| E. Arvanitis, G. Drakopoulos, L. Theodorakopoulos, S. Sioutas, and Ph. Mylonas |
| Extractive Document Summarization With Graph Neural Networks And Topic Modeling In PyTorch |
| IEEE International Conference on Big Data (IEEE BigData 2025), December 8-11, 2025, Macau, China |
|
ABSTRACT
|
| Algorithmic text summarization task in natural language processing aims to represent a given text in a shorter and suitable form for a human reader by locating sentences of interest while removing redundant ones. Regarding information retrieval, summary extraction can be beneficial in large document spaces by contributing to formulating queries with the least possible number of most representative terms, rendering thus a term augmentation more efficient. There are many approaches to this problem regarding the part of text summary model, such as sequence models, transformers, and autoencoders. These models are limited by the difficulty to capture inter-sentence structure, which is crucial especially in long texts. This importance stems from the fact that proximity in the sentence typically denotes semantic association. The proposed summarization framework relies on a combination of graph neural networks and topic modelling in order to represent texts with graphs, an intuitive way to capture relationships between sentences. Two strategies of obtaining attribute vectors, namely extracting them from each vertex and alternatively extracting them from each edge as a function of the features of the adjacent vertices, have been tested. Both strategies yielded comparable results as evaluated by the ROUGE family of metrics computed on a benchmark dataset with ground truth. Moreover, in all cases the resulting summaries were considerably shorter than the original text.
|
| 08 December , 2025 |
| E. Arvanitis, G. Drakopoulos, L. Theodorakopoulos, S. Sioutas, and Ph. Mylonas, "Extractive Document Summarization With Graph Neural Networks And Topic Modeling In PyTorch", IEEE International Conference on Big Data (IEEE BigData 2025), December 8-11, 2025, Macau, China |
[
BibTex] [
Print] [
 Back] |
