Introduction
Lexicalized Tree Adjoining Grammar (LTAG) is a powerful linguistic formalism where every word in a lexicon is associated with one or more "elementary trees" representing its syntactic structures. However, this "radical lexicalism" often leads to significant redundancy. For instance, in the English XTAG system, thousands of lexemes share identical tree families or undergo the same grammatical transformations, such as passivization or dative shifts.
Standard LTAG formalisms lack a native mechanism to capture these regularities efficiently. Previous solutions, such as "metarules" or monotonic inheritance, often required complex auxiliary formalisms to handle exceptions. This paper proposes using DATR, a nonmonotonic inheritance language, to represent the entire LTAG lexicon—including trees, rules, and exceptions—within a single, unified framework.
Methods: Bottom-Up Encoding and DATR
The core of the proposed method is a shift in how trees are represented. Instead of a traditional top-down description, the authors encode trees from the lexical anchor (the word itself) upwards.
Bottom-Up Tree Encoding
Using a variant of Kilbury’s encoding, trees are defined through binary relations: parent, left (sister), and right (sister). These relations are embedded directly into DATR feature structures. For example, a path like <parent parent cat> might return s, indicating that the grandparent of the current node is a Sentence node.
Nonmonotonic Inheritance
Unlike monotonic systems, DATR allows for "defaults" that can be overridden. This is essential for linguistic exceptions, such as a verb that follows a general transitive pattern but lacks a passive form. In this framework, lexical rules like the dative shift are treated as feature covariation constraints between <input> and <output> paths within the same inheritance node, rather than complex tree-rewriting operations.
Hierarchical Lexicon Design
The authors demonstrate that a hierarchical lexicon can drastically reduce redundancy. By organizing verbs into a hierarchy based on their subcategorization frames, common properties are inherited from higher-level nodes.
This hierarchy allows for a compact representation where a single line of DATR code can define a transformation across an entire class of verbs. The authors successfully integrated six major construction types—including passive, dative, subject-auxiliary inversion, and relative clauses—into this unified structure.
Key Insights
The central realization of this research is that because LTAG is radically lexicalist, a general-purpose lexical representation language like DATR is the most natural fit for it. By embedding tree structures into feature paths, the complexity of grammatical rules is reduced to simple feature constraints. Furthermore, the nonmonotonic approach allows for "total" default descriptions that are edited where necessary, which is often more intuitive and easier to implement than the "partial" descriptions required by monotonic systems.
Conclusion and Future Work
The paper demonstrates that DATR can provide a compact, extensible, and unified framework for LTAG lexicons. Future work aims to refine the handling of rule chaining (e.g., applying both passive and wh-question rules) and to provide more detailed accounts of complex cross-referencing for "null" elements. While the representation is theoretically efficient, empirical benchmarks on parsing speed and memory overhead in large-scale NLP tasks remain an area for further investigation.
Summary updated