texturizer 0.1.9

Python command line application to add text features to a CSV or TSV dataset.

texturizer

Status - Functional

This is an application to add features to a dataset that are derived from processing the content of existing columns of text data. It is specifically designed for adding somewhat bespoke and unusual features that are not particularly well identified by n-gram or word embedding approaches.

It will accept a CSV, TSV or XLS file and output an extended version of the dataset with additional columns appended. When run with default settings it will produce a small number of very simple numerical summaries.

Additional feature flags unlock features that are more computationally intensive and generally domain specific.

Released and distributed via setuptools/PyPI/pip for Python 3.

Additional detail available in the documentation

TODO

Current features are all derived from single records. Future development will add these
in some sense relative to a corpus.

* Add capacity to generate features relative to corpus averages
* Add capacity for comparison features to be generated relative to reference text(s)
* Investigate functionality for working with unix shell pipes and streams

Distribution

Released and distributed via setuptools/PyPI/pip for Python 3.

Resources & Dependencies

For Part of Speech Tagging and Word Embeddings we use spacy

Note: After install you will need to get spaCy to download the English model.

sudo python3 -m spacy download en

For string based text comparisons we use jellyfish and textdistance

Features

Each type of feature can be unlocked through the use of a specific command line switch:

  -topics            Default: False. Indicators for words from common topics.
  -topics=count                      Count matching words from common topics.
  -topics=normalize                  Count matching topic key words and normalize over topics.
  -traits            Default: False. Word usage for personality traits.
  -reason            Default: False. Word usage for reasoning: premises, conclusions.
  -rhetoric          Default: False. Word usage for rhetorical devices.
  -pos               Default: False. Part of speech proportions.
  -literacy          Default: False. Checks for simple literacy markers.
  -profanity         Default: False. Profanity check flags.
  -sentiment         Default: False. Words counts for positive and negative sentiment.
  -scarcity          Default: False. Word scarcity scores.
  -emoticons         Default: False. Emoticon check flags.
  -embedding         Default: False. Aggregate of Word Embedding Vectors.
  -embedding=normalize               Normalised Aggregate of Word Embedding Vectors.
  -comparison        Default: False. Cross-column comparisons using edit distances.

Usage

You can use this application multiple ways

Runner

Use the runner without installing the application. The following example will generate all features on the test data.

./texturizer-runner.py -columns=question,answer -pos -literacy -traits -reason -rhetoric -profanity -emoticons -embedding -sentiment -scarcity -comparison -topics=count data/test.csv > data/output.csv

This will send the time performance profile to STDERR as shown below:

Computation Time Profile for each Feature Set
---------------------------------------------
simple               0:00:00.498634
comparison           0:00:00.536637
profanity            0:00:00.496018
sentiment            0:00:03.310224
scarcity             0:00:00.523863
emoticons            0:00:00.219341
embedding            0:00:43.456939
topics               0:00:05.285120
traits               0:00:00.298902
reason               0:00:00.305391
rhetoric             0:00:02.988197
pos                  0:00:40.981175
literacy             0:00:00.371007

As you can see the part of speech (POS) features and word embeddings are the most time consuming to generate. In both instances these rely on the SpacY package to process the text block. For the moment it would be advised to avoid using them on very large datasets.

TODO: improve performance of these feature generators.

Directory as package

Alternatively, you can invoke the directory as a package:

python -m texturizer -columns=question,answer data/test.csv > data/output.csv

From Install

Or you can simply install the package and use the command line application directly

texturizer -h

Will print out the help

Installation

Installation from the source tree:

python setup.py install

(or via pip from PyPI):

pip install texturizer

You will then need to run the POST INSTALL SCRIPT to install the required Spacy Model (otherwise the POS features cannot be calculated).

Now, the texturizer command is available::

texturizer -columns=question,answer -topics data/test.csv > data/output.csv

This will take the Input CSV, calculate some simple summary features and produce an Output CSV with features appended as new columns.

For more complicated features see the additional options (outlined above).

Acknowledgements

Python package built using the bootstrap cmdline template by jgehrcke