Custom language detection

dateparser allows to customize the language detection behavior by using the detect_languages_function parameter. It currently supports two language detection libraries out of the box: fastText and langdetect, and allows you to implement your own custom language detection.


For short strings the language detection could fail, so it’s highly recommended to use detect_languages_function along with DEFAULT_LANGUAGES.

Built-in implementations


Language detection with fastText.

Import the fastText wrapper and pass it as detect_languages_function parameter. Example:

>>> from dateparser.custom_language_detection.fasttext import detect_languages
>>> dateparser.parse('12/12/12', detect_languages_function=detect_languages)

The fastText integration currently supports the large and the small models. Find more information about fasttext models. You can download your model of choice using dateparser-download.

Downloading small model:

>>> dateparser-download --fasttext small

Downloading large model:

>>> dateparser-download --fasttext large

Deleting all cached models:

>>> dateparser-download --clear_cache


If no model has been downloaded, the fastText wrapper downloads and uses the small model by default.


Language detection with langdetect.

Import the langdetect wrapper and pass it as detect_languages_function parameter. Example:

>>> from dateparser.custom_language_detection.langdetect import detect_languages
>>> dateparser.parse('12/12/12', detect_languages_function=detect_languages)


From some tests we did, we recommend to use fastText for faster and more accurate results.

Custom implementation

dateparser allows the integration of any library to detect languages by wrapping that library in a function that accepts 2 parameters, text and confidence_threshold, and returns a list of the detected language codes in ISO 639 standards.

Wrapper for boilerplate for implementing custom language detections:

def detect_languages(text, confidence_threshold):
    Takes 2 parameters, `text` and `confidence_threshold`, and returns
    a list of `languages codes`.

    * `text` is the input string whose language needs to be detected.

    * `confidence_threshold` is a number between 0 and 1 that indicates the
    minimum confidence required for language matches.

    For language detection libraries that, for each language, indicate how
    confident they are that the language matches the input text, you should
    filter out languages with a confidence lower than this value (adjusted,
    if needed, to the confidence range of the target library).

    This value comes from the dateparser setting

    The result must be a list of languages codes (strings).
    # here you can apply your own logic
    return language_codes