zoom_in Click to enlarge

LexRank

▲ ▼

description LexRank Overview

LexRank is an unsupervised, graph-based method for extractive text summarization. It represents sentences as nodes, connects sentences according to lexical similarity, and applies an eigenvector centrality calculation related to PageRank to identify sentences that are important within the resulting network. The method is used to select representative sentences from one document or a collection of documents without generating new wording.

help LexRank FAQ

What does LexRank actually select from a document?

LexRank is extractive: it selects existing sentences instead of writing new ones. Günes Erkan and Dragomir Radev introduced it in a 2004 Journal of Artificial Intelligence Research paper. [JAIR LexRank paper](https://mlanthology.org/jair/2004/erkan2004jair-lexrank/)

How does LexRank rank sentences?

Each sentence becomes a graph node, and edges reflect lexical similarity, commonly cosine similarity. Eigenvector centrality, related to PageRank, then gives high scores to sentences connected to other important sentences. [JAIR LexRank paper](https://mlanthology.org/jair/2004/erkan2004jair-lexrank/)

Does LexRank need labeled training examples?

No. LexRank is unsupervised, so it can rank sentences without human-written summaries or labeled training data, although tokenization and similarity settings still affect the result. [JAIR LexRank paper](https://mlanthology.org/jair/2004/erkan2004jair-lexrank/)

How is LexRank different from TextRank?

Both use graph ranking, but LexRank focuses on sentence similarity and eigenvector centrality for extractive summarization. The original LexRank evaluation included the DUC 2004 summarization tasks. [JAIR LexRank paper](https://mlanthology.org/jair/2004/erkan2004jair-lexrank/)

swap_horiz

Looking for LexRank alternatives? Compare top competitors ranked & scored

arrow_forward