🧑‍🔬 Donald Hindle

Donald Hindle's papers here apply statistical methods to large bodies of automatically parsed text to learn facts about English words and how they combine. A recurring question is how far word meaning and syntactic preference can be recovered from patterns of use alone, without hand-built dictionaries or hierarchies. The typical material is Associated Press newswire, processed by the partial parser Fidditch, from which grammatical relations between words are extracted and scored. The work touches on lexical semantics & word embeddings and on the predicate–argument relations studied under semantic role labeling.

The most cited of these papers is Noun Classification From Predicate-Argument Structures (1990). It puts Harris's distributional hypothesis to a larger test than earlier studies by comparing nouns through the verbs that take them as subjects or objects. Each link between a noun and a verb is weighted by mutual information, so that frequent but uninformative arguments count for little. The parser leaves uncertain relations unfilled, so its mistakes mostly drop evidence rather than add false evidence. The second paper, Structural Ambiguity and Lexical Relations (1993), addresses whether a prepositional phrase after a verb and its object attaches to the verb or to the noun. It notes that structural principles such as Right Association and Minimal Attachment predict human choices poorly. In their place it uses word-specific preferences, gathered from verb–noun–preposition triples in parsed newswire.

The later paper thus carries the corpus-based approach of the noun-classification study from word similarity over to syntactic disambiguation, with lexical association again doing the main work. The bylines of Hindle's papers point mainly to AT&T, spanning 1989 to 2002, and earlier to the University of Pennsylvania, between 1977 and 1982.

Donald Hindle has 2 papers in our library, in 1990–1993, most of them on Semantic role labeling, Lexical semantics & word embeddings.