Natural Language Expressions for Python

NatEx: Regular Expressions turbo-charged with notations for part-of-speech and dependency tree tags

In a Nutshell

from natex import natex

sentence = natex('Sloths eat steak in New York')

# check if string begins with noun:
sentence.match(r'@NOUN')
# returns <natex.Match object; span=(0, 6), match='Sloths'>

# find first occurrence of an adposition followed by a proper noun
sentence.search(r'@ADP <@PROPN>')  	
# returns <natex.Match object; span=(17, 28), match='in New York'>

# find all occurrences of nouns or proper nouns
sentence.findall(r'@(NOUN|PROPN)') 	
# returns ['Sloths', 'steak', 'New York']

# find all occurrences of nouns or proper nouns starting with an s (irregardless of casing)
sentence.findall(r's[^@]+@(NOUN|PROPN)', natex.I)
# returns ['Sloths', 'steak']

Goals & Design

The goal of NatEx is quick and simple parsing of tokens using their literal representation, part-of-speech, and dependency tree tags. Think of it as an extension of regular expressions for natural language processing. The generated part-of-speech and dependency tree tags are provided by stanza and merged into a string that can be searched through.

Why not Tregex, Semgrex, or Tsurgeon?

NatEx was designed primarily with simplicity in mind. Libraries like Tregex, Semgrex, or Tsurgeon may be able to match more complex patterns, but they have a steep learning curve and the patterns are hard to read. Plus NatEx is written for Python. It wraps the built-in re package with an abstraction layer and thus provides almost the same performance as any normal regex.

Examples

You can use it for simple tagging (NLU):

from natex import natex

sentence = natex('book flights from Munich to Chicago')
origin_location, destination_location = sentence.findall('<@PROPN>')
# origin_location ='Munich', destination_location = 'Chicago'

sentence = natex('I am being called Dan Borufka')
firstname, lastname = sentence.findall('<@PROPN>')
# firstname = 'Dan', lastname = 'Borufka'

sentence = natex('I need to go to Italy')
clause = sentence.search('<@ADP> <@PROPN>').match
# clause = 'to Italy'
destination = clause.split(' ')[1]

Or for simple response template generation (NLG):

from natex import natex

sentence = natex('Eat my shorts')

# look for token with imperative form
is_command = sentence.match(r'<!>')

if is_command:
	action_verb = sentence.search(r'<@VERB!>').lower()
	action_recipient = sentence.search(r'<#OBJ>')
	response = f'I will do my best to {action_verb} {action_recipient}!'

	# will contain 'I will do my best to eat shorts!'

Even more (random) examples:

from natex import natex

sentence = natex('Sloths eat steak in New York')

# find first occurrence of character sequence "ea" in nouns only
sentence.search(r'ea@NOUN')			
# returns <natex.Match object; span=(11, 16), match='steak'>

# find first occurrence of character sequence "ea"
sentence.search(r'ea')
# returns <natex.Match object; span=(7, 9), match='ea'>

# find all occurrences of nouns or proper nouns starting with a lowercase s
sentence.findall(r's[^@]+@(NOUN|PROPN)') 
# returns ['steak']

sentence = natex('Ein Hund isst keinen Gurkensalat in New York.', 'de')

# replace the nominal subject with the literal 'Affe'
sentence.sub(r'#NSUBJ', 'Affe')
# returns 'Ein Affe isst keinen Gurkensalat in New York.'

Check out test.py for some more examples!

Installation

Run:

pip install natex

By default, NatEx only installs the English models for stanza. Use the following command to download a model for another language:

python -m natex setup <language_code>`

e.g. for French use:

python -m natex setup fr`

Visit https://github.com/secretsauceai/natex-py for a full list of supported language codes.

Usage

NatEx provides the same API as the re package, but adds the following special characters:

If you combine features (e.g. seeking by part of speech and dependency tree simultaneously) make sure you provide them in the order of the table above.

✔ This will work:

natex('There goes a test sentence').findall(r'<@NOUN#OBJ>')

✘ But this won't:

natex('There goes a test sentence').findall(r'<#OBJ@NOUN>')

Calling the natex() function returns a NatEx instance. See API for a more detailed description. Just as the re.Match returning methods provided by Python's built-in re package, NatEx' equivalents will return a natex.Match object containing a span and a match property referring to position and substring of the sentence respectively.

Configuration

You can set the processing language of NatEx using the second parameter language_code (defaults to 'en'). It accepts a two-letter language-code, supporting all languages officially supported by stanza.

sentence = natex('Das Faultier isst keinen Gurkensalat', 'de')

When you run NatEx for the first time, it will check for the existence of the corresponding language models and download them if necessary. All subsequent calls to natex() will exclude that step.

API

The API is derived from Python's built-in re package:

NatEx

.match(pattern, flags)

Checks (from the beginning of the string) whether the sentence matches a pattern and returns a natex.Match object or None otherwise.

.search(pattern, flags) Returns a natex.Match object describing the first substring matching pattern.

.findall(pattern, flags) Returns all found strings matching pattern as a list.

.split(pattern, flags) Splits the sentence by all occurrences of the found pattern and returns a list of strings.

.sub(pattern, replacement, flags) Replaces all occurrences of the found pattern by replacement and returns the changed string.

Testing

You can use pytest in your terminal (simply type in pytest) to run the unit tests shipped with this package. Install it by running pip install pytest in your terminal.