This approach applies spectral clustering to compress graph for extraction based summarization. For that, they create graphs where nodes are sentences and edges are connections. A sentence node is connected to another one if there are one of the four conditions- 1) Deverbal Noun Reference 2) Entity Continuation 3) Discourse Markers (however, meanwhile, furthermore) 4) sentence similarity (embedding-> word movers distance of sentences )
After graph is created, they apply spectral clustering among the sentences of the documents and separate the clusters. Then in each cluster multi-sentence compression (MSC) applied to find single summary sentence from each cluster.
They evaluated their model using ROGUE analysis and found their system works better compared to LexRank, TextRank, MMR, Centroid.
For Human analysis they conducted survey based on four criteria's: Fluency(coherence of the sentence), Coverage (all aspects are covered), Redundancy (if some solutions are redundant)
Contributions of The Paper
First unsupervised summarization method which constructs sentence graphs by incorporating both linguistic knowledge and deep neural representations. It assumes that a summary sentence can be created by compressing a within-graph cluster.
Publisher
ACM
Link to The Paper
https://dl.acm.org/doi/abs/10.1145/3397271.3401327
Name of The Authors
Zhao JLiu MGao L et al.
Year of Publication
2020
Summary
This approach applies spectral clustering to compress graph for extraction based summarization. For that, they create graphs where nodes are sentences and edges are connections. A sentence node is connected to another one if there are one of the four conditions- 1) Deverbal Noun Reference 2) Entity Continuation 3) Discourse Markers (however, meanwhile, furthermore) 4) sentence similarity (embedding-> word movers distance of sentences ) After graph is created, they apply spectral clustering among the sentences of the documents and separate the clusters. Then in each cluster multi-sentence compression (MSC) applied to find single summary sentence from each cluster. They evaluated their model using ROGUE analysis and found their system works better compared to LexRank, TextRank, MMR, Centroid. For Human analysis they conducted survey based on four criteria's: Fluency(coherence of the sentence), Coverage (all aspects are covered), Redundancy (if some solutions are redundant)
Contributions of The Paper
First unsupervised summarization method which constructs sentence graphs by incorporating both linguistic knowledge and deep neural representations. It assumes that a summary sentence can be created by compressing a within-graph cluster.
Comments
No response