Text Summarization With Graph Attention Networks

Mohammadreza Ardestani; Yllias Chali

cs.CLApr 4, 2026

Text Summarization With Graph Attention Networks

Authors: Mohammadreza Ardestani, Yllias Chali

Organizations: Department of Computer Science, University of Lethbridge, Lethbridge, Alberta, Canada

Abstract

This study aimed to leverage graph information, particularly Rhetorical Structure Theory (RST) and Co-reference (Coref) graphs, to enhance the performance of our baseline summarization models. Specifically, we experimented with a Graph Attention Network architecture to incorporate graph information. However, this architecture did not enhance the performance. Subsequently, we used a simple Multi-layer Perceptron architecture, which improved the results in our proposed model on our primary dataset, CNN/DM. Additionally, we annotated XSum dataset with RST graph information, establishing a benchmark for future graph-based summarization models. This secondary dataset posed multiple challenges, revealing both the merits and limitations of our models.

Figures & tables

Figure 1: Proposed model overview and its MLP-based variation.

	Model	Precision	Recall	F1
Longformer	without graph	28.98	56.94	38.41
	w. GAT for RST	28.75	54.58	37.66
	w. GAT for Coref	26.57	57.26	36.30
	w. GAT for RST & Coref	27.62	57.94	37.40
	w. MLP for RST & Coref	29.81	54.94	38.65
BERT	without graph	28.61	65.00	39.74

Table 1: Test results of the encoder models on CNNDM dataset.

	Model	Precision	Recall	F1
Longformer
	without graph	30.71	50.04	38.06
	w. GAT for RST	30.89	48.06	37.61
	w. MLP for RST	30.86	49.74	38.09
BERT	without graph	28.98	56.94	38.41
	w. GAT for RST	28.75	54.58	37.66

Table 2: Test results of the encoder models on XSum dataset.

	Model	R-1	R-2	R-L
Longformer	without graph	43.60 ± 0.23	20.77 ± 0.25	40.60 ± 0.22
	w. GAT for RST	43.01 ± 0.23	20.35 ± 0.25	40.00 ± 0.21
	w. GAT for Coref	43.04 ± 0.22	20.32 ± 0.25	39.98 ± 0.24
	w. GAT for RST & Coref	43.30 ± 0.23	20.55 ± 0.23	40.30 ± 0.23
	w. MLP for RST & Coref	43.73 ± 0.22	20.91 ± 0.26	40.72 ± 0.24
BERT	without graph	43.43 ± 0.24	20.62 ± 0.24	40.40 ± 0.23

Table 3: Test results of the encoder-decoder models on CNNDM dataset.

	Model	R-1	R-2	R-L
Longformer
	without graph	36.56 ± 0.25	14.28 ± 0.23	28.29 ± 0.23
	w. GAT for RST	36.24 ± 0.26	14.14 ± 0.23	28.14 ± 0.24
	w. MLP for RST	36.82 ± 0.27	14.66 ± 0.24	28.67 ± 0.26
BERT	without graph	36.41 ± 0.28	14.23 ± 0.25	28.37 ± 0.27
	w. GAT for RST	35.33 ± 0.25	13.30 ± 0.23	27.45 ± 0.25

Table 4: Test results of encoder-decoder models on XSum dataset.

Oracle
Model	R-1	R-2	R-L
Lead3	40.42	17.62	36.67
Oracle (EDUs)	61.61	37.82	59.27
Extractive
DiscoBERT [ Xu et al., 2020 ]	43.38	20.44	40.21
DiscoBERT w Coref	43.58	20.64	40.42

Table 5: Comparing our best model with previous models as our baseline on CNNDM.

Oracle
Model	R-1	R-2	R-L
LEAD1 †	16.30	1.60	11.95
Oracle Sentence [ Liu and Lapata, 2019 ]	29.79	8.81	22.66
Oracle EDU (ours)	36.03	11.47	30.86
Abstractive
BERTSUMABS [ Liu and Lapata, 2019 ]	38.76	16.33	31.15

Table 6: Comparing our best model with baseline models on XSum.

CNNDM
Model	R-1	R-2	R-L
BART-base	41.44	18.49	38.32
Oracle (EDUs)	61.61 ( $\uparrow$ )	37.82 ( $\uparrow$ )	59.27 ( $\uparrow$ )
Zeroshot BART on oracle EDUs	53.43 ( $\uparrow$ )	32.20 ( $\uparrow$ )	51.05 ( $\uparrow$ )
Finetuned BART on oracle EDUs	60.89 ( $\uparrow$ )	37.59 ( $\uparrow$ )	57.47 ( $\uparrow$ )
XSum

Table 7: Comparing the performance of extractive with abstractive approaches on CNNDM and XSum.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2: Message aggregation and update in GNNs for a single node from its local neighborhood.

Text	n=1	n=2	n=3
Oracle EDUs	0.000	0.015	0.071
Selected EDUs	0.000	0.012	0.053
Predicted Summaries	0.005	0.060	0.171
Reference Summaries	0.008	0.173	0.491

Appendix

Table 8: Novel n-grams proportions of our best model on CNN/DM

Model	BERT Score	BART Score
Our Best Model	0.87	-3.73

Appendix

Table 9: BERT and BART scores of our best model on CNN/DM.

Figure 3: XSum data processing pipeline.

Dataset	Average		Empty			Total Records
	EDUs	Summary	RST Graph	Oracle Label	Both
Train	52.35	21.10	121	1944	102	204017
Test	41.85	21.10	7	85	5	11333
Valid	41.39	21.13	10	96	8	11327

Appendix

Table 10: Statistics of the Processed XSum.

Figure 4: Details of GNN and classification modules.

Figure 5: RST and Coref graph examples. Document A is selected from the test set of our processed XSum dataset. Part B, adapted from [ Xu et al., 2020 ] , shows Coref edges for a particular named entity.

Text Summarization

Text summarization develops methods for condensing documents while preserving the information, perspective, and evidence readers need.

See all 142 papers

CGS: Configurable Graph Summarization with Bounded Neighborhood Loss and Query Support
Jul 13, 2026Shubhadip Mitra, Sona Elza Simon, C Oswald +2
Blue Yonder India Pvt. Ltd., Bengaluru, India · Centre for Machine Intelligence and Data Science, Indian Institute of Technology Bombay, Mumbai, India · Dept. of Computer Science and Engineering, National Institute of Technology Tiruchirappalli, Tiruchirapalli, India +2
Position: Graph Condensation Needs a Reset -- Move Beyond Full-dataset Training and Model-Dependence
May 17, 2026Mridul Gupta, Samyak Jain, Vansh Ramani +2Graph Neural Networks Graph Representation Learning
Yardi School of Artificial Intelligence, IIT Delhi, India · Department of Computer Science and Engineering, IIT Delhi, India · Department of Chemical Engineering, IIT Delhi, India
Rethinking Message Passing as Retrieval for Text-Attributed Graph Learning
Aug 27, 2026Jintang Li, Yuhong Chen, Ruofan Wu +4Graph Neural Networks Text-Attributed Graph Learning
School of Computer Science and Engineering of Central South University, Changsha 410000, China

Jul 13, 2026J/K move · Enter open · S save

CGS: Configurable Graph Summarization with Bounded Neighborhood Loss and Query Support

Shubhadip Mitra, Sona Elza Simon, C Oswald, Arnab Bhattacharya+1

Blue Yonder India Pvt. Ltd., Bengaluru, India · Centre for Machine Intelligence and Data Science, Indian Institute of Technology Bombay, Mumbai, India · Dept. of Computer Science and Engineering, National Institute of Technology Tiruchirappalli, Tiruchirapalli, India+2

Open paper AskPDF arXiv

Given a large graph, how to generate a compact summary graph that is configurable by the user and supports multiple graph queries with either no loss or with high accuracy? The ever growing size of graph datasets makes the above question on graph summarization very pertinent. Although, there are several approaches, there does not exist a configurable graph summarization method that offers high compression along with support for multiple graph queries on the summary graph with high accuracy, and allows the user to configure the summarization based on: (1) lossless or lossy summarization, (2) amount of tolerable neighborhood loss, (3) the type of loss it can tolerate, in terms of false positive edges (i.e., extra edges), false negative edges (i.e., missing edges), or neither, in both the (a) reconstructed graph and the (b) query answers. To overcome these limitations, we propose a novel graph summarization framework CGS (Configurable Graph Summarizer) that builds upon the idea of aggregating nodes with common neighborhoods. The CGS framework consists of three summarization variants, CGS-E, CGS-I and CGS-U. While CGS-E is a lossless scheme, CGS-I and CGS-U are lossy schemes that allow reconstruction of the input graph with no false positive edges and no false negative edges, respectively. To bound the graph reconstruction loss, we introduce a user-specified parameter neighborhood loss tolerance threshold, that limits the maximum loss allowed in the neighborhood of each node. This allows graph reconstruction and neighborhood query evaluation with either no loss or with bounded loss guarantees. Empirical evaluation on several synthetic and real-world graphs shows that CGS offers superior summarization than the state-of-the-art methods, and can answer graph queries with fairly high accuracy and efficiency.

Abstract

Text Summarization

Explore similar work

CGS: Configurable Graph Summarization with Bounded Neighborhood Loss and Query Support

Position: Graph Condensation Needs a Reset -- Move Beyond Full-dataset Training and Model-Dependence

Rethinking Message Passing as Retrieval for Text-Attributed Graph Learning