StanzaGraphs is a Multilingual STANZA-based Summary and Keyword Extractor and Question-Answering System using TextGraphs and Neural Networks

The file summarizer.py generates dependency graphs of the form

<from, edge_label, to, sentence_number_in_which_it_occurs>

as .tsv (tab separated) or .pro (Prolog) files. Note that edge_labels are obtained by concatenating POS tags of the two nodes and the dependency_link label connecting them.

The file answerer.py encodes the tab separated outputs of summarizer.py into OneHotEncoded matrices, ready for use for machine learning tools as well as a simple question answering algorithm based on shared dependency edges with text graph of the sentences in the document.

A simple tensorflow keras based neural network implements a Trainer and an Inferencer using these matrices in file tf_answerer.py

A simple symbolic algorithm defines a baseline for answering questions related to a given text document (in folder texts) that has been processsed. A simple keras-based neural network is also included for the same task.

A proof-of-concept streamlit Web application is in file webapp.py. You can run it as shwon in the script:

webapp.sh

Before running, do:

pip3 install -U stanza

On the first run for a given language, a model (about 1GB) is downloaded, that might take a bit on slower networks. Once downloaded it goes in your home directory, in the the folder:

stanza_resources/

Also make sure dependencies in requirements.txt are all installed.

See more about Stanza at:

https://github.com/stanfordnlp/stanza