multilingual-rouge 0.0.1

A fork of Multilingual ROUGE (a fork of XL-Sum)

Multilingual ROUGE Scoring

This is a port of XL-Sum repository as a simple utility to utilize the multilingual RougeL utility for MIRAGE-Bench, with the following changes:

1. Added multilingual_rouge/bengali_stemmer from bengali-stemmer repository as it was not available in PyPI.

I do not own the codebase so any questions must be redirected to XL-Sum repository.

Installation

To install this library, please use the following:

python -m unidic download # for japanese segmentation
pip install -U multilingual-rouge

Alternatively to install from source:

python -m unidic download # for japanese segmentation
pip install -e .

Overview

ROUGE is the de facto evaluation metric used for text summarization. However, it was designed specifically for evaluating English texts. Due to the nature of the metric, scores are heavily dependent on text tokenization / stemming / unnecessary character removal, etc. This repo tries to address these issues by adding the following main features using an adaptation of rouge-score: Google's rouge implementation.

• Enables multilingual ROUGE scoring by making use of popular word segmentation / stemming algorithms for various languages.
• Removes only punctuation characters according to unicode data tables as part of text normalization. This enables basic rouge scoring even with the absence of a segmenter / stemmer for any language.
• Provides an easy to use interface for using custom tokenization / stemming implementations.

Supported language names for stemming

bengali, hindi, turkish, arabic, danish, dutch, english, finnish, french, german, hungarian, italian, norwegian, portuguese, romanian, russian, spanish, swedish

Supported language names for word segmentation

chinese, thai, japanese, burmese

Setup

pip3 install -r requirements.txt
python3 -m unidic download # for japanese segmentation
pip3 install --upgrade ./

Example Usage

Using CLI

python -m rouge_score.rouge \
    --target_filepattern=*.targets \
    --prediction_filepattern=*.decodes \
    --output_filename=scores.csv \
    --use_stemmer=true \ # optional
    --lang="bengali" # optional

Using python

• Default usage

from multilingual_rouge import rouge_scorer

scorer = rouge_scorer.RougeScorer(['rouge1', 'rougeL'], use_stemmer=True)
scores = scorer.score('The quick brown fox jumps over the lazy dog',
                      'The quick brown dog jumps on the log.')

• With provided language

from multilingual_rouge import rouge_scorer

scorer = rouge_scorer.RougeScorer(['rouge1', 'rougeL'], use_stemmer=True, lang="bengali")
scores = scorer.score('তোমার সাথে দেখা হয়ে ভালো লাগলো।',
                      'আপনার সাথে দেখা হয়ে ভালো লাগলো।')

• With your own stemming / word segmentation implementation

  Custom stemmer/ tokenizer implementations must be callable objects, i.e. functions or classes with __call__ method implemented. If lang is also given, user provided implementations take precedence over the library provided ones.

from multilingual_rouge import rouge_scorer

# example with custom stemming
class DummyStemmer(object):
    def __call__(self, token):
        stem = ""
        # your stemmer implementation
        return stem

# example with custom segmenter/tokenizer
def dummy_tokenize(text):
    tokens = []
    # your tokenizer implementation
    return tokens

scorer = rouge_scorer.RougeScorer(['rouge1', 'rougeL'], use_stemmer=True, 
                                    callable_stemmer=DummyStemmer(),
                                    callable_tokenizer=dummy_tokenize)
scores = scorer.score('The quick brown fox jumps over the lazy dog',
                      'The quick brown dog jumps on the log.')
                      

• To see list of all available keyword arguments and reference stemmer and segmenter implementations refer to rouge_scorer.py, stemmers.py and tokenizers.py

License

Originally licensed under the Apache 2.0 License.