Aligning Predicate Argument Structures in Monolingual Comparable Texts: A New Corpus for a New Task

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 218–227, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Aligning Predicate Argument Structures in Monolingual Comparable Texts: A New Corpus for a New Task Michael Roth and Anette Frank Department of Computational Linguistics Heidelberg University Germany {mroth,frank}@cl.uni-heidelberg.de Abstract Discourse coherence is an important aspect of natural language that is still understudied in computational linguistics. Our aim is to learn factors that constitute coherent discourse from data, with a focus on how to realize predicate- argument structures (PAS) in a model that ex- ceeds the sentence level. In particular, we aim to study the case of non-realized arguments as a coherence inducing factor. This task can be broken down into two subtasks. The first aligns predicates across comparable texts, ad- mitting partial argument structure correspon- dence. The resulting alignments and their con- texts can then be used for developing a coher- ence model for argument realization. This paper introduces a large corpus of com- parable monolingual texts as a prerequisite for approaching this task, including an evaluation set with manual predicate alignments. We il- lustrate the potential of this new resource for the empirical investigation of discourse coher- ence phenomena. Initial experiments on the task of predicting predicate alignments across text pairs show promising results. Our findings establish that manual and automatic predicate alignments across texts are feasible and that our data set holds potential for empirical re- search into a variety of discourse-related tasks. 1 Introduction Research in the fields of discourse and pragmatics has led to a number of theories that try to explain and formalize the effect of discourse coherence induc- ing elements either locally or globally. For exam- ple, Centering Theory (Grosz et al., 1995) provides a framework to model local coherence by relating the choice of referring expressions to the salience of an entity at certain stages of a discourse. An exam- ple for a global coherence model would be Rhetori- cal Structure Theory (Mann and Thompson, 1988), which addresses overall text structure by means of coherence relations between the parts of a text. In addition to such theories, computational ap- proaches have been proposed to capture correspond- ing phenomena empirically. A prominent example is the entity-based model by Barzilay and Lapata (2008). In their approach, local coherence is mod- eled by the observation of sentence-to-sentence re- alization patterns of individual entities. The learned model reflects a key idea from Centering Theory, namely that adjacent sentences in a coherent dis- course are likely to involve the same entities. One shortcoming of Barzilay and Lapata’s model (and extensions of it) is that it only investigates overt realization patterns in terms of grammatical func- tions. These functions reflect explicit realizations of predicate argument structures (PAS), but they do not capture the full range of salience factors. In partic- ular, the model does not reflect the importance of discourse entities that fill core roles of the predicate, but that remain implicit in the predicate’s local argu- ment …

Content truncated to 500 words