Subtopic-driven Multi-Document Summarization

Abstract

In multi-document summarization, a set of documents to be summarized is assumed to be on the same topic, known as the underlying topic in this paper. That is, the underlying topic can be collectively represented by all the documents in the set. Meanwhile, different documents may cover various different subtopics and the same subtopic can be across several documents. Inspired by topic model, the underlying topic of a document set can also be viewed as a collection of different subtopics of different importance. In this paper, we propose a summarization model called STDS. The model generates the underlying topic representation from both document view and subtopic view in parallel. The learning objective is to minimize the distance between the representations learned from the two views. The contextual information is encoded through a hierarchical RNN architecture. Sentence salience is estimated in a hierarchical way with subtopic salience and relative sentence salience, by considering the contextual information. Top ranked sentences are then extracted as a summary. Note that the notion of subtopic enables us to bring in additional information (e.g., comments to news articles) that is helpful for document summarization. Experimental results show that the proposed solution outperforms state-of-the-art methods on benchmark datasets.

A Spectral Method for Unsupervised Multi-Document Summarization

“…We adopt 100 words as the length limit in the DUC dataset, instead of 665 bytes specified by the official task. The change has also been made to provide the same setting for evaluating various methods in Hong et al (2014) and Zheng et al (2019). For the Yelp dataset, we set the limit to be the 99.5 th percentile less than the maximum length of any document; for Multi-News, the limit is set as 300 words.…”

“…In the experiments, we report the different combinations of ROUGE scores for each dataset, which have been recommended and adopted by previous works. Specifically, the recall scores of R-1,2,4 will be reported for the DUC 2004 dataset according to Hong et al (2014), Wang et al (2017) and Zheng et al (2019); the F1 scores of R-1,2,L will be reported for Yelp as in Chu and Liu (2019); the F1 scores of R-1,2,SU4 will be reported for Multi-News as in Fabbri et al (2019). The toolkit for computing ROUGE metrics is ROUGE-1.5.5 8 and its option is set to be '-m -c 95 -r 1000 -f A -p 0.5 -t 0'.…”

“…Regarding the document cluster as a manifold structure, (Wan et al, 2007;Cheng et al, 2011; use the manifold ranking process on data. There are also quite a few neural architecture based models for a hidden semantic representation of sentences, documents or subtopics, such as (Ma et al, 2016;Li et al, 2017a,b;Zheng et al, 2019). Spectral optimization.…”

“…For instance, Lin and Bilmes ( 2011) is a bi-criteria case that imposes the submodularity constraint on each criterion. Multi-criteria loss functions in neuralnetwork-based methods (Ma et al, 2016;Chu and Liu, 2019;Zheng et al, 2019) include the reconstruction errors from different spaces. In the above cases, the overall objective functions used include some hyperparameters for gluing singletons.…”

Enhancing Extractive Text Summarization with Topic-Aware Graph Neural Networks

“…Wei et al (2012) proposed to build a document graph consisting of words, sentences, and topic nodes and learn the graph with Markov chain. Zheng et al (2019) proposed to summarize multiple documents by mining cross-document subtopics. Narayan et al (2018) recommended enriching word representation with topical information.…”