Equilid

Socially-Eqiutable Language Identification


Keywords
langid, language, identification, indigenous, codeswitching, code-switching
License
Apache-2.0
Install
pip install Equilid==1.0.1

Documentation

Equilid: Socially-Equitable Language Identification

Equilid is a general purpose language identification library and command line utility built with the following goals:

  1. Identify a broad coverage of languages
  2. Recognize langugage in social media, with a particular emphasis on short text
  3. Recognize dialectic speech from a language's speakers
  4. Identify code-switched text in any language pairing at least at the phrase level
  5. Provide whole message and per-wor

Equilid currently comes pre-trained on 70 languages (ISO 639-3 codes given):

amh ara ben bos bul cat ces cym dan deu div ell
eng est eus fas fin fra guj hat heb hin hrv hun
hye ind isl ita jpn kan kat khm kor lao lat lava
lit mal mar mkd mon msa mya nep nld nor ori pan
pol por pus ron rus sin slk slv snd spa srp swe
tam tel tgl tha tur uig ukr urd vie zho 

Why use Equilid?

In global settings like Twitter, this text is written by authors from diverse linguistic backgrounds, who may communicate with regional dialects or even include parallel translations in the same message to address different audiences. Such dialectal variation is frequent in all languages and even macro-dialects such as American and British English are composed of local dialects that vary across city and socioeconomic development level. Yet current systems for broad-coverage LID—trained on dozens of languages—have largely leveraged Europeancentric corpora and not taken into account demographic and dialectal variation. As a result, these systems systematically misclassify texts from populations with millions of speakers whose local speech differs from the majority dialects. Equilid aims to be a socially equitable language identification system that operates at high precision in a massively multilingual, broad-coverage domains and that supports populations speaking underrepresented dialects, multilingual messages, and other linguistic varieties.

Short summary: If you are working with text from a global environment or especially if you are working with text from a country that has dialectic language, Equilid will provide superior language identification accuracy and help you find messages from underrepresented populations.

Installation

Under the hood, Equilid uses a neural seq2seq model. It depends on three libraries:

  • tensorflow 0.11.0
  • numpy
  • regex

Equilid may work with later versions of tensorflow but this hasn't been tested (yet).

Equilid can be installed via pip pip install equilid. However, this installs only the software and not the trained model. The trained model downloaded here [http://cs.stanford.edu/~jurgens/data/70lang.tar.gz] (559MB unarchived).

To install a trained model, create a directory models in the base Equilid directory and unpack the model's archive file into it.

Usage


Equilid can be used as both a stand-alone file and as a python library

equilid.py [options]

Options:
  -h, --help            show this help message and exit
  --predict             launch Equilid in per-token prediction mode
  --predict_file        reads unlabled instances from this file 
                        (if unspecified, STDIN is used)
  --predict_output_file writes per-token predictions to this file
                        (if unspecified, STDOUT is used)

You can also use Equilid as a Python library:

# python
Python 2.7.12 |Anaconda custom (64-bit)| (default, Jul  2 2016, 17:42:40) 
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux2
>>> import equilid
>>> equilid.classify("This is a test.")
['eng', 'eng', 'eng', 'eng']
>>> equilid.get_langs("Esto es una prueba.")
set(['spa'])
>>> Equilid.get_langs("This is a test.  Esto es una prueba.")
set(['spa', 'eng'])

What are my alternatives?

Model details

Equilid's training data was drawn from multiple sources:

Changelog

  • v1.0.1:

    • Fixed unicode issue
    • Added model download code (untested)
  • v1.0:

    • Initial release