NOVEL2GRAPH

A workflow to extract Knowledge Graph from literary text

By Mellace Simone, Kanjirangat Vani and Antonucci Alessandro

This pipeline is a semi-supervised algorithm, which receives a book as input and it discovers: the main characters, the main relations between characters, a Knowledge Graph including these information and more interesting material. As final result, you will obtain an interface with which you can monitor and explore your results.

This work is based on Temporal Embeddings and Transformer Models for Narrative Text Understanding, Relation Clustering in Narrative Knowledge Graphs and NOVEL2GRAPH: Visual Summariesof Narrative Text Enhanced by Machine Learning.

Quickstart

Clone the project
Download the latest version of https://nlp.stanford.edu/software/CRF-NER.shtml#Download
Unzip the downloaded folder stanford-ner-20XX-XX-XX and put it in ./libraries
The same as before following https://stanfordnlp.github.io/CoreNLP#Quickstart
Install requirements.txt
Start test_static_dynamic_embedding.py and test_relations_clustering.py providing a book in txt format
[Optional] Run test_relations_classifier.py or test_Bert_family_cls.py to discover more information about the book
Run GUI/test_dash.py to start the interface on your machine

Start

First perform the static and dynamic embedding of your book by typing:

$ python Code/test_static_dynamic_embedding.py myBook.txt

Secondly generate the Knowledge Graph of your book by typing:

$ python Code/test_relations_clustering.py myBook.txt

Finally run the interface to analyze and interact with your results:

$ python Code/GUI/test_dash.py myBook.txt

Output

Results are generated in Data folder, in particular this folder could contain these others sub-folders:

clust&Dealias/bookName/: which for each book you run the algorithm with, contains the following files:
- bookName_occurrences.csv contains a list of character names and occurrences;
- bookName_more_than_1.csv as before but only with names occurring more than once;
- bookName_clusters.csv contains all clusters (each row shows a cluster id, the contained names and the occurrences);
- bookName_out.txt contains the coreferenced story in which originals names are replaced with an identifier ("CHARACTERX")
- *.pkl are cached data.
embedding:
- embeddings cache data containing the embedding of the book;
- models/bookName/chapters contains dynamic and static models trained on the slices;
- slices/bookName/chapters contains slices (text divided in chapters) which are use to train each model.
embedRelations/book/:
- bookName_report_X.txt contains all relations involving X characters;
- bookName_embeddings_X.pkl contains embedded relations data;
- bookName_zero_char_sentences.pkl contains sentences with 0 characters;
- bookName_few_char_sentences.pkl contains sentences with fewer characters than desired (you can specify the desired amount in the code);
- bookName_right_char_sentences.pkl contains sentences with the right number of characters than desired;
- bookName_more_char_sentences.pkl contains sentences with more characters than desired.
family_relations:
- some material for relations extracted with test_Bert_family_cls.py
scraping/bookName:
- amazon.csv: a book's reviews from amazon books;
- goodreads.csv: a book's reviews from Goodreads;
- librarything.csv: a book's reviews from LibraryThing;
- wiki.csv: a book's reviews from Wikipedia.

Others

Remember to use the Shift-Reduce Parser for Stanford CoreNLP from https://nlp.stanford.edu/software/srparser.html
Some Stanford's extentions require Java x64 bit

Bugs fix

To install Graphviz: $ sudo apt install python-pydot python-pydot-ng graphviz
To install mysql-config: $ sudo apt-get install libmysqlclient-dev
To install resource stopwords (or punkt), please use the NLTK Downloader:

import nltk
nltk.download('stopwords')
nltk.download('punkt')

Often it is useful to download:
- python -m spacy download en
- python -m spacy download en_core_web_sm

Edge Labelling Code

Contains the scripts and files for Edge Labelling

IDSIA/novel2graph