Coherence-targeted text summarization

Abstract

For readers, coherence is no less important than informativeness for a summary. This paper is aimed to improve coherence in automatic text summaries by developing coherence models and related techniques. Different from most other efforts to improve summary coherence, my work treats coherence as an analyzable concept with multi-faceted and multi-disciplinary backgrounds. Specifically, I have explored the technical details of three kinds of coherence - shallow content-driven coherence, deep content-driven coherence, and cognitive model-driven coherence. Shallow content consists of words, phrases, sentences, and discourse units and their literal connections or co-occurrence patterns give rise to coherence. Experiments on single-document as well as multi-document news summarization show that coherence driven by words, entities, sentences, and events can help to better arrange selected summary sentences. Deep content is observed on a macro-text level, which is instantiated by news aspects and speech acts. Focusing on the relations among deep content units, I have applied coherence to both selecting and ordering summary sentences. Relying on human cognitive tendencies, cognitive model-driven coherence is understood as a necessary mechanism in text comprehension. The computational modeling of such coherence, coupled with proposition-level extractive summarization, works successfully for narrative text. To model coherence of different kinds, I have developed novel techniques that are suitable for different genres of text, including newswire, social media messages, and fairy tales. The extensive experimental results on benchmark or self-compiled datasets have validated the efficacy and robustness of the techniques in various circumstances. Among many of its contributions to the summarization community, my work shows that contrary to what is commonly held, coherence plays a pivotal, instead of ancillary, role in automatic summarization. As one of the few large-scale studies of coherence in summarization, my work is expected to herald a complete theory of coherence and more in-depth studies in coherence-targeted text summarization.