Treebank vs. Xbar-based Automatic F-Structure Annotation

Manual, large scale (computational) grammar development is time consuming, expensive and requires lots of linguistic expertise. More recently, a number of alternatives based on treebank resources (such as Penn-II, Susanne, AP treebank) have been explored. The idea is to automatically ``induce'' or rather read off (P)CFG grammars from the parse annotated treebank resources and to use the treebank grammars thus obtained in (probabilistic) parsing or as a starting point for further grammar development. The approach is cheap, fast, automatic, large scale, ``data driven'' and based on real language resources.

Treebank grammars typically involve large sets of lexical tags and non-lexical categories as syntactic information tends to be encoded in monadic category symbols. They feature flat rules (trees) that can ``underspecify'' attachment possibilities. Treebank grammars do not in general follow Xbar architectural design principles (this is not to say that treebank grammars do not have design principles). As a consequence, treebank grammars tend to have very large CFG rule bases (e.g. Penn-II > 17,000 CFG rules for about 1 million words of text) with often only minimally differing rules. Even though treebank grammars are large, they are still incomplete, exhibiting unabated rule accession rates. From a grammar engineering point of view, the size of the rule base poses problems for maintainability, extendability and, if a treebank grammar is to be used as a CF-base in a LFG grammar, for functional (feature-structure) annotations. From the point of view of theoretical linguistics, flat treebank trees and treebank grammars extracted from such trees do not express linguistic generalisations. From the perspective of empirical and corpus linguistics, flat trees are well-motivated as they allow underspecification of subtle and often time consuming attachment decisions. Indeed, it is sometimes doubted whether highly general Xbar schemata usefully scale to ``real'' language.

In previous work we developed methodologies for automatic feature-structure annotation of grammars extracted from treebanks. Automatic annotation of ``raw'' treebank grammars is difficult as annotation rules often need to identify subsequences in the RHSs of flat treebank rules as they explicitly encode head, complement and modifier relations. Xbar based CFG rules should substantially facilitate automatic feature-structure annotation of grammar rules.

In the present paper we conduct a number of experiments to explore a space of possible grammars based on a small fragment of the AP treebank resource. Starting with the original treebank fragment we automatically extract a CFG G. We then apply an automatic structure preserving grammar compaction step which generalises categories in the original treebank fragment and reduces the number of rules extracted, resulting in a generalised treebank fragment and in a compacted grammar G^c. The generalised fragment is then manually corrected to catch missed constituents (and the like) resulting in an automatically extracted, compacted and (effectively manually) corrected grammar G^c,m. Manual correction proceeds in the ``spirit'' of treebank grammars (we do not introduce Xbar analyses). We then explore how many of the manual correction steps on treebank trees can be achieved automatically. We develop, implement and test an automatic treebank ``grooming'' methodology which is applied to the generalised treebank fragment to yield a compacted and automatically corrected grammar G^c,a. Grammars G^c,m and G^c,a are very similar to compiled out ``flat'' LFG-82 style grammars. We explore regular expression based compaction (both manual and automatic) to relate G<sup>c,m</sup> to a LFG-82 style grammar design. Finally, we manually recode a subsection of the generalised and manually corrected treebank fragment into ``vanilla-flavour'' XBar based trees. From these we extract a compacted, manually corrected, XBar based grammar G^c,m,x. We evaluate our grammars and methods using standard labelled bracketing measures and according to how well they perform under automatic feature-structure annotation tasks.