Aligning Predicates across Monolingual Comparable Texts using Graph-based Clustering

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 171–182, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Aligning Predicates across Monolingual Comparable Texts using Graph-based Clustering Michael Roth and Anette Frank Department of Computational Linguistics Heidelberg University Germany {mroth,frank}@cl.uni-heidelberg.de Abstract Generating coherent discourse is an important aspect in natural language generation. Our aim is to learn factors that constitute coherent discourse from data, with a focus on how to re- alize predicate-argument structures in a model that exceeds the sentence level. We present an important subtask for this overall goal, in which we align predicates across compara- ble texts, admitting partial argument struc- ture correspondence. The contribution of this work is two-fold: We first construct a large corpus resource of comparable texts, includ- ing an evaluation set with manual predicate alignments. Secondly, we present a novel ap- proach for aligning predicates across compa- rable texts using graph-based clustering with Mincuts. Our method significantly outper- forms other alignment techniques when ap- plied to this novel alignment task, by a margin of at least 6.5 percentage points in F1-score. 1 Introduction Discourse coherence is an important aspect in natu- ral language generation (NLG) applications. A num- ber of theories have investigated coherence inducing factors. A prominent example is Centering Theory (Grosz et al., 1995), which models local coherence by relating the choice of referring expressions to the importance of an entity at a certain stage of a dis- course. A data-driven model based on this theory is the entity-based approach by Barzilay and Lap- ata (2008), which models coherence phenomena by observing sentence-to-sentence transitions of entity occurrences. Barzilay and Lapata show that their approach can discriminate between a coherent and a non-coherent set of ordered sentences. However, their model is not able to generate alternative entity realizations by itself. Furthermore, the entity-based approach only investigates realization patterns for individual enti- ties in discourse in terms of core grammatical func- tions. It does not investigate the interplay between entity transitions and realization patterns for full- fledged semantic structures. This interplay, how- ever, is an important factor for a semantics-based, generative model of discourse coherence. The main hypothesis of our work is that we can automatically learn context-specific realization pat- terns for predicate argument structures (PAS) from a semantically parsed corpus of comparable text pairs. Our assumption builds on the success of previous research, where comparable and parallel texts have been exploited for a range of related learning tasks, e.g., unsupervised discourse segmentation (Barzilay and Lee, 2004) and bootstrapping semantic analyz- ers (Titov and Kozhevnikov, 2010). For our purposes, we are interested in finding cor- responding PAS across comparable texts that are known to talk about the same events, and hence in- volve the same set of underlying event participants. By aligning predicates in such texts, we can inves- tigate the factors that determine discourse coher- ence in the realization patterns for the involved argu- ments. These include the specific …

Content truncated to 500 words