|References||[2}, , [1}|
|Other comments||temporarily restricted availability|
The system LAPIS implements an approach to learning LR parsers from tree-banks annotated with lexical semantic labels. The parsers so learned are based on grammars which are adapted to obtain optimal performance on the training data, i.e. to reduce the number of spurious parses by means of rule unfolding, lexicalisation, and the use of lexical semantic constraints. Such tailored grammars are extremely useful for restricted domains but a new one must be written for each of them, and their manual development can become very costly.
The fact that in LAPIS the concepts learned are parser actions rather than grammar rules, has a clear justification. The major possible criticism related to the difficulty to interpret the results obtained when parser actions are learned, is easily rejected, as the mapping of parser actions into grammar rules is straight-forward. In that way, the results of learning, the specialisations of parser actions are easily interpreted as partially unfolded and/or lexicalised grammar rules also containing lexical semantic constraints. One the other hand, although grammar rules and parser actions are closely related, the former are recursive concepts, whereas the latter are not. Learning recursive concepts is a notoriously hard task, therefore the shift in paradigm from learning grammar rules to learning parser actions allows to reduce the complexity of learning.
The integrated framework for speech recognition and parsing described in  traces the way of a number of possible applications in which the LR parsers learned with LAPIS are combined with the IBM commercial tools for speech recognition. An additional advantage of LAPIS in this context is the fact that the SRCL grammars required by the SR tools can be developed automatically, as a by-product of the parser learning.