Skip to main content

Parse natural language time expressions in python

Project description

ctparse - Parse natural language time expressions in python

Travis Coverage PyPi Updates Documentation Status https://img.shields.io/badge/code%20style-black-000000.svg

This code is in early alpha stage. There can and will be potentially breaking changes right on the ``master`` branch

Background

The package ctparse is a pure python package to parse time expressions from natural language (i.e. strings). In many ways it builds on similar concepts as Facebook’s duckling package (https://github.com/facebook/duckling). However, for the time being it only targets times and only German and English text.

In principle ctparse can be used to detect time expressions in a text, however its main use case is the semantic interpretation of such expressions. Detecting time expressions in the first place can - to our experience - be done more efficiently (and precisely) using e.g. CRFs or other models targeted at this specific task.

ctparse is designed with the use case in mind where interpretation of time expressions is done under the following assumptions:

  • All expressions are relative to some pre-defined reference times
  • Unless explicitly specified in the time expression, valid resolutions are in the future relative to the reference time (i.e. 12.5. will be the next 12th of May, but 12.5.2012 should correctly resolve to the 12th of May 2012).
  • If in doubt, resolutions in the near future are more likely than resolutions in the far future (not implemented yet, but any resolution more than i.e. 3 month in the future is extremely unlikely).

The specific comtravo use-case is resolving time expressions in booking requests which almost always refer to some point in time within the next 4-8 weeks.

ctparse currently is language agnostic and supports German and English expressions. This might get an extension in the future. The main reason is that in real world communication more often than not people write in one language (their business language) but use constructs to express times that are based on their mother tongue and/or what they believe to be the way to express dates in the target language. This leads to text in German with English time expressions and vice-versa. Using a language detection upfront on the complete original text is for obvious no solution - rather it would make the problem worse.

Example

from ctparse import ctparse
from datetime import datetime

# Set reference time
ts = datetime(2018, 3, 12, 14, 30)
ctparse('May 5th 2:30 in the afternoon', ts=ts)

This should return a Time object represented as Time[0-29]{2018-05-05 14:30 (X/X)}, indicating that characters 0-29 were used in the resolution, that the resolved date time is the 5th of May 2018 at 14:30 and that this resolution is neither based on a day of week (first X) nor a part of day (second X).

Latent time

Normally, ctparse will anchor time expressions to the reference time. For example, when parsing the time expression 8:00 pm, ctparse will resolve the expression to 8 pm after the reference time as follows

parse = ctparse("8:00 pm", ts=datetime(2020, 1, 1, 7, 0), latent_time=True) # default
# parse.resolution -> Time(2020, 1, 1, 20, 00)

This behavior can be customized using the option latent_time=False, which will return a time resolution not anchored to a particular date

parse = ctparse("8:00 pm", ts=datetime(2020, 1, 1, 7, 0), latent_time=False)
# parse.resolution -> Time(None, None, None, 20, 00)

Implementation

ctparse - as duckling - is a mixture of a rule and regular expression based system + some probabilistic modeling. In this sense it resembles a PCFG.

Rules

At the core ctparse is a collection of production rules over sequences of regular expressions and (intermediate) productions.

Productions are either of type Time, Interval or Duration and can have certain predicates (e.g. whether a Time is a part of day like 'afternoon').

A typical rule than looks like this:

@rule(predicate('isDate'), dimension(Interval))

I.e. this rule is applicable when the intermediate production resulted in something that has a date, followed by something that is in interval (like e.g. in 'May 5th 9-10').

The actual production is a python function with the following signature:

@rule(predicate('isDate'), dimension(Interval))
def ruleDateInterval(ts, d, i):
  """
  param ts: datetime - the current refenrence time
  d: Time - a time that contains at least a full date
  i: Interval - some Interval
  """
  if not (i.t_from.isTOD and i.t_to.isTOD):
    return None
  return Interval(
    t_from=Time(year=d.year, month=d.month, day=d.day,
                hour=i.t_from.hour, minute=i.t_from.minute),
    t_to=Time(year=d.year, month=d.month, day=d.day,
              hour=i.t_to.hour, minute=i.t_to.minute))

This production will return a new interval at the date of predicate('isDate') spanning the time coded in dimension(Interval). If the latter does code for something else than a time of day (TOD), no production is returned, e.g. the rule matched but failed.

Technical Background

Some observations on the problem:

  • Each rule is a combination of regular expressions and productions.
  • Consequently, each production must originate in a sequence of regular expressions that must have matched (parts of) the text.
  • Hence, only subsequence of all regular expressions in all rules can lead to a successful production.

To this end the algorithm proceeds as follows:

  1. Input a string and a reference time
  2. Find all matches of all regular expressions from all rules in the input strings. Each regular expression is assigned an identifier.
  3. Find all distinct sequences of these matches where two matches do not overlap nor have a gap inbetween
  4. To each such subsequence apply all rules at all possible positions until no further rules can be applied - in which case one solution is produced

Obviously, not all sequences of matching expressions and not all sequences of rules applied on top lead to meaningful results. Here the PCFG kicks in:

  • Based on example data (corpus.py) a model is calibrated to predict how likely a production is to lead to a/the correct result. Instead of doing a breadth first search, the most promising productions are applied first.
  • Resolutions are produced until there are no more resolutions or a timeout is hit.
  • Based on the same model from all resolutions the highest scoring is returned.

Credits

This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.

History

0.3.0 (2021-02-01)

  • Removed latent rules regarding times (latent rules regarding dates are still present)
  • Added latent_time option to customize the new behavior, defauld behavior is backwards-compatible

0.2.1 (2020-05-27)

  • Update development dependencies
  • Add flake8-bugbear and fixed issues

0.2.0 (2020-04-23)

  • Implemented new type Duration, to handle lengths of time
  • Adapted the dataset to include Duration
  • Implemented basic rule to merge Duration, Time and Interval in simple cases.
  • Created a make target to train the model make train

0.1.0 (2020-03-20)

  • Major refactor of code underlying predictive model
  • Based on a contribution from @bharathi-srini: replace naive bayes from sklearn by own implementation
  • Thus remove dependencies on numpy, scipy, scikit-learn
  • Predictions are much faster: 97/s in the old vs. 239/s in the new code base
  • Performance identical
  • Deprecate support for python 3.5, add 3.8
  • Add more strict type checking rules (mypy.ini)
  • Force black code formatting, make this a linter step, “black” all code

0.0.47 (2020-02-28)

  • Allow overlapping matches of regular expression when generating inital stack of “tokens”

0.0.46 (2020-02-26)

  • Implemented heuristics to detect (albeit imperfectly) military times

0.0.44 (2019-11-05)

  • Released time corpus
  • Implemented training model using ctparse corpus

0.0.43 (2019-11-01)

  • Added slash as a general separator
  • Added ruleTODTOD (to support expression like afternoon/evening)

0.0.42 (2019-10-30)

  • Removed nb module
  • Fix for two digit years
  • Freshly retrained model binary file

0.0.41 (2019-10-29)

  • Fix run_corpus refactoring bug
  • Implemented retraining utilities

0.0.40 (2019-10-25)

  • update develop dependencies
  • remove unused Protocol import from typing_extensions

0.0.39 (2019-10-24)

  • split ctparse file into several different modules
  • added types to public interface
  • introduced the Scorer abstraction to implement richer scoring strategies

0.0.38 (2018-11-05)

  • Added python 3.7 to supported versions (fix on travis available)

0.0.8 (2018-06-07)

  • First release on PyPI.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for ctparse, version 0.3.0
Filename, size File type Python version Upload date Hashes
Filename, size ctparse-0.3.0-py2.py3-none-any.whl (114.2 kB) File type Wheel Python version py2.py3 Upload date Hashes View
Filename, size ctparse-0.3.0.tar.gz (75.9 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page