/edm

Python package for understanding the difficulty of text classification datasets. (in CoNNL 2018)

Primary LanguagePythonGNU General Public License v2.0GPL-2.0

Wluper

Evolutionary Data Measures: Understanding the Difficulty of Text Classification Tasks

Authors: Ed Collins, Nikolai Rozanov, Bingbing Zhang

Contact: contact@wluper.com

In the paper of the corresponding name, we discuss how we used an evolutionary algorithm to discover which statistics about a text classification dataset most accurately represent how difficult that dataset is likely to be for machine learning models to learn. We presented there the difficulty measure which we discovered and have provided this Python package of code which can calculate it.

Installation

This code is pip-installable so can be installed on your machine by running:

pip3 install edm

The code requires Python 3 and NumPy.

It is recommended that you install this code in a virtualenv:

$ mkdir myvirtualenv/
$ virtualenv -p python3 myvirtualenv/
$ source bin/activate
(myvirtualenv) $ pip3 install edm

Running

To calculate the difficulty of a text classification dataset, you will need to provide two lists: one of sentences and one of labels. These two lists need to be the same length - i.e. every sentence has a label. Each item of data should be an untokenized string and each label a string.

>>> sents, labels = your_own_loading_function(PATH_TO_DATA_FILE)
>>> sents
["this is a positive sentence", "this is a negative sentence", ...]
>>> labels
["positive", "negative", ...]
>>> assert len(sents) == len(labels)
True

This code does not support the loading of data files (e.g. csv files) into memory - you will need to do this separately.

Once you have loaded your dataset into memory, you can receive a "difficulty report" by running the code as follows:

from edm import report

sents, labels = your_own_loading_function(PATH_TO_DATA_FILE)

print(report.get_difficulty_report(sents, labels))

Note that if your dataset is very large, then counting the words of the dataset may take several minutes. The Amazon Reviews dataset from Character-level Convolutional Networks for Text Classification by Xiang Zhang, Junbo Zhao and Yann LeCun, 2015 which contains 3.6 million Amazon reviews takes approximately 15 minutes to be processed and the difficulty report created. A loading bar will be displayed while the words are counted.

Citation

The official citation from CoNLL 2018 in Belgium. Please use this for citation:

@inproceedings{collins-etal-2018-evolutionary,
    title = "Evolutionary Data Measures: Understanding the Difficulty of Text Classification Tasks",
    author = "Collins, Edward  and
      Rozanov, Nikolai  and
      Zhang, Bingbing",
    booktitle = "Proceedings of the 22nd Conference on Computational Natural Language Learning",
    month = oct,
    year = "2018",
    address = "Brussels, Belgium",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/K18-1037",
    doi = "10.18653/v1/K18-1037",
    pages = "380--391",
    abstract = "Classification tasks are usually analysed and improved through new model architectures or hyperparameter optimisation but the underlying properties of datasets are discovered on an ad-hoc basis as errors occur. However, understanding the properties of the data is crucial in perfecting models. In this paper we analyse exactly which characteristics of a dataset best determine how difficult that dataset is for the task of text classification. We then propose an intuitive measure of difficulty for text classification datasets which is simple and fast to calculate. We empirically prove that this measure generalises to unseen data by comparing it to state-of-the-art datasets and results. This measure can be used to analyse the precise source of errors in a dataset and allows fast estimation of how difficult a dataset is to learn. We searched for this measure by training 12 classical and neural network based models on 78 real-world datasets, then use a genetic algorithm to discover the best measure of difficulty. Our difficulty-calculating code and datasets are publicly available.",
}