hmxv2/text-segmentation

Implementation of the paper: Text Segmentation as a Supervised Learning Task

Text Segmentation as a Supervised Learning Task

This repository contains code and supplementary materials which are required to train and evaluate a model as described in the paper Text Segmentation as a Supervised Learning Task

Downalod required resources

wiki-727K, wiki-50 datasets:

https://www.dropbox.com/sh/k3jh0fjbyr0gw0a/AADzAd9SDTrBnvs1qLCJY5cza?dl=0

word2vec:

https://drive.google.com/a/audioburst.com/uc?export=download&confirm=zrin&id=0B7XkCwpI5KDYNlNUTTlSS21pQmM

Fill relevant paths in configgenerator.py, and execute the script (git repository includes Choi dataset)

Creating an environment:

conda create -n textseg python=2.7 numpy scipy gensim ipython 
source activate textseg
pip install http://download.pytorch.org/whl/cu80/torch-0.3.0-cp27-cp27mu-linux_x86_64.whl 
pip install tqdm pathlib2 segeval tensorboard_logger flask flask_wtf nltk
pip install pandas xlrd xlsxwriter termcolor

How to run training process?

python run.py --help

Example:

python run.py --cuda --model max_sentence_embedding --wiki 

How to evaluate trained model (on wiki-727/choi dataset)?

python test_accuracy.py  --help

Example:

python test_accuracy.py --cuda --model <path_to_model> --wiki

How to create a new wikipedia dataset:

python wiki_processor.py --input <input> --temp <temp_files_folder> --output <output_folder> --train <ratio> --test <ratio>

Input is the full path to the wikipedia dump, temp is the path to the temporary files folder, and output is the path to the newly generated wikipedia dataset.

Wikipedia dump can be downloaded from following url:

https://dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2