A Treebank Query System Based on an Extracted Tree Grammar

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 661–664, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics A Treebank Query System Based on an Extracted Tree Grammar Seth Kulick and Ann Bies Linguistic Data Consortium University of Pennsylvania 3600 Market St., Suite 810 Philadelphia, PA 19104 {skulick,bies}@ldc.upenn.edu Abstract Recent work has proposed the use of an ex- tracted tree grammar as the basis for treebank analysis and search queries, in which queries are stated over the elementary trees, which are small chunks of syntactic structure. However, this work was lacking in two crucial ways. First, it did not allow for including lexical properties of tokens in the search. Second, it did not allow for using the derivation tree in the search, describing how the elementary trees are connected together. In this work we describe an implementation that overcomes these problems. 1 Introduction (Kulick and Bies, 2009) describe the need for tree- bank search that compares two sets of trees over the same tokens. Their motivation is the problem of comparing different annotations of the same data, such as with inter-annotator agreement evaluation during corpus construction. The typical need is to recognize which annotation decisions the annotators are disagreeing on. This is similar to the problem of determining where the gold trees and parser output differ, which can also be viewed as two annotations of the same data. As they point out, for this purpose it would be use- ful to be able to state queries in a way that relates to the decisions that annotators actually make, or that a parser mimics. They provide examples suggesting that (parent, head, sister) relations as in e.g. (Collins, 2003) are not sufficient, and that what is needed is the ability to state queries in terms of small chunks of syntactic structure. Their solution is to use an extracted tree gram- mar, inspired by Tree Adjoining Grammar (Joshi and Schabes, 1997). The “elementary trees” of the TAG-like grammar become the objects on which queries can be stated. They demonstrate how the “lexicalization” property of the grammar, in which each elementary tree is associated with one or more token, allows for the the queries to be carried out in parallel across the two sets of trees. However, the work was lacking in two crucial ways. First, it did not allow for including lexical properties of a token, such as its Part-of-Speech tag, together with the elementary tree search. This made it impossible to formulate such queries as “find all ADVP elementary trees for which the head of the tree is a NOUN NUM”. Even more seriously, there was no way to search over the “derivation tree”, which encodes how the extracted elementary trees combine together to create the original tree. This made it impossible to carry out searches such as “find all verb frames with a PP-LOC modifying it”, and in general to search for the crucial question of where annotators disagree on attachment decisions. …

Content truncated to 500 words