From Treebanks to Tree-Adjoining Grammars

Grammars are valuable resources for natural language processing. A large-scale grammar may incorporate a vast amount of information on morphology, syntax, and semantics. Traditionally, grammars are built manually. Hand-crafted grammars often contain rich information, but require tremendous human effort to build and maintain. As large-scale treebanks become available in the last decade, there has been much work on extracting grammars automatically from treebanks. Such grammars are called treebank grammars. Many of the previous work on grammar extraction such as (Shirai, Tokunaga, and Tanaka, 1995; Charniak, 1996; Krotov et al., 1998) generate context-free grammars (CFGs). In this chapter, we present a system, LexTract, which generates both CFGs and lexicalized tree adjoining grammars (LTAGs). Extracting LTAGs is more complicated than extracting CFGs because of the differences between LTAGs and CFGs. First, the primitive elements of an LTAG are lexicalized tree structures (called elementary trees), rather than context-free rules. Therefore, an LTAG extraction algorithm needs to examine a larger portion of a phrase structure to build an elementary tree. Second, because the adjoining operation in LTAG allows an elementary tree to be inserted within another elementary tree, an elementary tree is often formed by gluing together several disconnected parts of a phrase structure. Third, unlike in CFGs, parse trees (also known as derived trees in the LTAG formalism) and derivation trees (which describe how elementary trees are combined to form parse trees) are distinct in the LTAG formalism in the sense that a parse tree can Treebank grammars may contain less information (e.g., feature structures associated with non-terminal nodes) than some hand-crafted grammars, but they are sufficient for some NLP tasks such as Supertagging and parsing as described in Section 4.