An effective discourse parser that uses rich linguistic information
NAACL '09 Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics
| Hi-index | 0.00 |
Previous work in discourse parsing have overlooked the role of relational data and have relied mostly on syntactic and lexical information. This thesis investigates the impact of a highly structured relational learning model for discourse parsing of natural language text. Our relational learning model makes use of compositional semantics and segment discourse structure, in addition to lexical and linguistic cues. A new resource for discourse structure based on informational relations was developed. Two lexical semantic resources for verbs and nouns were coupled with a symbolic parser to build semantic representations. The problem of identifying informational relations was cast as a multi-classification problem. Results from experiments show that a relational learning model based on Inductive Logic Programming (ILP) outperforms propositional rule learners like Decision Tree and RIPPER as well as the probabilistic Naive Bayes model. To build the entire discourse structure of a given natural language text, a shift-reduce discourse parser was developed that identifies informational relations between text segments and builds a hierarchical tree structure of text. The shift-reduce parser that uses the ILP based relation classifier provides a significant improvement over a right branching majority class baseline model.