zoom_in Click to enlarge

BertSum

▲ ▼

language

description BertSum Overview

BertSum is a neural network summarization approach utilizing pre-trained BERT embeddings to generate abstractive summaries of input texts, often prioritizing semantic understanding over direct extraction.

help BertSum FAQ

Is BertSum an extractive or abstractive summarizer?

The original BERTSUM work by Yang Liu focused on extractive summarization, selecting complete sentences from the source. A later 2019 paper by Liu and Mirella Lapata extended the framework to both extractive and abstractive models.

How does BertSum adapt BERT to a document with many sentences?

It inserts sentence-boundary markers and uses their contextual representations to classify sentences for inclusion in the summary. The model also modifies segment embeddings so adjacent sentences can be distinguished.

Which datasets were used to evaluate BertSum?

The published work evaluated summarization on news datasets including CNN/DailyMail and the New York Times corpus. Performance was commonly reported with ROUGE metrics against human-written reference summaries.

Can BertSum summarize documents longer than BERT's input limit?

Not in a single unmodified pass, because standard BERT has a limited token window. Long reports generally require truncation, chunking, a hierarchical approach, or a model designed for longer contexts.

swap_horiz

Looking for BertSum alternatives? Compare top competitors ranked & scored

arrow_forward