UBC Theses and Dissertations
Exploration on the synergy between discourse and neural summarizers Xiao, Wen
Automatic Text Summarization is the challenging NLP task of summarizing some source input text - a single document, a cluster of multiple related documents, or conversations - into a shorter text, covering the main points succinctly. Since the 1950s, researchers have been exploring with some success both extractive solutions - simply picking salient sentences from the source, and abstractive ones - directly generating the summary word by word. Before the neural revolution, summarizers relied on explicit linguistic properties of the input, including lexical, syntactic and discourse information, but their performance was rather poor. With the development of powerful machine learning and the availability of large-scale datasets, neural models have delivered considerable gains in performance regarding both automatic evaluation metrics and human evaluation. Nevertheless, they still suffer from difficulties in understanding the long input document(s), and the generated summaries often contain factual inconsistencies and hallucinations, namely the content not existing in, or even contradicting the source document. One promising solution that we explore in this thesis is to inject linguistic knowledge into the neural models as a guidance. In particular, we focus on discourse, which reflects how the text is aggregated as a coherent document with specific central focus and fluent topical transport. While in the past decades, discourse has been shown to be helpful in non-neural summarizers, here we investigate the synergy between discourse and neural summarizers, with a special focus on the discourse structure (both explicit structures and RST structures) and the entities of the document(s). Specifically, we propose a series of methods to inject discourse information into neural summarizers regarding both the document understanding and summary generation steps, covering both extractive and abstractive methods. A large set of experimental results indicate that our methods effectively improve not only the summarizers overall performance, but also their efficiency, generality, and factualness. Conversely, by evaluating the discourse trees induced by neural summarizers on several human-annotated discourse tree datasets, we show that such summarizers do capture discourse structural information implicitly.
Item Citations and Data
Attribution-NonCommercial-NoDerivatives 4.0 International