texturizer 0.1.7

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 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. Indicators for presence of words from common topics.
• -topics=count. Counts of all word matches from common topics.
• -pos. Part of speech proportions in the text.
• -literacy. Checks for common literacy markers.
• -traits. Checks for common stylistic elements or traits that suggest personality type.
• -rhetoric. Checks for rhetorical devices used for persuasion
• -profanity. Profanity check flags.
• -sentiment. Sentiment word counts and score.
• -scarcity. Word scarcity scores.
• -emoticons. Emoticon check flags.
• -embedding. Word embedding vectors from the Spacy Package.
• -comparison. Cross-column comparisons using edit distance metrics

Usage

You can use this application multiple ways

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 -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               0:00:00.639998
comparison           0:00:00.492709
profanity            0:00:00.584252
sentiment            0:00:03.144435
scarcity             0:00:00.641485
emoticons            0:00:00.411252
embedding            0:00:21.821642
topics               0:00:03.554538
traits               0:00:00.619522
rhetoric             0:00:03.902830
pos                  0:00:27.074761
literacy             0:00:01.301165

As you can see the part of speech (POS) features and word embeddings are the most time consuming to generate. This is because they are both using the SpacY package to process the text block.

It is worth avoiding them on very large datasets.

Alternatively, you can invoke the directory as a package:

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

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

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