quantulum3

Python library for information extraction of quantities, measurements and their units from unstructured text. It is able to disambiguate between similar looking units based on their k-nearest neighbours in their GloVe vector representation and their Wikipedia page.

This is the Python 3 compatible fork of recastrodiaz' fork of grhawks' fork of the original by Marco Lagi. The compatibility with the newest version of sklearn is based on the fork of sohrabtowfighi.

User Guide

Installation

pip install quantulum3

To install dependencies for using or training the disambiguation classifier, use

pip install quantulum3[classifier]

The disambiguation classifier is used when the parser find two or more units that are a match for the text.

Usage

>>> from quantulum3 import parser
>>> quants = parser.parse('I want 2 liters of wine')
>>> quants
[Quantity(2, 'litre')]

The Quantity class stores the surface of the original text it was extracted from, as well as the (start, end) positions of the match:

>>> quants[0].surface
u'2 liters'
>>> quants[0].span
(7, 15)

The value attribute provides the parsed numeric value and the unit.name attribute provides the name of the parsed unit:

>>> quants[0].value
2.0
>>> quants[0].unit.name
'litre'

An inline parser that embeds the parsed quantities in the text is also available (especially useful for debugging):

>>> print parser.inline_parse('I want 2 liters of wine')
I want 2 liters {Quantity(2, "litre")} of wine

As the parser is also able to parse dimensionless numbers, this library can also be used for simple number extraction.

>>> print parser.parse('I want two')
[Quantity(2, 'dimensionless')]

Units and entities

All units (e.g. litre) and the entities they are associated to (e.g. volume) are reconciled against WikiPedia:

>>> quants[0].unit
Unit(name="litre", entity=Entity("volume"), uri=https://en.wikipedia.org/wiki/Litre)

>>> quants[0].unit.entity
Entity(name="volume", uri=https://en.wikipedia.org/wiki/Volume)

This library includes more than 290 units and 75 entities. It also parses spelled-out numbers, ranges and uncertainties:

>>> parser.parse('I want a gallon of beer')
[Quantity(1, 'gallon')]

>>> parser.parse('The LHC smashes proton beams at 12.8–13.0 TeV')
[Quantity(12.8, "teraelectronvolt"), Quantity(13, "teraelectronvolt")]

>>> quant = parser.parse('The LHC smashes proton beams at 12.9±0.1 TeV')
>>> quant[0].uncertainty
0.1

Non-standard units usually don't have a WikiPedia page. The parser will still try to guess their underlying entity based on their dimensionality:

>>> parser.parse('Sound travels at 0.34 km/s')[0].unit
Unit(name="kilometre per second", entity=Entity("speed"), uri=None)

Disambiguation

If the parser detects an ambiguity, a classifier based on the WikiPedia pages of the ambiguous units or entities tries to guess the right one:

>>> parser.parse('I spent 20 pounds on this!')
[Quantity(20, "pound sterling")]

>>> parser.parse('It weighs no more than 20 pounds')
[Quantity(20, "pound-mass")]

or:

>>> text = 'The average density of the Earth is about 5.5x10-3 kg/cm³'
>>> parser.parse(text)[0].unit.entity
Entity(name="density", uri=https://en.wikipedia.org/wiki/Density)

>>> text = 'The amount of O₂ is 2.98e-4 kg per liter of atmosphere'
>>> parser.parse(text)[0].unit.entity
Entity(name="concentration", uri=https://en.wikipedia.org/wiki/Concentration)

In addition to that, the classifier is trained on the most similar words to all of the units surfaces, according to their distance in GloVe vector representation.

Spoken version

Quantulum classes include methods to convert them to a speakable unit.

>>> parser.parse("Gimme 10e9 GW now!")[0].to_spoken()
ten billion gigawatts
>>> parser.inline_parse_and_expand("Gimme $1e10 now and also 1 TW and 0.5 J!")
Gimme ten billion dollars now and also one terawatt and zero point five joules!

Manipulation

While quantities cannot be manipulated within this library, there are many great options out there:

• pint
• natu
• quantities

Extension

Training the classifier

If you want to train the classifier yourself, you will need the dependencies for the classifier (see installation).

Use quantulum3-training on the command line, the script quantulum3/scripts/train.py or the method train_classifier in quantulum3.classifier to train the classifier.

quantulum3-training --lang <language> --data <path/to/training/file.json> --output <path/to/output/file.joblib>

You can pass multiple training files in to the training command. The output is in joblib format.

To use your custom model, pass the path to the trained model file to the parser:

parser = Parser.parse(<text>, classifier_path="path/to/model.joblib")

Example training files can be found in quantulum3/_lang/<language>/train.

If you want to create a new or different similars.json, install pymagnitude.

For the extraction of nearest neighbours from a vector word representation file, use scripts/extract_vere.py. It automatically extracts the k nearest neighbours in vector space of the vector representation for each of the possible surfaces of the ambiguous units. The resulting neighbours are stored in quantulum3/similars.json and automatically included for training.

The file provided should be in .magnitude format as other formats are first converted to a .magnitude file on-the-run. Check out pre-formatted Magnitude formatted word-embeddings and Magnitude for more information.

Additional units

It is possible to add additional entities and units to be parsed by quantulum. These will be added to the default units and entities. See below code for an example invocation:

>>> from quantulum3.load import add_custom_unit, remove_custom_unit
>>> add_custom_unit(name="schlurp", surfaces=["slp"], entity="dimensionless")
>>> parser.parse("This extremely sharp tool is precise up to 0.5 slp")
[Quantity(0.5, "Unit(name="schlurp", entity=Entity("dimensionless"), uri=None)")]

The keyword arguments to the function add_custom_unit are directly translated to the properties of the unit to be created.

Custom Units and Entities

It is possible to load a completely custom set of units and entities. This can be done by passing a list of file paths to the load_custom_units and load_custom_entities functions. Loading custom untis and entities will replace the default units and entities that are normally loaded.

The recomended way to load quantities is via a context manager:

>>> from quantulum3 import load, parser
>>> with load.CustomQuantities(["path/to/units.json"], ["path/to/entities.json"]):
>>>     parser.parse("This extremely sharp tool is precise up to 0.5 slp")

[Quantity(0.5, "Unit(name="schlurp", entity=Entity("dimensionless"), uri=None)")]

>>> # default units and entities are loaded again

But it is also possible to load custom units and entities manually:

>>> from quantulum3 import load, parser

>>> load.load_custom_units(["path/to/units.json"])
>>> load.load_custom_entities(["path/to/entities.json"])
>>> parser.parse("This extremely sharp tool is precise up to 0.5 slp")

[Quantity(0.5, "Unit(name="schlurp", entity=Entity("dimensionless"), uri=None)")]

>>> # remove custom units and entities and load default units and entities
>>> load.reset_quantities()

See the Developer Guide below for more information about the format of units and entities files.

Developer Guide

Adding Units and Entities

See units.json for the complete list of units and entities.json for the complete list of entities. The criteria for adding units have been:

• the unit has (or is redirected to) a WikiPedia page
• the unit is in common use (e.g. not the premetric Swedish units of measurement).

It's easy to extend these two files to the units/entities of interest. Here is an example of an entry in entities.json:

"speed": {
    "dimensions": [{"base": "length", "power": 1}, {"base": "time", "power": -1}],
    "URI": "https://en.wikipedia.org/wiki/Speed"
}

• The name of an entity is its key. Names are required to be unique.
• URI is the name of the wikipedia page of the entity. (i.e. https://en.wikipedia.org/wiki/Speed => Speed)
• dimensions is the dimensionality, a list of dictionaries each having a base (the name of another entity) and a power (an integer, can be negative).

Here is an example of an entry in units.json:

"metre per second": {
    "surfaces": ["metre per second", "meter per second"],
    "entity": "speed",
    "URI": "Metre_per_second",
    "dimensions": [{"base": "metre", "power": 1}, {"base": "second", "power": -1}],
    "symbols": ["mps"]
},
"year": {
    "surfaces": [ "year", "annum" ],
    "entity": "time",
    "URI": "Year",
    "dimensions": [],
    "symbols": [ "a", "y", "yr" ],
    "prefixes": [ "k", "M", "G", "T", "P", "E" ]
}

• The name of a unit is its key. Names are required to be unique.
• URI follows the same scheme as in the entities.json
• surfaces is a list of strings that refer to that unit. The library takes care of plurals, no need to specify them.
• entity is the name of an entity in entities.json
• dimensions follows the same schema as in entities.json, but the base is the name of another unit, not of another entity.
• symbols is a list of possible symbols and abbreviations for that unit.
• prefixes is an optional list. It can contain Metric and Binary prefixes and automatically generates according units. If you want to add specifics (like different surfaces) you need to create an entry for that prefixes version on its own.

All fields are case sensitive.

Contributing

dev build:

If you'd like to contribute follow these steps:

1. Clone a fork of this project into your workspace
2. Run pip install -e . at the root of your development folder.
3. pip install pipenv and pipenv shell
4. Inside the project folder run pipenv install --dev
5. Make your changes
6. Run scripts/format.sh and scripts/build.py from the package root directory.
7. Test your changes with python3 setup.py test (Optional, will be done automatically after pushing)
8. Create a Pull Request when having commited and pushed your changes

Language support

There is a branch for language support, namely language_support. From inspecting the README file in the _lang subdirectory and the functions and values given in the new _lang.en_US submodule, one should be able to create own language submodules. The new language modules should automatically be invoked and be available, both through the lang= keyword argument in the parser functions as well as in the automatic unittests.

No changes outside the own language submodule folder (i.e. _lang.de_DE) should be necessary. If there are problems implementing a new language, don't hesitate to open an issue.