talon

Mailgun library to extract message quotations and signatures.

Plain-text quotation headers are recognized in several languages, including Ukrainian reply headers produced by localized mail clients.

If you ever tried to parse message quotations or signatures you know that absence of any formatting standards in this area could make this task a nightmare. Hopefully this library will make your life much easier. The name of the project is inspired by TALON - multipurpose robot designed to perform missions ranging from reconnaissance to combat and operate in a number of hostile environments. That’s what a good quotations and signature parser should be like :smile:

Installation

Talon supports Python 3.7 and newer.

Quotations and brute-force signature extraction:

.. code:: sh

pip install talon

Machine-learning signature extraction needs numpy, scipy, scikit-learn, and joblib. Install those extras with:

.. code:: sh

pip install "talon[ml]"

Talon includes inline type annotations (PEP 561_), so type checkers such as mypy pick them up automatically.

.. _PEP 561: https://peps.python.org/pep-0561/

Usage

Here’s how you initialize the library and extract a reply from a text message:

.. code:: python

import talon
from talon import quotations

talon.init()

text =  """Reply

-----Original Message-----

Quote"""

reply = quotations.extract_from(text, 'text/plain')
# -- OR --
reply = quotations.extract_from_plain(text)
# reply == "Reply"

A signature can follow the quote. extract_from keeps that signature with the reply. extract_quotation returns the quoted span between them, including the splitter line:

.. code:: python

text = """Reply

On 11-Apr-2011, at 6:54 PM, Bob <[email protected]> wrote:

> Quote

--
Bob Smith"""

reply = quotations.extract_from_plain(text)
# reply == "Reply\n\n--\nBob Smith"

quotation = quotations.extract_quotation(text, 'text/plain')
# -- OR --
quotation = quotations.extract_quotation_plain(text)
# quotation == "On 11-Apr-2011, at 6:54 PM, Bob <[email protected]> wrote:\n\n> Quote"

To extract a reply from html:

.. code:: python

html = """Reply
<blockquote>

  <div>
    On 11-Apr-2011, at 6:54 PM, Bob &lt;[email protected]&gt; wrote:
  </div>

  <div>
    Quote
  </div>

</blockquote>"""

reply = quotations.extract_from(html, 'text/html')
# -- OR --
reply = quotations.extract_from_html(html)
# reply == "<html><body><p>Reply</p></body></html>"

quotation = quotations.extract_quotation_html(html)
# quotation is the removed <blockquote>, including "Quote"

Often the best way is the easiest one. Here’s how you can extract signature from email message without any machine learning fancy stuff:

.. code:: python

from talon.signature.bruteforce import extract_signature


message = """Wow. Awesome!
--
Bob Smith"""

text, signature = extract_signature(message)
# text == "Wow. Awesome!"
# signature == "--\nBob Smith"

Quick and works like a charm 90% of the time. For other 10% you can use the power of machine learning algorithms (requires talon[ml]):

.. code:: python

import talon
# don't forget to init the library first
# it loads machine learning classifiers
talon.init()

from talon import signature


message = """Thanks Sasha, I can't go any higher and is why I limited it to the
homepage.

John Doe
via mobile"""

text, signature = signature.extract(message, sender='[email protected]')
# text == "Thanks Sasha, I can't go any higher and is why I limited it to the\nhomepage."
# signature == "John Doe\nvia mobile"

For machine learning talon currently uses the scikit-learn_ library to build SVM classifiers. The core of machine learning algorithm lays in talon.signature.learning package. It defines a set of features to apply to a message (featurespace.py), how data sets are built (dataset.py), classifier’s interface (classifier.py).

Currently the data used for training is taken from our personal email conversations and from ENRON_ dataset. As a result of applying our set of features to the dataset we provide files classifier and train.data that don’t have any personal information but could be used to load trained classifier. Those files should be regenerated every time the feature/data set is changed.

To regenerate the model files, you can run

.. code:: sh

python train.py

or

.. code:: python

from talon.signature import EXTRACTOR_FILENAME, EXTRACTOR_DATA
from talon.signature.learning.classifier import train, init
train(init(), EXTRACTOR_DATA, EXTRACTOR_FILENAME)

Open-source Dataset

Recently we started a forge_ project to create an open-source, annotated dataset of raw emails. In the project we used a subset of ENRON_ data, cleansed of private, health and financial information by EDRM_. At the moment over 190 emails are annotated. Any contribution and collaboration on the project are welcome. Once the dataset is ready we plan to start using it for talon.

.. _scikit-learn: http://scikit-learn.org .. _ENRON: https://www.cs.cmu.edu/~enron/ .. _EDRM: http://www.edrm.net/resources/data-sets/edrm-enron-email-data-set .. _forge: https://github.com/mailgun/forge

Training on your dataset

talon comes with a pre-processed dataset and a pre-trained classifier. To retrain the classifier on your own dataset of raw emails, structure and annotate them in the same way the forge_ project does. Then do:

.. code:: python

from talon.signature.learning.dataset import build_extraction_dataset
from talon.signature.learning import classifier as c 

build_extraction_dataset("/path/to/your/P/folder", "/path/to/talon/signature/data/train.data")
c.train(c.init(), "/path/to/talon/signature/data/train.data", "/path/to/talon/signature/data/classifier")

Note that for signature extraction you need just the folder with the positive samples with annotated signature lines (P folder).

.. _forge: https://github.com/mailgun/forge

Research

The library is inspired by the following research papers and projects:

• http://www.cs.cmu.edu/~vitor/papers/sigFilePaper_finalversion.pdf
• http://www.cs.cornell.edu/people/tj/publications/joachims_01a.pdf