Sampson 1986 - A stochastic approach to parsing Learns statistical rules from a manually annotated corpus. Uses simulated annealing to find the most probable parse. (Randomized inference, similar to later work in NLP in 2014
here) “We have built up a database of manually-parsed sentences, from which we extract statistics that allow a likelihood measure to be determined for any logically possible non-leaf constituent of a parse-tree. That is, given a pairing of a mother-label with a sequence of daughter-labels, say the pair <J, NN JJ P>, the likelihood function will return a figure for the relative frequency with which (in this case) an adjective phrase consists of singular common noun + adjective + prepositional phrase.” “The most direct way… would be to generate all possible tree-structures for a given sentence taken as a sequence of word-tags, and all possible labellings of each of those structures, and choose the tree whose overall plausibility figure is highest. Unlike in the case of word-tagging, however, for parsing this approach is wholly impractical… I have therefore begun to experiment with simulated annealing as a solution to the problem.”