wil.py

# Copyright 2021 The HuggingFace Evaluate Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Word Information Loss (WIL) metric. """

import datasets
from jiwer import process_words

import evaluate


_CITATION = """\
@inproceedings{inproceedings,
    author = {Morris, Andrew and Maier, Viktoria and Green, Phil},
    year = {2004},
    month = {01},
    pages = {},
    title = {From WER and RIL to MER and WIL: improved evaluation measures for connected speech recognition.}
}y
"""

_DESCRIPTION = """\
Word Information Loss can be used to evaluate the performance of an automatic speech recognizer. It has information-theoretic backings, is symmetric between predictions and targets, and is bounded between 0 and 1.
The formula for WIL is
WIL = 1 - (C/P)(C/T)
where
C is the number of correct words,
P is the number of words in the prediction,
T is the number of words in the target.

This value measures the amount of information loss between two sentences. A score of 0 indicates that the prediction and target match perfectly.
"""

_KWARGS_DESCRIPTION = """
Compute the WIL between two sets of words.
Args:
    targets: List of target words.
    predictions: List of transcriptions to evaluate.
    concatenate_texts (bool, default=False): Whether to concatenate the WIL of the concanated strings or the mean WIL for each pair.
Returns:
    (float): the word information loss
Examples:
    >>> predictions = ["this is a prediction", "there is an other sample"]
    >>> targets = ["this is the target", "there is another one"]
    >>> wil = evaluate.load("wil")
    >>> wil_score = wil.compute(predictions=predictions, targets=targets)
    >>> print(wil_score)
    0.775
    >>> wil_score = wil.compute(predictions=targets, targets=predictions)
    >>> print(wil_score)
    0.775
"""


@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class WIL(evaluate.Metric):
    def _info(self):
        return evaluate.MetricInfo(
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            features=datasets.Features(
                {
                    "predictions": datasets.Value("string", id="sequence"),
                    "references": datasets.Value("string", id="sequence"),
                }
            ),
            codebase_urls=["https://github.com/jitsi/jiwer/"],
            reference_urls=[
                "https://en.wikipedia.org/wiki/Word_error_rate",
            ],
        )

    def _compute(self, predictions=None, references=None, concatenate_texts=False):
        if concatenate_texts:
            return process_words(references, predictions).wil
        else:
            total = 0
            for prediction, reference in zip(predictions, references):
                measures = process_words(reference, prediction).wil
                total += measures
            return total/len(predictions)