Some notes on how to make parsing with Pet more robust.

Preprocessing

Token Mapping

Different assumptions made by preprocessing tools and unhandled characteristics of the text to be processed can influence parse success or failure. In order to adapt the output of the preprocessing tools to the assumptions made by the grammar, you can use Chart_Mapping, a mechanism for the non-monotonic, rule-based manipulation of chart items that are described by feature structures. One of the applications of Chart Mapping is Token Mapping, which gives users a means for manipulating tokens using the known formal devices of the grammar like unification and coindexation.

Unknown words

For both methods, it holds that it is unclear which REL it gets. Also, is it a word or a lexeme? There is no possibility in general (if the surface string is not stamped into CARG, as for proper names) to generate back again from the MRS to the surface string, if unknown word handling has been applied. Such information should be included in the MRS somehow, and this can be achieved in the user-fns.lsp

As to PET, it will be possible to construct the PRED string of an unknown word with chart mapping. However, since inflectional analysis could return many possible stems for a given word and since there is no way to pick the correct one reliably by only looking at this one form, running inflectional analysis before lexical instantiation in order to provide a stem in token mapping doesn't seem to be a strategy that is general enough. The surface form would have to be used instead to build the PRED value, leading to separate relations for each form of a lexeme.

Stephan points out that partial lexical gaps (is verb, but only encoded as noun) is not captured by these mechanisms, but this will be solved when the chart mapping is merged into main.

Generics triggered from POS mapping

cheap -default-les

Instantiates generic lexical entries for lexical gaps in the chart using a mapping from pos tags to generic les that has to be specified in the grammar. The pos tags have to be delivered by an external pos tagger and can be passed to PET by most of the input formats. See PetInput for more information.

Generics triggered from token fs

cheap  -default-les=all -cm

This is the new way of generic lexical instantiation that has been introduced together with token feature structures and Chart_Mapping. In this new setup, the parser tries to instantiate all generic lexical entries for each word. Upon lexical instantiation, the token feature is unified into a designated path of the lexical entry. Only if this unification succeeds, the lexical item is instantiated. In order to control the instantiation of generic lexical entries, the token feature structures are appropriatly constrained in the generic lexical entry, for instance by requiring that a generic verbal entry is only applicable for token feature structures where the highest ranked part-of-speech tag is a verb. Please see the page on Chart_Mapping for more information.

Lexical type prediction

By Yi.

cheap  -predict-les

Maximum Entropy Model has to be created from a script that Yi, that needs a treebanked profile as input. In e.g. pred-lex.tdl, it should be listed which lexical types should be able to be predicted. At least 2000 sentences are needed.

Supertagger

A supertagger, similar in nature to the CCG's, and based on CRF's, is currently under construction.

Grammar Internal Solutions

Pet settings

The amount of items that are parsed of a corpus, depends heavily on the high number of parameters passed onto the parser. Some of those settings involves restricting the search space (like reducing the maximum number of edges). The constraint that Dan uses now is the mem option in cheap (although you should half it, for some strange reason!). This seems to be the most reasonable setting.

See also PetParameters.

Robuster Partial Parsing

Yi created a two-phase parsing algorithm that, in case the deep grammar does not succeed, a CFG backbone is used to still get a reasonable parse. This still has to be integrated in the main branch.

PetRobustness (last edited 2011-10-08 21:12:07 by localhost)

(The DELPH-IN infrastructure is hosted at the University of Oslo)