Dual Encoding for Abstractive Text Summarization.

Area of Science:

• Natural Language Processing
• Artificial Intelligence

Background:

• Sequence-to-sequence attentional models are effective for abstractive text summarization.
• Existing models often use a single encoder, potentially limiting summary diversity.

Purpose of the Study:

• To propose and evaluate a novel dual encoding model for abstractive text summarization.
• To enhance summary quality by reducing repetition in long sequence generation.

Main Methods:

• A dual encoder architecture comprising primary and secondary encoders is introduced.
• The primary encoder performs coarse encoding, while the secondary encoder refines encoding based on input and previously generated text.
• Combined encodings are fed into a decoder for summary generation.

Main Results:

• The dual encoding model demonstrated improved performance against existing methods on CNN/DailyMail and DUC 2004 datasets.
• The model effectively decreases repetition phenomena in long sequence generation.

Conclusions:

• The proposed dual encoding model offers a promising advancement in abstractive text summarization.
• This approach generates more diverse summaries and mitigates repetition issues.