Mailgun library to extract message quotations and signatures.
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 😄
Filters have been put in place to allow you to customize talon.
- talon_email_signature_patterns
- talon_email_signature_words
- talon_email_footer_patterns
- talon_email_footer_words
- talon_email_footer_lines
- talon_email_footer_lines_ratio
Here’s how you initialize the library and extract a reply from a text message:
import talon
from talon import quotations
talon.init()
text = """Reply
-----Original Message-----
Quote"""
reply = quotations.extract_from(text, 'text/plain')
reply = quotations.extract_from_plain(text)
# reply == "Reply"To extract a reply from html:
html = """Reply
<blockquote>
<div>
On 11-Apr-2011, at 6:54 PM, Bob <[email protected]> wrote:
</div>
<div>
Quote
</div>
</blockquote>"""
reply = quotations.extract_from(html, 'text/html')
reply = quotations.extract_from_html(html)
# reply == "<html><body><p>Reply</p></body></html>"Often the best way is the easiest one. Here’s how you can extract signature from email message without any machine learning fancy stuff:
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:
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
python train.pyor
from talon.signature import EXTRACTOR_FILENAME, EXTRACTOR_DATA
from talon.signature.learning.classifier import train, init
train(init(), EXTRACTOR_DATA, EXTRACTOR_FILENAME)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.
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:
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).
The library is inspired by the following research papers and projects: