it right and annotated data becomes the field's currency, for better and for worse.
Who this problem belongs to
The two figures whose methods fit it best, out of 46 in contention.
Manning's own doctoral and early career work in the 1990s was substantially built on exactly this problem's scenario: the Penn Treebank, released with its first million-word annotated corpus by 1992-93, and the statistical parsers researchers built against it through the mid-1990s. His and Schutze's later Foundations of Statistical Natural Language Processing systematized the entire methodology this problem asks for — learning probabilistic context-free grammars from annotated trees, resolving structural ambiguity like prepositional-phrase attachment through learned statistics rather than hand-coded rules, and evaluating with precision and recall against held-out treebank sections rather than by elegance of formalism. He also wrote seriously about what the treebank's Penn Treebank annotation conventions quietly encoded as linguistic assumptions. This is close to Manning's actual formative research problem, not an applied analogy from a different field.
Jelinek's statistical approach at IBM through the 1970s and 80s — replace hand-built linguistic rules with statistics learned from data — is the philosophical and methodological ancestor of exactly what this problem demands: treat parsing as a learnable statistical problem rather than a hand-crafted grammatical exercise. His famous remark that firing linguists improved his speech recognizer captures precisely the empirical, data-driven stance the statistical parsers built on the Penn Treebank in the mid-1990s adopted wholesale, measuring themselves by held-out accuracy rather than linguistic elegance. Jelinek's own group worked primarily in speech recognition and language modeling rather than syntactic parsing specifically, and he did not build treebank parsers himself, though his methodological approach directly enabled the field that did. His relevance is as the statistical-NLP tradition's founding figure whose creed this problem's parser embodies.
In the mind map
The same ideas, as concepts rather than history — in John's ML knowledge map.
46 figures are scored on this problem. Draw it in a battle to see where you land.