blester125 / text-rank

Text Rank with Python

Text Rank

Implementation of text rank algorithms described in (Mihalcea and Tarau, 2004)

text-rank's Issues

Add unittests

The following things need to be unittested

• Graph Building functions

• Graph classes themselves

• Text rank utilties

• text rank

• similarity functions

• task utilities

update the sentence graph constrution to support POS filters

Add documentation

There should be information in the read me about how text rank works and comments in the code about what part of the math each section is.

There should also be some comments about the differences between the delayed updates and updating as they go.

There should so be performance tables

Control the seed

We should be able to control the seed used to generate the initial ws for reproducibility

add pyproject.toml

we need a pyproject.toml to manage black configuration and to be future proof as the python packing community moves to it

Scipy sparse

Eventually we may need a SparseAdjacencyMatrix graph class to handle large graphs, the AdjacencyList might be enough to handle it though, it depends on speed.

This is low priority and should only be done when we find a graph we can't handle.

Add evaluation harness

Right now the results are just qualitative, I want to add an evaluation script that at least can be used for keyword extraction with F1. A more complex one for summarization (rouge, etc) would also be nice

Transistion Matrix

Look into storing the transistion scores for text rank instead of the denoms.

np.sum(adj / denom * ws, axis=0)

To

np.sum(trans * ws, axis=0)