Core Concepts
Relation triples are utilized for interpretable summarization, enhancing understanding and context.
Abstract
Abstract:
RTSUM framework uses relation triples for summarization.
Salient relation triples selected via multi-level salience scoring.
Web demo provides fine-grained interpretations.
Introduction:
Text summarization crucial in information overload era.
Abstractive methods lack interpretability compared to extractive methods.
Interpretability in Summarization:
Extractive methods offer advantages in interpretability.
Selecting entire sentences may include unnecessary information.
Leveraging Relation Triples:
Relation triples used as basic unit for summarization.
Selection-and-sentencification approach combines extractive and abstractive methods.
Unsupervised Relation Triple-based Summarization Framework:
RTSUM identifies salient relation triples from various textual units.
Neural text-to-text architecture used for sentencification.
Demo: Interpretable Summarizing Tool:
Tool visualizes salience at three levels: sentences, relation triples, phrases.
Customization options available for users.
Implementation Details:
Text graph construction filters out less confident relation triples.
RTSUM ranks and selects top-K salient relation triples based on final scores.
Related Work:
Unsupervised extractive summarization focuses on key sentences using text graphs.
Unsupervised abstractive summarization utilizes autoencoding architecture or PLMs.
Conclusion:
RTSUM framework leverages relation triples for effective and interpretable summarization.
Stats
"Formally, we present an unsupervised Relation Triple-based Summarization framework, named RTSUM."
"The number of relation triples to be selected is set to K = 3."