Merge pull request #177 from mailgun/obukhov-sergey-patch-1

Update Readme with how to retrain on your own data

.gitignore

@@ -39,6 +39,8 @@ nosetests.xml
 /.emacs.desktop
 /.emacs.desktop.lock
 .elc
+.idea
+.cache
 auto-save-list
 tramp
 .\#*

README.rst

@@ -129,6 +129,22 @@ start using it for talon.
 .. _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
 --------

setup.py

@@ -29,7 +29,7 @@ class InstallCommand(install):
 
 
 setup(name='talon',
-      version='1.3.7',
+      version='1.4.5',
       description=("Mailgun library "
                    "to extract message quotations and signatures."),
       long_description=open("README.rst").read(),

talon/html_quotations.py

@@ -94,6 +94,12 @@ def cut_microsoft_quote(html_message):
         #outlook 2007, 2010 (american)
         "//div[@style='border:none;border-top:solid #B5C4DF 1.0pt;"
         "padding:3.0pt 0in 0in 0in']|"
+        #outlook 2013 (international)
+        "//div[@style='border:none;border-top:solid #E1E1E1 1.0pt;"
+        "padding:3.0pt 0cm 0cm 0cm']|"
+        #outlook 2013 (american)
+        "//div[@style='border:none;border-top:solid #E1E1E1 1.0pt;"
+        "padding:3.0pt 0in 0in 0in']|"
         #windows mail
         "//div[@style='padding-top: 5px; "
         "border-top-color: rgb(229, 229, 229); "

talon/quotations.py

@@ -38,10 +38,14 @@ RE_ON_DATE_SMB_WROTE = re.compile(
             'Op',
             # German
             'Am',
+            # Portuguese
+            'Em',
             # Norwegian
             u'På',
             # Swedish, Danish
             'Den',
+            # Vietnamese
+            u'Vào',
         )),
         # Date and sender separator
         u'|'.join((
@@ -62,8 +66,12 @@ RE_ON_DATE_SMB_WROTE = re.compile(
             'schreef','verzond','geschreven',
             # German
             'schrieb',
+            # Portuguese
+            'escreveu',
             # Norwegian, Swedish
             'skrev',
+            # Vietnamese
+            u'đã viết',
         ))
     ))
 # Special case for languages where text is translated like this: 'on {date} wrote {somebody}:'
@@ -143,7 +151,7 @@ RE_FROM_COLON_OR_DATE_COLON = re.compile(u'(_+\r?\n)?[\s]*(:?[*]?{})[\s]?:[*]?.*
 RE_ANDROID_WROTE = re.compile(u'[\s]*[-]+.*({})[ ]*[-]+'.format(
     u'|'.join((
         # English
-        'wrote'
+        'wrote',
     ))), re.I)
 
 # Support polymail.io reply format
@@ -161,15 +169,15 @@ SPLITTER_PATTERNS = [
     RE_FROM_COLON_OR_DATE_COLON,
     # 02.04.2012 14:20 пользователь "bob@example.com" <
     # bob@xxx.mailgun.org> написал:
-    re.compile("(\d+/\d+/\d+|\d+\.\d+\.\d+).*@", re.S),
+    re.compile("(\d+/\d+/\d+|\d+\.\d+\.\d+).*\s\S+@\S+", re.S),
     # 2014-10-17 11:28 GMT+03:00 Bob <
     # bob@example.com>:
-    re.compile("\d{4}-\d{2}-\d{2}\s+\d{2}:\d{2}\s+GMT.*@", re.S),
+    re.compile("\d{4}-\d{2}-\d{2}\s+\d{2}:\d{2}\s+GMT.*\s\S+@\S+", re.S),
     # Thu, 26 Jun 2014 14:00:51 +0400 Bob <bob@example.com>:
     re.compile('\S{3,10}, \d\d? \S{3,10} 20\d\d,? \d\d?:\d\d(:\d\d)?'
                '( \S+){3,6}@\S+:'),
     # Sent from Samsung MobileName <address@example.com> wrote:
-    re.compile('Sent from Samsung .*@.*> wrote'),
+    re.compile('Sent from Samsung.* \S+@\S+> wrote'),
     RE_ANDROID_WROTE,
     RE_POLYMAIL
     ]
@@ -282,7 +290,7 @@ def process_marked_lines(lines, markers, return_flags=[False, -1, -1]):
     # inlined reply
     # use lookbehind assertions to find overlapping entries e.g. for 'mtmtm'
     # both 't' entries should be found
-    for inline_reply in re.finditer('(?<=m)e*((?:t+e*)+)m', markers):
+    for inline_reply in re.finditer('(?<=m)e*(t[te]*)m', markers):
         # long links could break sequence of quotation lines but they shouldn't
         # be considered an inline reply
         links = (
@@ -426,6 +434,9 @@ def _extract_from_html(msg_body):
     Extract not quoted message from provided html message body
     using tags and plain text algorithm.
 
+    Cut out first some encoding html tags such as xml and doctype
+    for avoiding conflict with unicode decoding
+
     Cut out the 'blockquote', 'gmail_quote' tags.
     Cut Microsoft quotations.
 
@@ -441,6 +452,9 @@ def _extract_from_html(msg_body):
         return msg_body
 
     msg_body = msg_body.replace(b'\r\n', b'\n')
+
+    msg_body = re.sub(r"\<\?xml.+\?\>|\<\!DOCTYPE.+]\>", "", msg_body)
+
     html_tree = html_document_fromstring(msg_body)
 
     if html_tree is None:

talon/signature/bruteforce.py

@@ -1,15 +1,15 @@
 from __future__ import absolute_import
+
 import logging
 
 import regex as re
 
-from talon.utils import get_delimiter
 from talon.signature.constants import (SIGNATURE_MAX_LINES,
                                        TOO_LONG_SIGNATURE_LINE)
+from talon.utils import get_delimiter
 
 log = logging.getLogger(__name__)
 
-
 # regex to fetch signature based on common signature words
 RE_SIGNATURE = re.compile(r'''
                (
@@ -28,7 +28,6 @@ RE_SIGNATURE = re.compile(r'''
                )
                ''', re.I | re.X | re.M | re.S)
 
-
 # signatures appended by phone email clients
 RE_PHONE_SIGNATURE = re.compile(r'''
                (
@@ -45,7 +44,6 @@ RE_PHONE_SIGNATURE = re.compile(r'''
                )
                ''', re.I | re.X | re.M | re.S)
 
-
 # see _mark_candidate_indexes() for details
 # c - could be signature line
 # d - line starts with dashes (could be signature or list item)
@@ -112,7 +110,7 @@ def extract_signature(msg_body):
 
             return (stripped_body.strip(),
                     signature.strip())
-    except Exception as e:
+    except Exception:
         log.exception('ERROR extracting signature')
         return (msg_body, None)
 
@@ -163,7 +161,7 @@ def _mark_candidate_indexes(lines, candidate):
     'cdc'
     """
     # at first consider everything to be potential signature lines
-    markers = bytearray('c'*len(candidate))
+    markers = list('c' * len(candidate))
 
     # mark lines starting from bottom up
     for i, line_idx in reversed(list(enumerate(candidate))):
@@ -174,7 +172,7 @@ def _mark_candidate_indexes(lines, candidate):
             if line.startswith('-') and line.strip("-"):
                 markers[i] = 'd'
 
-    return markers
+    return "".join(markers)
 
 
 def _process_marked_candidate_indexes(candidate, markers):

talon/signature/extraction.py

@@ -1,16 +1,15 @@
 # -*- coding: utf-8 -*-
 
 from __future__ import absolute_import
+
 import logging
 
-import regex as re
 import numpy
-
+import regex as re
-from talon.signature.learning.featurespace import features, build_pattern
-from talon.utils import get_delimiter
 from talon.signature.bruteforce import get_signature_candidate
+from talon.signature.learning.featurespace import features, build_pattern
 from talon.signature.learning.helpers import has_signature
-
+from talon.utils import get_delimiter
 
 log = logging.getLogger(__name__)
 
@@ -33,7 +32,7 @@ RE_REVERSE_SIGNATURE = re.compile(r'''
 
 def is_signature_line(line, sender, classifier):
     '''Checks if the line belongs to signature. Returns True or False.'''
-    data = numpy.array(build_pattern(line, features(sender)))
+    data = numpy.array(build_pattern(line, features(sender))).reshape(1, -1)
     return classifier.predict(data) > 0
 
 
@@ -58,7 +57,7 @@ def extract(body, sender):
                 text = delimiter.join(text)
                 if text.strip():
                     return (text, delimiter.join(signature))
-    except Exception:
+    except Exception as e:
         log.exception('ERROR when extracting signature with classifiers')
 
     return (body, None)
@@ -81,7 +80,7 @@ def _mark_lines(lines, sender):
     candidate = get_signature_candidate(lines)
 
     # at first consider everything to be text no signature
-    markers = bytearray('t'*len(lines))
+    markers = list('t' * len(lines))
 
     # mark lines starting from bottom up
     # mark only lines that belong to candidate
@@ -96,7 +95,7 @@ def _mark_lines(lines, sender):
         elif is_signature_line(line, sender, EXTRACTOR):
             markers[j] = 's'
 
-    return markers
+    return "".join(markers)
 
 
 def _process_marked_lines(lines, markers):
@@ -111,3 +110,4 @@ def _process_marked_lines(lines, markers):
         return (lines[:-signature.end()], lines[-signature.end():])
 
     return (lines, None)
+

talon/signature/learning/classifier.py

@@ -6,9 +6,10 @@ body belongs to the signature.
 """
 
 from __future__ import absolute_import
+
 from numpy import genfromtxt
-from sklearn.svm import LinearSVC
 from sklearn.externals import joblib
+from sklearn.svm import LinearSVC
 
 
 def init():
@@ -29,4 +30,40 @@ def train(classifier, train_data_filename, save_classifier_filename=None):
 
 def load(saved_classifier_filename, train_data_filename):
     """Loads saved classifier. """
+    try:
         return joblib.load(saved_classifier_filename)
+    except Exception:
+        import sys
+        if sys.version_info > (3, 0):
+            return load_compat(saved_classifier_filename)
+
+        raise
+
+
+def load_compat(saved_classifier_filename):
+    import os
+    import pickle
+    import tempfile
+
+    # we need to switch to the data path to properly load the related _xx.npy files
+    cwd = os.getcwd()
+    os.chdir(os.path.dirname(saved_classifier_filename))
+
+    # convert encoding using pick.load and write to temp file which we'll tell joblib to use
+    pickle_file = open(saved_classifier_filename, 'rb')
+    classifier = pickle.load(pickle_file, encoding='latin1')
+
+    try:
+        # save our conversion if permissions allow
+        joblib.dump(classifier, saved_classifier_filename)
+    except Exception:
+        # can't write to classifier, use a temp file
+        tmp = tempfile.SpooledTemporaryFile()
+        joblib.dump(classifier, tmp)
+        saved_classifier_filename = tmp
+
+    # important, use joblib.load before switching back to original cwd
+    jb_classifier = joblib.load(saved_classifier_filename)
+    os.chdir(cwd)
+
+    return jb_classifier

talon/signature/learning/dataset.py

@@ -17,13 +17,14 @@ suffix which should be `_sender`.
 """
 
 from __future__ import absolute_import
+
 import os
+
 import regex as re
+from six.moves import range
 
 from talon.signature.constants import SIGNATURE_MAX_LINES
 from talon.signature.learning.featurespace import build_pattern, features
-from six.moves import range
-
 
 SENDER_SUFFIX = '_sender'
 BODY_SUFFIX = '_body'
@@ -57,9 +58,14 @@ def parse_msg_sender(filename, sender_known=True):
     algorithm:
     >>> parse_msg_sender(filename, False)
     """
+    import sys
+    kwargs = {}
+    if sys.version_info > (3, 0):
+        kwargs["encoding"] = "utf8"
+
     sender, msg = None, None
     if os.path.isfile(filename) and not is_sender_filename(filename):
-        with open(filename) as f:
+        with open(filename, **kwargs) as f:
             msg = f.read()
             sender = u''
             if sender_known:

talon/utils.py

@@ -1,19 +1,18 @@
 # coding:utf-8
 
 from __future__ import absolute_import
-import logging
+
 from random import shuffle
-import chardet
+
 import cchardet
-import regex as re
+import chardet
-
-from lxml.html import html5parser
-from lxml.cssselect import CSSSelector
-
 import html5lib
+import regex as re
+import six
+from lxml.cssselect import CSSSelector
+from lxml.html import html5parser
 
 from talon.constants import RE_DELIMITER
-import six
 
 
 def safe_format(format_string, *args, **kwargs):
@@ -177,6 +176,8 @@ def html_to_text(string):
 def html_fromstring(s):
     """Parse html tree from string. Return None if the string can't be parsed.
     """
+    if isinstance(s, six.text_type):
+        s = s.encode('utf8')
     try:
         if html_too_big(s):
             return None
@@ -189,6 +190,8 @@ def html_fromstring(s):
 def html_document_fromstring(s):
     """Parse html tree from string. Return None if the string can't be parsed.
     """
+    if isinstance(s, six.text_type):
+        s = s.encode('utf8')
     try:
         if html_too_big(s):
             return None
@@ -203,7 +206,9 @@ def cssselect(expr, tree):
 
 
 def html_too_big(s):
-    return s.count('<') > _MAX_TAGS_COUNT
+    if isinstance(s, six.text_type):
+        s = s.encode('utf8')
+    return s.count(b'<') > _MAX_TAGS_COUNT
 
 
 def _contains_charset_spec(s):
@@ -248,7 +253,6 @@ def _html5lib_parser():
 _UTF8_DECLARATION = (b'<meta http-equiv="Content-Type" content="text/html;'
                      b'charset=utf-8">')
 
-
 _BLOCKTAGS = ['div', 'p', 'ul', 'li', 'h1', 'h2', 'h3']
 _HARDBREAKS = ['br', 'hr', 'tr']
 

tests/html_quotations_test.py

@@ -1,13 +1,13 @@
 # -*- coding: utf-8 -*-
 
 from __future__ import absolute_import
-from . import *
-from . fixtures import *
 
-import regex as re
+# noinspection PyUnresolvedReferences
+import re
 
 from talon import quotations, utils as u
-
+from . import *
+from .fixtures import *
 
 RE_WHITESPACE = re.compile("\s")
 RE_DOUBLE_WHITESPACE = re.compile("\s")
@@ -303,7 +303,12 @@ Reply
 
 
 def extract_reply_and_check(filename):
-    f = open(filename)
+    import sys
+    kwargs = {}
+    if sys.version_info > (3, 0):
+        kwargs["encoding"] = "utf8"
+
+    f = open(filename, **kwargs)
 
     msg_body = f.read()
     reply = quotations.extract_from_html(msg_body)

tests/signature/extraction_test.py

@@ -1,16 +1,16 @@
 # -*- coding: utf-8 -*-
 
 from __future__ import absolute_import
-from .. import *
 
 import os
 
-from talon.signature.learning import dataset
-from talon import signature
-from talon.signature import extraction as e
-from talon.signature import bruteforce
 from six.moves import range
 
+from talon.signature import bruteforce, extraction, extract
+from talon.signature import extraction as e
+from talon.signature.learning import dataset
+from .. import *
+
 
 def test_message_shorter_SIGNATURE_MAX_LINES():
     sender = "bob@foo.bar"
@@ -18,20 +18,25 @@ def test_message_shorter_SIGNATURE_MAX_LINES():
 
 Thanks in advance,
 Bob"""
-    text, extracted_signature = signature.extract(body, sender)
+    text, extracted_signature = extract(body, sender)
     eq_('\n'.join(body.splitlines()[:2]), text)
     eq_('\n'.join(body.splitlines()[-2:]), extracted_signature)
 
 
 def test_messages_longer_SIGNATURE_MAX_LINES():
+    import sys
+    kwargs = {}
+    if sys.version_info > (3, 0):
+        kwargs["encoding"] = "utf8"
+
     for filename in os.listdir(STRIPPED):
         filename = os.path.join(STRIPPED, filename)
         if not filename.endswith('_body'):
             continue
         sender, body = dataset.parse_msg_sender(filename)
-        text, extracted_signature = signature.extract(body, sender)
+        text, extracted_signature = extract(body, sender)
         extracted_signature = extracted_signature or ''
-        with open(filename[:-len('body')] + 'signature') as ms:
+        with open(filename[:-len('body')] + 'signature', **kwargs) as ms:
             msg_signature = ms.read()
             eq_(msg_signature.strip(), extracted_signature.strip())
             stripped_msg = body.strip()[:len(body.strip()) - len(msg_signature)]
@@ -47,7 +52,7 @@ Thanks in advance,
 some text which doesn't seem to be a signature at all
 Bob"""
 
-    text, extracted_signature = signature.extract(body, sender)
+    text, extracted_signature = extract(body, sender)
     eq_('\n'.join(body.splitlines()[:2]), text)
     eq_('\n'.join(body.splitlines()[-3:]), extracted_signature)
 
@@ -60,7 +65,7 @@ Thanks in advance,
 some long text here which doesn't seem to be a signature at all
 Bob"""
 
-    text, extracted_signature = signature.extract(body, sender)
+    text, extracted_signature = extract(body, sender)
     eq_('\n'.join(body.splitlines()[:-1]), text)
     eq_('Bob', extracted_signature)
 
@@ -68,13 +73,13 @@ Bob"""
 
     some *long* text here which doesn't seem to be a signature at all
     """
-    ((body, None), signature.extract(body, "david@example.com"))
+    ((body, None), extract(body, "david@example.com"))
 
 
 def test_basic():
     msg_body = 'Blah\r\n--\r\n\r\nSergey Obukhov'
     eq_(('Blah', '--\r\n\r\nSergey Obukhov'),
-        signature.extract(msg_body, 'Sergey'))
+        extract(msg_body, 'Sergey'))
 
 
 def test_capitalized():
@@ -99,7 +104,7 @@ Doe Inc
 Doe Inc
 555-531-7967"""
 
-    eq_(sig, signature.extract(msg_body, 'Doe')[1])
+    eq_(sig, extract(msg_body, 'Doe')[1])
 
 
 def test_over_2_text_lines_after_signature():
@@ -110,25 +115,25 @@ def test_over_2_text_lines_after_signature():
     2 non signature lines in the end
     It's not signature
     """
-    text, extracted_signature = signature.extract(body, "Bob")
+    text, extracted_signature = extract(body, "Bob")
     eq_(extracted_signature, None)
 
 
 def test_no_signature():
     sender, body = "bob@foo.bar", "Hello"
-    eq_((body, None), signature.extract(body, sender))
+    eq_((body, None), extract(body, sender))
 
 
 def test_handles_unicode():
     sender, body = dataset.parse_msg_sender(UNICODE_MSG)
-    text, extracted_signature = signature.extract(body, sender)
+    text, extracted_signature = extract(body, sender)
 
 
-@patch.object(signature.extraction, 'has_signature')
+@patch.object(extraction, 'has_signature')
 def test_signature_extract_crash(has_signature):
     has_signature.side_effect = Exception('Bam!')
     msg_body = u'Blah\r\n--\r\n\r\nСергей'
-    eq_((msg_body, None), signature.extract(msg_body, 'Сергей'))
+    eq_((msg_body, None), extract(msg_body, 'Сергей'))
 
 
 def test_mark_lines():

tests/text_quotations_test.py

@@ -119,6 +119,38 @@ On 11-Apr-2011, at 6:54 PM, Roman Tkachenko <romant@example.com> sent:
     eq_("Test reply", quotations.extract_from_plain(msg_body))
 
 
+def test_appointment():
+    msg_body = """Response
+
+10/19/2017 @ 9:30 am for physical therapy
+Bla
+1517 4th Avenue Ste 300
+London CA 19129, 555-421-6780
+
+John Doe, FCLS
+Mailgun Inc
+555-941-0697
+
+From: from@example.com [mailto:from@example.com]
+Sent: Wednesday, October 18, 2017 2:05 PM
+To: John Doer - SIU <jd@example.com>
+Subject: RE: Claim # 5551188-1
+
+Text"""
+
+    expected = """Response
+
+10/19/2017 @ 9:30 am for physical therapy
+Bla
+1517 4th Avenue Ste 300
+London CA 19129, 555-421-6780
+
+John Doe, FCLS
+Mailgun Inc
+555-941-0697"""
+    eq_(expected, quotations.extract_from_plain(msg_body))
+
+
 def test_line_starts_with_on():
     msg_body = """Blah-blah-blah
 On blah-blah-blah"""
@@ -401,6 +433,14 @@ Op 17-feb.-2015, om 13:18 heeft Julius Caesar <pantheon@rome.com> het volgende g
 Small batch beard laboris tempor, non listicle hella Tumblr heirloom.
 """))
 
+def test_vietnamese_from_block():
+    eq_('Hello', quotations.extract_from_plain(
+    u"""Hello
+
+Vào 14:24 8 tháng 6, 2017, Hùng Nguyễn <hungnguyen@xxx.com> đã viết:
+
+> Xin chào
+"""))
 
 def test_quotation_marker_false_positive():
     msg_body = """Visit us now for assistance...
@@ -770,3 +810,16 @@ def test_split_email():
     expected_markers = "stttttsttttetesetesmmmmmmssmmmmmmsmmmmmmmm"
     markers = quotations.split_emails(msg)
     eq_(markers, expected_markers)
+
+
+
+def test_feedback_below_left_unparsed():
+    msg_body = """Please enter your feedback below. Thank you.
+
+------------------------------------- Enter Feedback Below -------------------------------------
+
+The user experience was unparallelled. Please continue production. I'm sending payment to ensure
+that this line is intact."""
+
+    parsed = quotations.extract_from_plain(msg_body)
+    eq_(msg_body, parsed.decode('utf8'))

tests/utils_test.py

@@ -1,12 +1,12 @@
 # coding:utf-8
 
 from __future__ import absolute_import
-from . import *
 
-from talon import utils as u
 import cchardet
 import six
-from lxml import html
+
+from talon import utils as u
+from . import *
 
 
 def test_get_delimiter():
@@ -115,15 +115,16 @@ font: 13px 'Lucida Grande', Arial, sans-serif;
 
 
 def test_comment_no_parent():
-    s = "<!-- COMMENT 1 --> no comment"
+    s = b'<!-- COMMENT 1 --> no comment'
     d = u.html_document_fromstring(s)
-    eq_("no comment", u.html_tree_to_text(d))
+    eq_(b"no comment", u.html_tree_to_text(d))
 
 
 @patch.object(u.html5parser, 'fromstring', Mock(side_effect=Exception()))
 def test_html_fromstring_exception():
     eq_(None, u.html_fromstring("<html></html>"))
 
+
 @patch.object(u, 'html_too_big', Mock())
 @patch.object(u.html5parser, 'fromstring')
 def test_html_fromstring_too_big(fromstring):
@@ -158,5 +159,5 @@ def test_html_too_big():
 
 @patch.object(u, '_MAX_TAGS_COUNT', 3)
 def test_html_to_text():
-    eq_("Hello", u.html_to_text("<div>Hello</div>"))
+    eq_(b"Hello", u.html_to_text("<div>Hello</div>"))
     eq_(None, u.html_to_text("<div><span>Hi</span></div>"))