Document-Aware Graph Models for Query-Oriented Multi-document Summarization

Sentence ranking is the issue of most concern in document summarization. In recent years, graph-based summarization models and sentence ranking algorithms have drawn considerable attention from the extractive summarization community due to their capability of recursively calculating sentence significance from the entire text graph that links sentences together rather than relying on single sentence alone. However, when dealing with multi-document summarization, existing sentence ranking algorithms often assemble a set of documents into one large file. The document dimension is ignored. In this work, we develop two alternative models to integrate the document dimension into existing sentence ranking algorithms. They are the one-layer (i.e. sentence layer) document-sensitive model and the two-layer (i.e. document and sentence layers) mutual reinforcement model. While the former implicitly incorporates the document’s influence in sentence ranking, the latter explicitly formulates the mutual reinforcement among sentence and document during ranking. The effectiveness of the proposed models and algorithms are examined on the DUC query-oriented multi-document summarization data sets.