| John M. Zelle and Raymond J. Mooney
"Learning Semantic Grammars with Constructive
Inductive Logic Programming"
in "Proceedings of the Eleventh National Conference
on Artificial Intelligence (AAAI-93)", pp. 817-822,
AAI Press/MIT Press, 1993
John M. Zelle and Raymond J. Mooney
"Inducing Deterministic {P}rolog Parsers
from TreeBanks: A Machine Learning Approach"
"Proceedings of the Twelfth National Conference
on Artificial Intelligence (AAAI-94)", pp.748-753,
AAAI Press/MIT Press, 1994
|
| This dataset has been used for construction of semantic grammars. This is a difficult and interesting problem which has been treated by machine learning techniques recently [Zelle 93]. In the mentioned approach the semantic-grammar acquisition problem is viewed as learning of search-control heuristics. Appropriate control rules are learned using a new first-order induction algorithm that automatically invents useful syntactic and semantic categories. The logic programming formalism is used to represent these control rules. The task can be also viewed as an n-way categorization problem for complex tuples. Empirical results show that the learned parsers generalize well to novel sentences and out-perform previous approaches based on connectionnist techniques [Zelle 94].
The data come in two ILP formats.
1) Examples of the predicate parse which are pairs of sentences with their case-role analysis. For instance:
parse([the,man,ate],[ate,agt:[man,det:the]]).
2) Control examples for parsing the input sentences using a shift/reduce parser. Transitions of the parser are described by 4-tuples (Stack,Input,NewStack,NewInput). In this data set, lists of states of the parser, where first two atributes Stack, Input are instantiated, are matched to the corresponding semantic attachment of stack items. For instance:
op([S1,S2|SRest],Inp,[SNew|SRest],Inp):-
attach(S1,prep,S2,SNew).
op([[rock,det:the],with,[hit,pat:[rock,det:the],agt:[girl,det:the]]],[],NewStack,NewInput).
op([[hammer,det:the],with,[hit,pat:[rock,det:the],agt:[girl, det:the]]],[],NewStack,NewInput).
..
See the papers on CHILL in ftp://ftp.cs.utexas.edu/pub/mooney/papers/chill* for more details on this approach. (mk-control-xs)
Note:
M&K stands for McCelland and Kawamoto since this artificially generated data originally derives from:
J. L. McClelland and A. H. Kawamoto
"Mechanisms of Sentence Processing: Assigning Roles
to Constituents of Sentences"
in D. E. Rumelhart and J. L. McClelland (ed.):
Parallel Distributed Processing,
Vol. II, pp. 318-362
MITP, Cambridge, MA 1986
|
and GMD 