Semantic Parsing for High-Precision Semantic Role Labelling

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 1–8 Manchester, August 2008 Semantic Parsing for High-Precision Semantic Role Labelling Paola Merlo Linguistics Department University of Geneva 5 rue de Candolle 1211 Gen`eve 4 Switzerland merlo@lettres.unige.ch Gabriele Musillo Depts of Linguistics and Computer Science University of Geneva 5 Rue de Candolle 1211 Gen`eve 4 Switzerland musillo4@etu.unige.ch Abstract In this paper, we report experiments that explore learning of syntactic and seman- tic representations. First, we extend a state-of-the-art statistical parser to pro- duce a richly annotated tree that identi- fies and labels nodes with semantic role la- bels as well as syntactic labels. Secondly, we explore rule-based and learning tech- niques to extract predicate-argument struc- tures from this enriched output. The learn- ing method is competitive with previous single-system proposals for semantic role labelling, yields the best reported preci- sion, and produces a rich output. In com- bination with other high recall systems it yields an F-measure of 81%. 1 Introduction In statistical natural language processing, consid- erable ingenuity and insight have been devoted to developing models of syntactic information, such as statistical parsers and taggers. Successes in these syntactic tasks have recently paved the way to applying novel statistical learning techniques to levels of semantic representation, such as re- covering the logical form of a sentence for in- formation extraction and question-answering ap- plications (Miller et al., 2000; Ge and Mooney, 2005; Zettlemoyer and Collins, 2007; Wong and Mooney, 2007). In this paper, we also focus our interest on learn- ing semantic information. Differently from other work that has focussed on logical form, however, we explore the problem of recovering the syn- tactic structure of the sentence, the propositional c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. argument-structure of its main predicates, and the substantive labels assigned to the arguments in the propositional structure, the semantic roles. This rich output can be useful for information extrac- tion and question-answering, but also for anaphora resolution and other tasks for which the structural information provided by full syntactic parsing is necessary. The task of semantic role labelling (SRL), as has been defined by previous researchers (Gildea and Jurafsky, 2002), requires collecting all the argu- ments that together with a verb form a predicate- argument structure. In most previous work, the task has been decomposed into the argument iden- tification and argument labelling subtasks: first the arguments of each specific verb in the sentence are identified by classifying constituents in the sen- tence as arguments or not arguments. The argu- ments are then labelled in a second step. We propose to produce the rich syntactic- semantic output in two steps, which are different from the argument identification and argument la- belling subtasks. First, we generate trees that bear both syntactic and semantic annotation, such as those in Figure 1. The parse tree, however, does not explicitly encode information about predicate- argument structure, because it does not explicitly associate each semantic …

Content truncated to 500 words