Vorschau

PDF, Englisch - Hauptdokument Download (1MB) | Nutzungsbedingungen

Abstract

Natural language texts frequently contain related information in different positions in discourse. As human readers, we can recognize such information across sentence boundaries and correctly infer relations between them. Given this inference capability, we understand texts that describe complex dependencies even if central aspects are not repeated in every sentence. In linguistics, certain omissions of redundant information are known under the term ellipsis and have been studied as cohesive devices in discourse (Halliday and Hasan, 1976). For computational approaches to semantic processing, such cohesive devices are problematic because methods are traditionally applied on the sentence level and barely take surrounding context into account.

In this dissertation, we investigate omission phenomena on the level of predicate-argument structures. In particular, we examine instances of structures involving arguments that are not locally realized but inferable from context. The goal of this work is to automatically acquire and process such instances, which we also refer to as implicit arguments, to improve natural language processing applications. Our main contribution is a framework that identifies implicit arguments by aligning and comparing predicate-argument structures across pairs of comparable texts. As part of this framework, we develop a novel graph-based clustering approach, which detects corresponding predicate-argument structures using pairwise similarity metrics. To find discourse antecedents of implicit arguments, we further design a heuristic method that utilizes automatic annotations from various linguistic pre-processing tools.

We empirically validate the utility of automatically induced instances of implicit arguments and discourse antecedents in three extrinsic evaluation scenarios. In the first scenario, we show that our induced pairs of arguments and antecedents can successfully be applied to improve a pre-existing model for linking implicit arguments in discourse. In two further evaluation settings, we show that induced instances of implicit arguments, together with their aligned explicit counterparts, can be used as training material for a novel model of local coherence. Given discourse-level and semantic features, this model can predict whether a specific argument should be explicitly realized to establish local coherence or whether it is inferable and hence redundant in context.