Skip to content

GitLab

Commit 25991f98 authored by hepj's avatar hepj
Browse files

修改readme

parent ac192496
Pipeline #1415 failed with stages
in 0 seconds
Hide whitespace changes
Inline Side-by-side
Showing with 1958 additions and 0 deletions
evaluate-0.4.2/metrics/code_eval/README.md 0 → 100644
View file @ 25991f98
---
title: Code Eval
emoji: 🤗
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false
tags:
- evaluate
- metric
description: >-
This metric implements the evaluation harness for the HumanEval problem solving dataset
described in the paper "Evaluating Large Language Models Trained on Code"
(https://arxiv.org/abs/2107.03374).
---
# Metric Card for Code Eval
## Metric description
The CodeEval metric estimates the pass@k metric for code synthesis.
It implements the evaluation harness for the HumanEval problem solving dataset described in the paper ["Evaluating Large Language Models Trained on Code"](https://arxiv.org/abs/2107.03374).
## How to use
The Code Eval metric calculates how good are predictions given a set of references. Its arguments are:
`predictions`: a list of candidates to evaluate. Each candidate should be a list of strings with several code candidates to solve the problem.
`references`: a list with a test for each prediction. Each test should evaluate the correctness of a code candidate.
`k`: number of code candidates to consider in the evaluation. The default value is `[1, 10, 100]`.
`num_workers`: the number of workers used to evaluate the candidate programs (The default value is `4`).
`timeout`: The maximum time taken to produce a prediction before it is considered a "timeout". The default value is `3.0` (i.e. 3 seconds).
```python
from evaluate import load
code_eval = load("code_eval")
test_cases = ["assert add(2,3)==5"]
candidates = [["def add(a,b): return a*b", "def add(a, b): return a+b"]]
pass_at_k, results = code_eval.compute(references=test_cases, predictions=candidates, k=[1, 2])
```
N.B.
This metric exists to run untrusted model-generated code. Users are strongly encouraged not to do so outside of a robust security sandbox. Before running this metric and once you've taken the necessary precautions, you will need to set the `HF_ALLOW_CODE_EVAL` environment variable. Use it at your own risk:
```python
import os
os.environ["HF_ALLOW_CODE_EVAL"] = "1"`
```
## Output values
The Code Eval metric outputs two things:
`pass_at_k`: a dictionary with the pass rates for each k value defined in the arguments.
`results`: a dictionary with granular results of each unit test.
### Values from popular papers
The [original CODEX paper](https://arxiv.org/pdf/2107.03374.pdf) reported that the CODEX-12B model had a pass@k score of 28.8% at `k=1`, 46.8% at `k=10` and 72.3% at `k=100`. However, since the CODEX model is not open source, it is hard to verify these numbers.
## Examples
Full match at `k=1`:
```python
from evaluate import load
code_eval = load("code_eval")
test_cases = ["assert add(2,3)==5"]
candidates = [["def add(a, b): return a+b"]]
pass_at_k, results = code_eval.compute(references=test_cases, predictions=candidates, k=[1])
print(pass_at_k)
{'pass@1': 1.0}
```
No match for k = 1:
```python
from evaluate import load
code_eval = load("code_eval")
test_cases = ["assert add(2,3)==5"]
candidates = [["def add(a,b): return a*b"]]
pass_at_k, results = code_eval.compute(references=test_cases, predictions=candidates, k=[1])
print(pass_at_k)
{'pass@1': 0.0}
```
Partial match at k=1, full match at k=2:
```python
from evaluate import load
code_eval = load("code_eval")
test_cases = ["assert add(2,3)==5"]
candidates = [["def add(a, b): return a+b", "def add(a,b): return a*b"]]
pass_at_k, results = code_eval.compute(references=test_cases, predictions=candidates, k=[1, 2])
print(pass_at_k)
{'pass@1': 0.5, 'pass@2': 1.0}
```
## Limitations and bias
As per the warning included in the metric code itself:
> This program exists to execute untrusted model-generated code. Although it is highly unlikely that model-generated code will do something overtly malicious in response to this test suite, model-generated code may act destructively due to a lack of model capability or alignment. Users are strongly encouraged to sandbox this evaluation suite so that it does not perform destructive actions on their host or network. For more information on how OpenAI sandboxes its code, see the accompanying paper. Once you have read this disclaimer and taken appropriate precautions, uncomment the following line and proceed at your own risk:
More information about the limitations of the code can be found on the [Human Eval Github repository](https://github.com/openai/human-eval).
## Citation
```bibtex
@misc{chen2021evaluating,
title={Evaluating Large Language Models Trained on Code},
author={Mark Chen and Jerry Tworek and Heewoo Jun and Qiming Yuan \
and Henrique Ponde de Oliveira Pinto and Jared Kaplan and Harri Edwards \
and Yuri Burda and Nicholas Joseph and Greg Brockman and Alex Ray \
and Raul Puri and Gretchen Krueger and Michael Petrov and Heidy Khlaaf \
and Girish Sastry and Pamela Mishkin and Brooke Chan and Scott Gray \
and Nick Ryder and Mikhail Pavlov and Alethea Power and Lukasz Kaiser \
and Mohammad Bavarian and Clemens Winter and Philippe Tillet \
and Felipe Petroski Such and Dave Cummings and Matthias Plappert \
and Fotios Chantzis and Elizabeth Barnes and Ariel Herbert-Voss \
and William Hebgen Guss and Alex Nichol and Alex Paino and Nikolas Tezak \
and Jie Tang and Igor Babuschkin and Suchir Balaji and Shantanu Jain \
and William Saunders and Christopher Hesse and Andrew N. Carr \
and Jan Leike and Josh Achiam and Vedant Misra and Evan Morikawa \
and Alec Radford and Matthew Knight and Miles Brundage and Mira Murati \
and Katie Mayer and Peter Welinder and Bob McGrew and Dario Amodei \
and Sam McCandlish and Ilya Sutskever and Wojciech Zaremba},
year={2021},
eprint={2107.03374},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
```
## Further References
- [Human Eval Github repository](https://github.com/openai/human-eval)
- [OpenAI Codex website](https://openai.com/blog/openai-codex/)
evaluate-0.4.2/metrics/code_eval/app.py 0 → 100644
View file @ 25991f98
import evaluate
from evaluate.utils import launch_gradio_widget
module = evaluate.load("code_eval")
launch_gradio_widget(module)
evaluate-0.4.2/metrics/code_eval/code_eval.py 0 → 100644
View file @ 25991f98
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# 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.
"""The CodeEval metric estimates the pass@k metric for code synthesis.
This is an evaluation harness for the HumanEval problem solving dataset
described in the paper "Evaluating Large Language Models Trained on Code"
(https://arxiv.org/abs/2107.03374)."""
import itertools
import os
from collections import Counter, defaultdict
from concurrent.futures import ThreadPoolExecutor, as_completed
import datasets
import numpy as np
import evaluate
from .execute import check_correctness
_CITATION = """\
@misc{chen2021evaluating,
title={Evaluating Large Language Models Trained on Code},
author={Mark Chen and Jerry Tworek and Heewoo Jun and Qiming Yuan \
and Henrique Ponde de Oliveira Pinto and Jared Kaplan and Harri Edwards \
and Yuri Burda and Nicholas Joseph and Greg Brockman and Alex Ray \
and Raul Puri and Gretchen Krueger and Michael Petrov and Heidy Khlaaf \
and Girish Sastry and Pamela Mishkin and Brooke Chan and Scott Gray \
and Nick Ryder and Mikhail Pavlov and Alethea Power and Lukasz Kaiser \
and Mohammad Bavarian and Clemens Winter and Philippe Tillet \
and Felipe Petroski Such and Dave Cummings and Matthias Plappert \
and Fotios Chantzis and Elizabeth Barnes and Ariel Herbert-Voss \
and William Hebgen Guss and Alex Nichol and Alex Paino and Nikolas Tezak \
and Jie Tang and Igor Babuschkin and Suchir Balaji and Shantanu Jain \
and William Saunders and Christopher Hesse and Andrew N. Carr \
and Jan Leike and Josh Achiam and Vedant Misra and Evan Morikawa \
and Alec Radford and Matthew Knight and Miles Brundage and Mira Murati \
and Katie Mayer and Peter Welinder and Bob McGrew and Dario Amodei \
and Sam McCandlish and Ilya Sutskever and Wojciech Zaremba},
year={2021},
eprint={2107.03374},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
"""
_DESCRIPTION = """\
This metric implements the evaluation harness for the HumanEval problem solving dataset
described in the paper "Evaluating Large Language Models Trained on Code"
(https://arxiv.org/abs/2107.03374).
"""
_KWARGS_DESCRIPTION = """
Calculates how good are predictions given some references, using certain scores
Args:
predictions: list of candidates to evaluate. Each candidates should be a list
of strings with several code candidates to solve the problem.
references: a list with a test for each prediction. Each test should evaluate the
correctness of a code candidate.
k: number of code candidates to consider in the evaluation (Default: [1, 10, 100])
num_workers: number of workers used to evaluate the canidate programs (Default: 4).
timeout:
Returns:
pass_at_k: dict with pass rates for each k
results: dict with granular results of each unittest
Examples:
>>> code_eval = evaluate.load("code_eval")
>>> test_cases = ["assert add(2,3)==5"]
>>> candidates = [["def add(a,b): return a*b", "def add(a, b): return a+b"]]
>>> pass_at_k, results = code_eval.compute(references=test_cases, predictions=candidates, k=[1, 2])
>>> print(pass_at_k)
{'pass@1': 0.5, 'pass@2': 1.0}
"""
_WARNING = """
################################################################################
!!!WARNING!!!
################################################################################
The "code_eval" metric executes untrusted model-generated code in Python.
Although it is highly unlikely that model-generated code will do something
overtly malicious in response to this test suite, model-generated code may act
destructively due to a lack of model capability or alignment.
Users are strongly encouraged to sandbox this evaluation suite so that it
does not perform destructive actions on their host or network. For more
information on how OpenAI sandboxes its code, see the paper "Evaluating Large
Language Models Trained on Code" (https://arxiv.org/abs/2107.03374).
Once you have read this disclaimer and taken appropriate precautions,
set the environment variable HF_ALLOW_CODE_EVAL="1". Within Python you can to this
with:
>>> import os
>>> os.environ["HF_ALLOW_CODE_EVAL"] = "1"
################################################################################\
"""
_LICENSE = """The MIT License
Copyright (c) OpenAI (https://openai.com)
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE."""
@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class CodeEval(evaluate.Metric):
def _info(self):
return evaluate.MetricInfo(
# This is the description that will appear on the metrics page.
description=_DESCRIPTION,
citation=_CITATION,
inputs_description=_KWARGS_DESCRIPTION,
# This defines the format of each prediction and reference
features=datasets.Features(
{
"predictions": datasets.Sequence(datasets.Value("string")),
"references": datasets.Value("string"),
}
),
homepage="https://github.com/openai/human-eval",
codebase_urls=["https://github.com/openai/human-eval"],
reference_urls=["https://github.com/openai/human-eval"],
license=_LICENSE,
)
def _compute(self, predictions, references, k=[1, 10, 100], num_workers=4, timeout=3.0):
"""Returns the scores"""
if os.getenv("HF_ALLOW_CODE_EVAL", 0) != "1":
raise ValueError(_WARNING)
if os.name == "nt":
raise NotImplementedError("This metric is currently not supported on Windows.")
with ThreadPoolExecutor(max_workers=num_workers) as executor:
futures = []
completion_id = Counter()
n_samples = 0
results = defaultdict(list)
for task_id, (candidates, test_case) in enumerate(zip(predictions, references)):
for candidate in candidates:
test_program = candidate + "\n" + test_case
args = (test_program, timeout, task_id, completion_id[task_id])
future = executor.submit(check_correctness, *args)
futures.append(future)
completion_id[task_id] += 1
n_samples += 1
for future in as_completed(futures):
result = future.result()
results[result["task_id"]].append((result["completion_id"], result))
total, correct = [], []
for result in results.values():
result.sort()
passed = [r[1]["passed"] for r in result]
total.append(len(passed))
correct.append(sum(passed))
total = np.array(total)
correct = np.array(correct)
ks = k
pass_at_k = {f"pass@{k}": estimate_pass_at_k(total, correct, k).mean() for k in ks if (total >= k).all()}
return pass_at_k, results
def estimate_pass_at_k(num_samples, num_correct, k):
"""Estimates pass@k of each problem and returns them in an array."""
def estimator(n: int, c: int, k: int) -> float:
"""Calculates 1 - comb(n - c, k) / comb(n, k)."""
if n - c < k:
return 1.0
return 1.0 - np.prod(1.0 - k / np.arange(n - c + 1, n + 1))
if isinstance(num_samples, int):
num_samples_it = itertools.repeat(num_samples, len(num_correct))
else:
assert len(num_samples) == len(num_correct)
num_samples_it = iter(num_samples)
return np.array([estimator(int(n), int(c), k) for n, c in zip(num_samples_it, num_correct)])
evaluate-0.4.2/metrics/code_eval/execute.py 0 → 100644
View file @ 25991f98
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# 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.
# This code is adapted from OpenAI's release
# https://github.com/openai/human-eval/blob/master/human_eval/execution.py
import contextlib
import faulthandler
import io
import multiprocessing
import os
import platform
import signal
import tempfile
def check_correctness(check_program, timeout, task_id, completion_id):
"""
Evaluates the functional correctness of a completion by running the test
suite provided in the problem.
:param completion_id: an optional completion ID so we can match
the results later even if execution finishes asynchronously.
"""
manager = multiprocessing.Manager()
result = manager.list()
p = multiprocessing.Process(target=unsafe_execute, args=(check_program, result, timeout))
p.start()
p.join(timeout=timeout + 1)
if p.is_alive():
p.kill()
if not result:
result.append("timed out")
return dict(
task_id=task_id,
passed=result[0] == "passed",
result=result[0],
completion_id=completion_id,
)
def unsafe_execute(check_program, result, timeout):
with create_tempdir():
# These system calls are needed when cleaning up tempdir.
import os
import shutil
rmtree = shutil.rmtree
rmdir = os.rmdir
chdir = os.chdir
# Disable functionalities that can make destructive changes to the test.
reliability_guard()
# Run program.
try:
exec_globals = {}
with swallow_io():
with time_limit(timeout):
exec(check_program, exec_globals)
result.append("passed")
except TimeoutException:
result.append("timed out")
except BaseException as e:
result.append(f"failed: {e}")
# Needed for cleaning up.
shutil.rmtree = rmtree
os.rmdir = rmdir
os.chdir = chdir
@contextlib.contextmanager
def time_limit(seconds):
def signal_handler(signum, frame):
raise TimeoutException("Timed out!")
signal.setitimer(signal.ITIMER_REAL, seconds)
signal.signal(signal.SIGALRM, signal_handler)
try:
yield
finally:
signal.setitimer(signal.ITIMER_REAL, 0)
@contextlib.contextmanager
def swallow_io():
stream = WriteOnlyStringIO()
with contextlib.redirect_stdout(stream):
with contextlib.redirect_stderr(stream):
with redirect_stdin(stream):
yield
@contextlib.contextmanager
def create_tempdir():
with tempfile.TemporaryDirectory() as dirname:
with chdir(dirname):
yield dirname
class TimeoutException(Exception):
pass
class WriteOnlyStringIO(io.StringIO):
"""StringIO that throws an exception when it's read from"""
def read(self, *args, **kwargs):
raise OSError
def readline(self, *args, **kwargs):
raise OSError
def readlines(self, *args, **kwargs):
raise OSError
def readable(self, *args, **kwargs):
"""Returns True if the IO object can be read."""
return False
class redirect_stdin(contextlib._RedirectStream): # type: ignore
_stream = "stdin"
@contextlib.contextmanager
def chdir(root):
if root == ".":
yield
return
cwd = os.getcwd()
os.chdir(root)
try:
yield
except BaseException as exc:
raise exc
finally:
os.chdir(cwd)
def reliability_guard(maximum_memory_bytes=None):
"""
This disables various destructive functions and prevents the generated code
from interfering with the test (e.g. fork bomb, killing other processes,
removing filesystem files, etc.)
WARNING
This function is NOT a security sandbox. Untrusted code, including, model-
generated code, should not be blindly executed outside of one. See the
Codex paper for more information about OpenAI's code sandbox, and proceed
with caution.
"""
if maximum_memory_bytes is not None:
import resource
resource.setrlimit(resource.RLIMIT_AS, (maximum_memory_bytes, maximum_memory_bytes))
resource.setrlimit(resource.RLIMIT_DATA, (maximum_memory_bytes, maximum_memory_bytes))
if not platform.uname().system == "Darwin":
resource.setrlimit(resource.RLIMIT_STACK, (maximum_memory_bytes, maximum_memory_bytes))
faulthandler.disable()
import builtins
builtins.exit = None
builtins.quit = None
import os
os.environ["OMP_NUM_THREADS"] = "1"
os.kill = None
os.system = None
os.putenv = None
os.remove = None
os.removedirs = None
os.rmdir = None
os.fchdir = None
os.setuid = None
os.fork = None
os.forkpty = None
os.killpg = None
os.rename = None
os.renames = None
os.truncate = None
os.replace = None
os.unlink = None
os.fchmod = None
os.fchown = None
os.chmod = None
os.chown = None
os.chroot = None
os.fchdir = None
os.lchflags = None
os.lchmod = None
os.lchown = None
os.getcwd = None
os.chdir = None
import shutil
shutil.rmtree = None
shutil.move = None
shutil.chown = None
import subprocess
subprocess.Popen = None # type: ignore
__builtins__["help"] = None
import sys
sys.modules["ipdb"] = None
sys.modules["joblib"] = None
sys.modules["resource"] = None
sys.modules["psutil"] = None
sys.modules["tkinter"] = None
evaluate-0.4.2/metrics/code_eval/requirements.txt 0 → 100644
View file @ 25991f98
git+https://github.com/huggingface/evaluate@{COMMIT_PLACEHOLDER}
\ No newline at end of file
evaluate-0.4.2/metrics/comet/README.md 0 → 100644
View file @ 25991f98
---
title: COMET
emoji: 🤗
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false
tags:
- evaluate
- metric
description: >-
Crosslingual Optimized Metric for Evaluation of Translation (COMET) is an open-source framework used to train Machine Translation metrics that achieve high levels of correlation with different types of human judgments (HTER, DA's or MQM).
With the release of the framework the authors also released fully trained models that were used to compete in the WMT20 Metrics Shared Task achieving SOTA in that years competition.
See the [README.md] file at https://unbabel.github.io/COMET/html/models.html for more information.
---
# Metric Card for COMET
## Metric description
Crosslingual Optimized Metric for Evaluation of Translation (COMET) is an open-source framework used to train Machine Translation metrics that achieve high levels of correlation with different types of human judgments.
## How to use
COMET takes 3 lists of strings as input: `sources` (a list of source sentences), `predictions` (a list of candidate translations) and `references` (a list of reference translations).
```python
from evaluate import load
comet_metric = load('comet')
source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
hypothesis = ["The fire could be stopped", "Schools and kindergartens were open"]
reference = ["They were able to control the fire.", "Schools and kindergartens opened"]
comet_score = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
```
It has several configurations, named after the COMET model to be used. For versions below 2.0 it will default to `wmt20-comet-da` (previously known as `wmt-large-da-estimator-1719`) and for the latest versions (>= 2.0) it will default to `Unbabel/wmt22-comet-da`.
Alternative models that can be chosen include `wmt20-comet-qe-da`, `wmt21-comet-mqm`, `wmt21-cometinho-da`, `wmt21-comet-qe-mqm` and `emnlp20-comet-rank`. Notably, a distilled model is also available, which is 80% smaller and 2.128x faster while performing close to non-distilled alternatives. You can use it with the identifier `eamt22-cometinho-da`. This version, called Cometinho, was elected as [the best paper](https://aclanthology.org/2022.eamt-1.9) at the annual European conference on Machine Translation.
> NOTE: In `unbabel-comet>=2.0` all models were moved to Hugging Face Hub and you need to add the suffix `Unbabel/` to be able to download and use them. For example for the distilled version replace `eamt22-cometinho-da` with `Unbabel/eamt22-cometinho-da`.
It also has several optional arguments:
`gpus`: optional, an integer (number of GPUs to train on) or a list of integers (which GPUs to train on). Set to 0 to use CPU. The default value is `None` (uses one GPU if possible, else use CPU).
`progress_bar`a boolean -- if set to `True`, progress updates will be printed out. The default value is `False`.
More information about model characteristics can be found on the [COMET website](https://unbabel.github.io/COMET/html/index.html).
## Output values
The COMET metric outputs two lists:
`scores`: a list of COMET scores for each of the input sentences, ranging from 0-1.
`mean_score`: the mean value of COMET scores `scores` over all the input sentences, ranging from 0-1.
### Values from popular papers
The [original COMET paper](https://arxiv.org/pdf/2009.09025.pdf) reported average COMET scores ranging from 0.4 to 0.6, depending on the language pairs used for evaluating translation models. They also illustrate that COMET correlates well with human judgements compared to other metrics such as [BLEU](https://huggingface.co/metrics/bleu) and [CHRF](https://huggingface.co/metrics/chrf).
## Examples
Full match:
```python
from evaluate import load
comet_metric = load('comet')
source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
hypothesis = ["They were able to control the fire.", "Schools and kindergartens opened"]
reference = ["They were able to control the fire.", "Schools and kindergartens opened"]
results = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
print([round(v, 1) for v in results["scores"]])
[1.0, 1.0]
```
Partial match:
```python
from evaluate import load
comet_metric = load('comet')
source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
hypothesis = ["The fire could be stopped", "Schools and kindergartens were open"]
reference = ["They were able to control the fire", "Schools and kindergartens opened"]
results = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
print([round(v, 2) for v in results["scores"]])
[0.19, 0.92]
```
No match:
```python
from evaluate import load
comet_metric = load('comet')
source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
hypothesis = ["The girl went for a walk", "The boy was sleeping"]
reference = ["They were able to control the fire", "Schools and kindergartens opened"]
results = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
print([round(v, 2) for v in results["scores"]])
[0.00, 0.00]
```
## Limitations and bias
The models provided for calculating the COMET metric are built on top of XLM-R and cover the following languages:
Afrikaans, Albanian, Amharic, Arabic, Armenian, Assamese, Azerbaijani, Basque, Belarusian, Bengali, Bengali Romanized, Bosnian, Breton, Bulgarian, Burmese, Burmese, Catalan, Chinese (Simplified), Chinese (Traditional), Croatian, Czech, Danish, Dutch, English, Esperanto, Estonian, Filipino, Finnish, French, Galician, Georgian, German, Greek, Gujarati, Hausa, Hebrew, Hindi, Hindi Romanized, Hungarian, Icelandic, Indonesian, Irish, Italian, Japanese, Javanese, Kannada, Kazakh, Khmer, Korean, Kurdish (Kurmanji), Kyrgyz, Lao, Latin, Latvian, Lithuanian, Macedonian, Malagasy, Malay, Malayalam, Marathi, Mongolian, Nepali, Norwegian, Oriya, Oromo, Pashto, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Sanskri, Scottish, Gaelic, Serbian, Sindhi, Sinhala, Slovak, Slovenian, Somali, Spanish, Sundanese, Swahili, Swedish, Tamil, Tamil Romanized, Telugu, Telugu Romanized, Thai, Turkish, Ukrainian, Urdu, Urdu Romanized, Uyghur, Uzbek, Vietnamese, Welsh, Western, Frisian, Xhosa, Yiddish.
Thus, results for language pairs containing uncovered languages are unreliable, as per the [COMET website](https://github.com/Unbabel/COMET)
Also, calculating the COMET metric involves downloading the model from which features are obtained -- the default model, `wmt22-comet-da`, takes over 2.32GB of storage space and downloading it can take a significant amount of time depending on the speed of your internet connection. If this is an issue, choose a smaller model; for instance `eamt22-cometinho-da` is 344MB.
### Interpreting Scores:
When using COMET to evaluate machine translation, it's important to understand how to interpret the scores it produces.
In general, COMET models are trained to predict quality scores for translations. These scores are typically normalized using a z-score transformation to account for individual differences among annotators. While the raw score itself does not have a direct interpretation, it is useful for ranking translations and systems according to their quality.
However, for the latest COMET models like `Unbabel/wmt22-comet-da`, we have introduced a new training approach that scales the scores between 0 and 1. This makes it easier to interpret the scores: a score close to 1 indicates a high-quality translation, while a score close to 0 indicates a translation that is no better than random chance.
It's worth noting that when using COMET to compare the performance of two different translation systems, it's important to run statistical significance measures to reliably compare scores between systems.
## Citation
```bibtex
@inproceedings{rei-etal-2022-comet,
title = "{COMET}-22: Unbabel-{IST} 2022 Submission for the Metrics Shared Task",
author = "Rei, Ricardo and
C. de Souza, Jos{\'e} G. and
Alves, Duarte and
Zerva, Chrysoula and
Farinha, Ana C and
Glushkova, Taisiya and
Lavie, Alon and
Coheur, Luisa and
Martins, Andr{\'e} F. T.",
booktitle = "Proceedings of the Seventh Conference on Machine Translation (WMT)",
month = dec,
year = "2022",
address = "Abu Dhabi, United Arab Emirates (Hybrid)",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.wmt-1.52",
pages = "578--585",
}
```
```bibtex
@inproceedings{rei-EtAl:2020:WMT,
author = {Rei, Ricardo and Stewart, Craig and Farinha, Ana C and Lavie, Alon},
title = {Unbabel's Participation in the WMT20 Metrics Shared Task},
booktitle = {Proceedings of the Fifth Conference on Machine Translation},
month = {November},
year = {2020},
address = {Online},
publisher = {Association for Computational Linguistics},
pages = {909--918},
}
```
```bibtex
@inproceedings{rei-etal-2020-comet,
title = "{COMET}: A Neural Framework for {MT} Evaluation",
author = "Rei, Ricardo and
Stewart, Craig and
Farinha, Ana C and
Lavie, Alon",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.emnlp-main.213",
pages = "2685--2702",
}
```
For the distilled version:
```bibtex
@inproceedings{rei-etal-2022-searching,
title = "Searching for {COMETINHO}: The Little Metric That Could",
author = "Rei, Ricardo and
Farinha, Ana C and
de Souza, Jos{\'e} G.C. and
Ramos, Pedro G. and
Martins, Andr{\'e} F.T. and
Coheur, Luisa and
Lavie, Alon",
booktitle = "Proceedings of the 23rd Annual Conference of the European Association for Machine Translation",
month = jun,
year = "2022",
address = "Ghent, Belgium",
publisher = "European Association for Machine Translation",
url = "https://aclanthology.org/2022.eamt-1.9",
pages = "61--70",
}
```
## Further References
- [COMET website](https://unbabel.github.io/COMET/html/index.html)
- [Hugging Face Tasks - Machine Translation](https://huggingface.co/tasks/translation)
evaluate-0.4.2/metrics/comet/app.py 0 → 100644
View file @ 25991f98
import sys
import evaluate
from evaluate.utils import launch_gradio_widget
sys.path = [p for p in sys.path if p != "/home/user/app"]
module = evaluate.load("comet")
sys.path = ["/home/user/app"] + sys.path
launch_gradio_widget(module)
evaluate-0.4.2/metrics/comet/comet.py 0 → 100644
View file @ 25991f98
# Copyright 2020 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.
""" COMET metric.
Requirements:
pip install unbabel-comet
Usage:
```python
from evaluate import load
comet_metric = load('metrics/comet/comet.py')
#comet_metric = load('comet')
#comet_metric = load('comet', 'Unbabel/wmt20-comet-da')
source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
hypothesis = ["The fire could be stopped", "Schools and kindergartens were open"]
reference = ["They were able to control the fire.", "Schools and kindergartens opened"]
predictions = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
predictions['scores']
```
"""
import comet # From: unbabel-comet
import datasets
import torch
from packaging import version
import evaluate
logger = evaluate.logging.get_logger(__name__)
_CITATION = """\
@inproceedings{rei-etal-2022-comet,
title = "{COMET}-22: Unbabel-{IST} 2022 Submission for the Metrics Shared Task",
author = "Rei, Ricardo and
C. de Souza, Jos{\'e} G. and
Alves, Duarte and
Zerva, Chrysoula and
Farinha, Ana C and
Glushkova, Taisiya and
Lavie, Alon and
Coheur, Luisa and
Martins, Andr{\'e} F. T.",
booktitle = "Proceedings of the Seventh Conference on Machine Translation (WMT)",
month = dec,
year = "2022",
address = "Abu Dhabi, United Arab Emirates (Hybrid)",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.wmt-1.52",
pages = "578--585",
}
@inproceedings{rei-EtAl:2020:WMT,
author = {Rei, Ricardo and Stewart, Craig and Farinha, Ana C and Lavie, Alon},
title = {Unbabel's Participation in the WMT20 Metrics Shared Task},
booktitle = {Proceedings of the Fifth Conference on Machine Translation},
month = {November},
year = {2020},
address = {Online},
publisher = {Association for Computational Linguistics},
pages = {909--918},
}
@inproceedings{rei-etal-2020-comet,
title = "{COMET}: A Neural Framework for {MT} Evaluation",
author = "Rei, Ricardo and
Stewart, Craig and
Farinha, Ana C and
Lavie, Alon",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.emnlp-main.213",
pages = "2685--2702",
}
"""
_DESCRIPTION = """\
Crosslingual Optimized Metric for Evaluation of Translation (COMET) is an open-source framework used to train Machine Translation metrics that achieve high levels of correlation with different types of human judgments (HTER, DA's or MQM).
With the release of the framework the authors also released fully trained models that were used to compete in the WMT20 Metrics Shared Task achieving SOTA in that years competition.
See the [README.md] file at https://unbabel.github.io/COMET/html/models.html for more information.
"""
_KWARGS_DESCRIPTION = """
COMET score.
Args:
`sources` (list of str): Source sentences
`predictions` (list of str): candidate translations
`references` (list of str): reference translations
`gpus` (bool): Number of GPUs to use. 0 for CPU
`progress_bar` (bool): Flag that turns on and off the predict progress bar. Defaults to True
Returns:
Dict with all sentence-level scores (`scores` key) a system-level score (`mean_score` key).
Examples:
>>> comet_metric = evaluate.load('comet')
>>> # comet_metric = load('comet', 'wmt20-comet-da') # you can also choose which model to use
>>> source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
>>> hypothesis = ["The fire could be stopped", "Schools and kindergartens were open"]
>>> reference = ["They were able to control the fire.", "Schools and kindergartens opened"]
>>> results = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
>>> print([round(v, 3) for v in results["scores"]])
[0.839, 0.972]
"""
@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class COMET(evaluate.Metric):
def _info(self):
return evaluate.MetricInfo(
description=_DESCRIPTION,
citation=_CITATION,
homepage="https://unbabel.github.io/COMET/html/index.html",
inputs_description=_KWARGS_DESCRIPTION,
features=datasets.Features(
{
"sources": datasets.Value("string", id="sequence"),
"predictions": datasets.Value("string", id="sequence"),
"references": datasets.Value("string", id="sequence"),
}
),
codebase_urls=["https://github.com/Unbabel/COMET"],
reference_urls=[
"https://github.com/Unbabel/COMET",
"https://aclanthology.org/2022.wmt-1.52/",
"https://www.aclweb.org/anthology/2020.emnlp-main.213/",
"http://www.statmt.org/wmt20/pdf/2020.wmt-1.101.pdf6",
],
)
def _download_and_prepare(self, dl_manager):
if self.config_name == "default":
if version.parse(comet.__version__) >= version.parse("2.0.0"):
self.scorer = comet.load_from_checkpoint(comet.download_model("Unbabel/wmt22-comet-da"))
else:
self.scorer = comet.load_from_checkpoint(comet.download_model("wmt20-comet-da"))
else:
self.scorer = comet.load_from_checkpoint(comet.download_model(self.config_name))
def _compute(self, sources, predictions, references, gpus=None, progress_bar=False):
if gpus is None:
gpus = 1 if torch.cuda.is_available() else 0
data = {"src": sources, "mt": predictions, "ref": references}
data = [dict(zip(data, t)) for t in zip(*data.values())]
if version.parse(comet.__version__) >= version.parse("2.0.0"):
output = self.scorer.predict(data, gpus=gpus, progress_bar=progress_bar)
scores, mean_score = output.scores, output.system_score
else:
scores, mean_score = self.scorer.predict(data, gpus=gpus, progress_bar=progress_bar)
return {"mean_score": mean_score, "scores": scores}
evaluate-0.4.2/metrics/comet/requirements.txt 0 → 100644
View file @ 25991f98
git+https://github.com/huggingface/evaluate@{COMMIT_PLACEHOLDER}
unbabel-comet
torch
\ No newline at end of file
evaluate-0.4.2/metrics/competition_math/README.md 0 → 100644
View file @ 25991f98
---
title: Competition MATH
emoji: 🤗
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false
tags:
- evaluate
- metric
description: >-
This metric is used to assess performance on the Mathematics Aptitude Test of Heuristics (MATH) dataset.
It first canonicalizes the inputs (e.g., converting "1/2" to "\frac{1}{2}") and then computes accuracy.
---
# Metric Card for Competition MATH
## Metric description
This metric is used to assess performance on the [Mathematics Aptitude Test of Heuristics (MATH) dataset](https://huggingface.co/datasets/competition_math).
It first canonicalizes the inputs (e.g., converting `1/2` to `\\frac{1}{2}`) and then computes accuracy.
## How to use
This metric takes two arguments:
`predictions`: a list of predictions to score. Each prediction is a string that contains natural language and LaTeX.
`references`: list of reference for each prediction. Each reference is a string that contains natural language and LaTeX.
```python
>>> from evaluate import load
>>> math = load("competition_math")
>>> references = ["\\frac{1}{2}"]
>>> predictions = ["1/2"]
>>> results = math.compute(references=references, predictions=predictions)
```
N.B. To be able to use Competition MATH, you need to install the `math_equivalence` dependency using `pip install git+https://github.com/hendrycks/math.git`.
## Output values
This metric returns a dictionary that contains the [accuracy](https://huggingface.co/metrics/accuracy) after canonicalizing inputs, on a scale between 0.0 and 1.0.
### Values from popular papers
The [original MATH dataset paper](https://arxiv.org/abs/2103.03874) reported accuracies ranging from 3.0% to 6.9% by different large language models.
More recent progress on the dataset can be found on the [dataset leaderboard](https://paperswithcode.com/sota/math-word-problem-solving-on-math).
## Examples
Maximal values (full match):
```python
>>> from evaluate import load
>>> math = load("competition_math")
>>> references = ["\\frac{1}{2}"]
>>> predictions = ["1/2"]
>>> results = math.compute(references=references, predictions=predictions)
>>> print(results)
{'accuracy': 1.0}
```
Minimal values (no match):
```python
>>> from evaluate import load
>>> math = load("competition_math")
>>> references = ["\\frac{1}{2}"]
>>> predictions = ["3/4"]
>>> results = math.compute(references=references, predictions=predictions)
>>> print(results)
{'accuracy': 0.0}
```
Partial match:
```python
>>> from evaluate import load
>>> math = load("competition_math")
>>> references = ["\\frac{1}{2}","\\frac{3}{4}"]
>>> predictions = ["1/5", "3/4"]
>>> results = math.compute(references=references, predictions=predictions)
>>> print(results)
{'accuracy': 0.5}
```
## Limitations and bias
This metric is limited to datasets with the same format as the [Mathematics Aptitude Test of Heuristics (MATH) dataset](https://huggingface.co/datasets/competition_math), and is meant to evaluate the performance of large language models at solving mathematical problems.
N.B. The MATH dataset also assigns levels of difficulty to different problems, so disagregating model performance by difficulty level (similarly to what was done in the [original paper](https://arxiv.org/abs/2103.03874) can give a better indication of how a given model does on a given difficulty of math problem, compared to overall accuracy.
## Citation
```bibtex
@article{hendrycksmath2021,
title={Measuring Mathematical Problem Solving With the MATH Dataset},
author={Dan Hendrycks
and Collin Burns
and Saurav Kadavath
and Akul Arora
and Steven Basart
and Eric Tang
and Dawn Song
and Jacob Steinhardt},
journal={arXiv preprint arXiv:2103.03874},
year={2021}
}
```
## Further References
- [MATH dataset](https://huggingface.co/datasets/competition_math)
- [MATH leaderboard](https://paperswithcode.com/sota/math-word-problem-solving-on-math)
- [MATH paper](https://arxiv.org/abs/2103.03874)
\ No newline at end of file
evaluate-0.4.2/metrics/competition_math/app.py 0 → 100644
View file @ 25991f98
import evaluate
from evaluate.utils import launch_gradio_widget
module = evaluate.load("competition_math")
launch_gradio_widget(module)
evaluate-0.4.2/metrics/competition_math/competition_math.py 0 → 100644
View file @ 25991f98
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# 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.
"""Accuracy metric for the Mathematics Aptitude Test of Heuristics (MATH) dataset."""
import datasets
import math_equivalence # From: git+https://github.com/hendrycks/math.git
import evaluate
_CITATION = """\
@article{hendrycksmath2021,
title={Measuring Mathematical Problem Solving With the MATH Dataset},
author={Dan Hendrycks
and Collin Burns
and Saurav Kadavath
and Akul Arora
and Steven Basart
and Eric Tang
and Dawn Song
and Jacob Steinhardt},
journal={arXiv preprint arXiv:2103.03874},
year={2021}
}
"""
_DESCRIPTION = """\
This metric is used to assess performance on the Mathematics Aptitude Test of Heuristics (MATH) dataset.
It first canonicalizes the inputs (e.g., converting "1/2" to "\\frac{1}{2}") and then computes accuracy.
"""
_KWARGS_DESCRIPTION = r"""
Calculates accuracy after canonicalizing inputs.
Args:
predictions: list of predictions to score. Each prediction
is a string that contains natural language and LaTex.
references: list of reference for each prediction. Each
reference is a string that contains natural language
and LaTex.
Returns:
accuracy: accuracy after canonicalizing inputs
(e.g., converting "1/2" to "\\frac{1}{2}")
Examples:
>>> metric = evaluate.load("competition_math")
>>> results = metric.compute(references=["\\frac{1}{2}"], predictions=["1/2"])
>>> print(results)
{'accuracy': 1.0}
"""
@datasets.utils.file_utils.add_end_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class CompetitionMathMetric(evaluate.Metric):
"""Accuracy metric for the MATH dataset."""
def _info(self):
return evaluate.MetricInfo(
description=_DESCRIPTION,
citation=_CITATION,
inputs_description=_KWARGS_DESCRIPTION,
features=datasets.Features(
{
"predictions": datasets.Value("string"),
"references": datasets.Value("string"),
}
),
# Homepage of the metric for documentation
homepage="https://github.com/hendrycks/math",
# Additional links to the codebase or references
codebase_urls=["https://github.com/hendrycks/math"],
)
def _compute(self, predictions, references):
"""Returns the scores"""
n_correct = 0.0
for i, j in zip(predictions, references):
n_correct += 1.0 if math_equivalence.is_equiv(i, j) else 0.0
accuracy = n_correct / len(predictions)
return {
"accuracy": accuracy,
}
evaluate-0.4.2/metrics/competition_math/requirements.txt 0 → 100644
View file @ 25991f98
git+https://github.com/huggingface/evaluate@{COMMIT_PLACEHOLDER}
git+https://github.com/hendrycks/math.git
\ No newline at end of file
evaluate-0.4.2/metrics/confusion_matrix/README.md 0 → 100644
View file @ 25991f98
---
title: Confusion Matrix
emoji: 🤗
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false
tags:
- evaluate
- metric
description: >-
The confusion matrix evaluates classification accuracy.
Each row in a confusion matrix represents a true class and each column represents the instances in a predicted class.
---
# Metric Card for Confusion Matrix
## Metric Description
The confusion matrix evaluates classification accuracy. Each row in a confusion matrix represents a true class and each column represents the instances in a predicted class. Let's look at an example:
| | setosa | versicolor | virginica |
| ---------- | ------ | ---------- | --------- |
| setosa | 13 | 0 | 0 |
| versicolor | 0 | 10 | 6 |
| virginica | 0 | 0 | 9 |
What information does this confusion matrix provide?
* All setosa instances were properly predicted as such (true positives).
* The model always correctly classifies the setosa class (there are no false positives).
* 10 versicolor instances were properly classified, but 6 instances were misclassified as virginica.
* All virginica insances were properly classified as such.
## How to Use
At minimum, this metric requires predictions and references as inputs.
```python
>>> confusion_metric = evaluate.load("confusion_matrix")
>>> results = confusion_metric.compute(references=[0, 1, 1, 2, 0, 2, 2], predictions=[0, 2, 1, 1, 0, 2, 0])
>>> print(results)
{'confusion_matrix': [[2, 0, 0], [0, 1, 1], [1, 1, 1]]}
```
### Inputs
- **predictions** (`list` of `int`): Predicted labels.
- **references** (`list` of `int`): Ground truth labels.
- **labels** (`list` of `int`): List of labels to index the matrix. This may be used to reorder or select a subset of labels.
- **sample_weight** (`list` of `float`): Sample weights.
- **normalize** (`str`): Normalizes confusion matrix over the true (rows), predicted (columns) conditions or all the population.
### Output Values
- **confusion_matrix**(`list` of `list` of `str`): Confusion matrix. Minimum possible value is 0. Maximum possible value is 1.0, or the number of examples input, if `normalize` is set to `True`.
Output Example(s):
```python
{'confusion_matrix': [[2, 0, 0], [0, 1, 1], [1, 1, 1]]}
```
This metric outputs a dictionary, containing the confusion matrix.
### Examples
Example 1 - A simple example
```python
>>> confusion_metric = evaluate.load("confusion_matrix")
>>> results = confusion_metric.compute(references=[0, 1, 1, 2, 0, 2, 2], predictions=[0, 2, 1, 1, 0, 2, 0])
>>> print(results)
{'confusion_matrix': [[2, 0, 0], [0, 1, 1], [1, 1, 1]]}
```
## Citation(s)
```bibtex
@article{scikit-learn,
title={Scikit-learn: Machine Learning in {P}ython},
author={Pedregosa, F. and Varoquaux, G. and Gramfort, A. and Michel, V.
and Thirion, B. and Grisel, O. and Blondel, M. and Prettenhofer, P.
and Weiss, R. and Dubourg, V. and Vanderplas, J. and Passos, A. and
Cournapeau, D. and Brucher, M. and Perrot, M. and Duchesnay, E.},
journal={Journal of Machine Learning Research},
volume={12},
pages={2825--2830},
year={2011}
}
```
## Further References
* https://scikit-learn.org/stable/modules/generated/sklearn.metrics.confusion_matrix.html
* https://en.wikipedia.org/wiki/Confusion_matrix
\ No newline at end of file
evaluate-0.4.2/metrics/confusion_matrix/app.py 0 → 100644
View file @ 25991f98
import evaluate
from evaluate.utils import launch_gradio_widget
module = evaluate.load("confusion_matrix")
launch_gradio_widget(module)
evaluate-0.4.2/metrics/confusion_matrix/confusion_matrix.py 0 → 100644
View file @ 25991f98
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# 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.
"""Confusion Matrix."""
import datasets
from sklearn.metrics import confusion_matrix
import evaluate
_DESCRIPTION = """
The confusion matrix evaluates classification accuracy. Each row in a confusion matrix represents a true class and each column represents the instances in a predicted class
"""
_KWARGS_DESCRIPTION = """
Args:
predictions (`list` of `int`): Predicted labels.
references (`list` of `int`): Ground truth labels.
labels (`list` of `int`): List of labels to index the matrix. This may be used to reorder or select a subset of labels.
sample_weight (`list` of `float`): Sample weights.
normalize (`str`): Normalizes confusion matrix over the true (rows), predicted (columns) conditions or all the population.
Returns:
confusion_matrix (`list` of `list` of `int`): Confusion matrix whose i-th row and j-th column entry indicates the number of samples with true label being i-th class and predicted label being j-th class.
Examples:
Example 1-A simple example
>>> confusion_matrix_metric = evaluate.load("confusion_matrix")
>>> results = confusion_matrix_metric.compute(references=[0, 1, 2, 0, 1, 2], predictions=[0, 1, 1, 2, 1, 0])
>>> print(results) # doctest: +ELLIPSIS +NORMALIZE_WHITESPACE
{'confusion_matrix': array([[1, 0, 1], [0, 2, 0], [1, 1, 0]][...])}
"""
_CITATION = """
@article{scikit-learn,
title={Scikit-learn: Machine Learning in {P}ython},
author={Pedregosa, F. and Varoquaux, G. and Gramfort, A. and Michel, V.
and Thirion, B. and Grisel, O. and Blondel, M. and Prettenhofer, P.
and Weiss, R. and Dubourg, V. and Vanderplas, J. and Passos, A. and
Cournapeau, D. and Brucher, M. and Perrot, M. and Duchesnay, E.},
journal={Journal of Machine Learning Research},
volume={12},
pages={2825--2830},
year={2011}
}
"""
@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class ConfusionMatrix(evaluate.Metric):
def _info(self):
return evaluate.MetricInfo(
description=_DESCRIPTION,
citation=_CITATION,
inputs_description=_KWARGS_DESCRIPTION,
features=datasets.Features(
{
"predictions": datasets.Sequence(datasets.Value("int32")),
"references": datasets.Sequence(datasets.Value("int32")),
}
if self.config_name == "multilabel"
else {
"predictions": datasets.Value("int32"),
"references": datasets.Value("int32"),
}
),
reference_urls=["https://scikit-learn.org/stable/modules/generated/sklearn.metrics.confusion_matrix.html"],
)
def _compute(self, predictions, references, labels=None, sample_weight=None, normalize=None):
return {
"confusion_matrix": confusion_matrix(
references, predictions, labels=labels, sample_weight=sample_weight, normalize=normalize
)
}
evaluate-0.4.2/metrics/confusion_matrix/requirements.txt 0 → 100644
View file @ 25991f98
git+https://github.com/huggingface/evaluate@{COMMIT_PLACEHOLDER}
scikit-learn
\ No newline at end of file
evaluate-0.4.2/metrics/coval/README.md 0 → 100644
View file @ 25991f98
---
title:
emoji: 🤗
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false
tags:
- evaluate
- metric
description: >-
CoVal is a coreference evaluation tool for the CoNLL and ARRAU datasets which
implements of the common evaluation metrics including MUC [Vilain et al, 1995],
B-cubed [Bagga and Baldwin, 1998], CEAFe [Luo et al., 2005],
LEA [Moosavi and Strube, 2016] and the averaged CoNLL score
(the average of the F1 values of MUC, B-cubed and CEAFe)
[Denis and Baldridge, 2009a; Pradhan et al., 2011].
This wrapper of CoVal currently only work with CoNLL line format:
The CoNLL format has one word per line with all the annotation for this word in column separated by spaces:
Column Type Description
1 Document ID This is a variation on the document filename
2 Part number Some files are divided into multiple parts numbered as 000, 001, 002, ... etc.
3 Word number
4 Word itself This is the token as segmented/tokenized in the Treebank. Initially the *_skel file contain the placeholder [WORD] which gets replaced by the actual token from the Treebank which is part of the OntoNotes release.
5 Part-of-Speech
6 Parse bit This is the bracketed structure broken before the first open parenthesis in the parse, and the word/part-of-speech leaf replaced with a *. The full parse can be created by substituting the asterix with the "([pos] [word])" string (or leaf) and concatenating the items in the rows of that column.
7 Predicate lemma The predicate lemma is mentioned for the rows for which we have semantic role information. All other rows are marked with a "-"
8 Predicate Frameset ID This is the PropBank frameset ID of the predicate in Column 7.
9 Word sense This is the word sense of the word in Column 3.
10 Speaker/Author This is the speaker or author name where available. Mostly in Broadcast Conversation and Web Log data.
11 Named Entities These columns identifies the spans representing various named entities.
12:N Predicate Arguments There is one column each of predicate argument structure information for the predicate mentioned in Column 7.
N Coreference Coreference chain information encoded in a parenthesis structure.
More informations on the format can be found here (section "*_conll File Format"): http://www.conll.cemantix.org/2012/data.html
Details on the evaluation on CoNLL can be found here: https://github.com/ns-moosavi/coval/blob/master/conll/README.md
CoVal code was written by @ns-moosavi.
Some parts are borrowed from https://github.com/clarkkev/deep-coref/blob/master/evaluation.py
The test suite is taken from https://github.com/conll/reference-coreference-scorers/
Mention evaluation and the test suite are added by @andreasvc.
Parsing CoNLL files is developed by Leo Born.
---
## Metric description
CoVal is a coreference evaluation tool for the [CoNLL](https://huggingface.co/datasets/conll2003) and [ARRAU](https://catalog.ldc.upenn.edu/LDC2013T22) datasets which implements of the common evaluation metrics including MUC [Vilain et al, 1995](https://aclanthology.org/M95-1005.pdf), B-cubed [Bagga and Baldwin, 1998](https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.34.2578&rep=rep1&type=pdf), CEAFe [Luo et al., 2005](https://aclanthology.org/H05-1004.pdf), LEA [Moosavi and Strube, 2016](https://aclanthology.org/P16-1060.pdf) and the averaged CoNLL score (the average of the F1 values of MUC, B-cubed and CEAFe).
CoVal code was written by [`@ns-moosavi`](https://github.com/ns-moosavi), with some parts borrowed from [Deep Coref](https://github.com/clarkkev/deep-coref/blob/master/evaluation.py). The test suite is taken from the [official CoNLL code](https://github.com/conll/reference-coreference-scorers/), with additions by [`@andreasvc`](https://github.com/andreasvc) and file parsing developed by Leo Born.
## How to use
The metric takes two lists of sentences as input: one representing `predictions` and `references`, with the sentences consisting of words in the CoNLL format (see the [Limitations and bias](#Limitations-and-bias) section below for more details on the CoNLL format).
```python
from evaluate import load
coval = load('coval')
words = ['bc/cctv/00/cctv_0005 0 0 Thank VBP (TOP(S(VP* thank 01 1 Xu_li * (V*) * -',
... 'bc/cctv/00/cctv_0005 0 1 you PRP (NP*) - - - Xu_li * (ARG1*) (ARG0*) (116)',
... 'bc/cctv/00/cctv_0005 0 2 everyone NN (NP*) - - - Xu_li * (ARGM-DIS*) * (116)',
... 'bc/cctv/00/cctv_0005 0 3 for IN (PP* - - - Xu_li * (ARG2* * -',
... 'bc/cctv/00/cctv_0005 0 4 watching VBG (S(VP*)))) watch 01 1 Xu_li * *) (V*) -',
... 'bc/cctv/00/cctv_0005 0 5 . . *)) - - - Xu_li * * * -']
references = [words]
predictions = [words]
results = coval.compute(predictions=predictions, references=references)
```
It also has several optional arguments:
`keep_singletons`: After extracting all mentions of key or system file mentions whose corresponding coreference chain is of size one are considered as singletons. The default evaluation mode will include singletons in evaluations if they are included in the key or the system files. By setting `keep_singletons=False`, all singletons in the key and system files will be excluded from the evaluation.
`NP_only`: Most of the recent coreference resolvers only resolve NP mentions and leave out the resolution of VPs. By setting the `NP_only` option, the scorer will only evaluate the resolution of NPs.
`min_span`: By setting `min_span`, the scorer reports the results based on automatically detected minimum spans. Minimum spans are determined using the [MINA algorithm](https://arxiv.org/pdf/1906.06703.pdf).
## Output values
The metric outputs a dictionary with the following key-value pairs:
`mentions`: number of mentions, ranges from 0-1
`muc`: MUC metric, which expresses performance in terms of recall and precision, ranging from 0-1.
`bcub`: B-cubed metric, which is the averaged precision of all items in the distribution, ranging from 0-1.
`ceafe`: CEAFe (Constrained Entity Alignment F-Measure) is computed by aligning reference and system entities with the constraint that a reference entity is aligned with at most one reference entity. It ranges from 0-1
`lea`: LEA is a Link-Based Entity-Aware metric which, for each entity, considers how important the entity is and how well it is resolved. It ranges from 0-1.
`conll_score`: averaged CoNLL score (the average of the F1 values of `muc`, `bcub` and `ceafe`), ranging from 0 to 100.
### Values from popular papers
Given that many of the metrics returned by COVAL come from different sources, is it hard to cite reference values for all of them.
The CoNLL score is used to track progress on different datasets such as the [ARRAU corpus](https://paperswithcode.com/sota/coreference-resolution-on-the-arrau-corpus) and [CoNLL 2012](https://paperswithcode.com/sota/coreference-resolution-on-conll-2012).
## Examples
Maximal values
```python
from evaluate import load
coval = load('coval')
words = ['bc/cctv/00/cctv_0005 0 0 Thank VBP (TOP(S(VP* thank 01 1 Xu_li * (V*) * -',
... 'bc/cctv/00/cctv_0005 0 1 you PRP (NP*) - - - Xu_li * (ARG1*) (ARG0*) (116)',
... 'bc/cctv/00/cctv_0005 0 2 everyone NN (NP*) - - - Xu_li * (ARGM-DIS*) * (116)',
... 'bc/cctv/00/cctv_0005 0 3 for IN (PP* - - - Xu_li * (ARG2* * -',
... 'bc/cctv/00/cctv_0005 0 4 watching VBG (S(VP*)))) watch 01 1 Xu_li * *) (V*) -',
... 'bc/cctv/00/cctv_0005 0 5 . . *)) - - - Xu_li * * * -']
references = [words]
predictions = [words]
results = coval.compute(predictions=predictions, references=references)
print(results)
{'mentions/recall': 1.0, 'mentions/precision': 1.0, 'mentions/f1': 1.0, 'muc/recall': 1.0, 'muc/precision': 1.0, 'muc/f1': 1.0, 'bcub/recall': 1.0, 'bcub/precision': 1.0, 'bcub/f1': 1.0, 'ceafe/recall': 1.0, 'ceafe/precision': 1.0, 'ceafe/f1': 1.0, 'lea/recall': 1.0, 'lea/precision': 1.0, 'lea/f1': 1.0, 'conll_score': 100.0}
```
## Limitations and bias
This wrapper of CoVal currently only works with [CoNLL line format](https://huggingface.co/datasets/conll2003), which has one word per line with all the annotation for this word in column separated by spaces:
| Column | Type | Description |
|:-------|:----------------------|:--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| 1 | Document ID | This is a variation on the document filename |
| 2 | Part number | Some files are divided into multiple parts numbered as 000, 001, 002, ... etc. |
| 3 | Word number | |
| 4 | Word | This is the token as segmented/tokenized in the Treebank. Initially the *_skel file contain the placeholder [WORD] which gets replaced by the actual token from the Treebank which is part of the OntoNotes release. |
| 5 | Part-of-Speech | |
| 6 | Parse bit | This is the bracketed structure broken before the first open parenthesis in the parse, and the word/part-of-speech leaf replaced with a *. The full parse can be created by substituting the asterix with the "([pos] [word])" string (or leaf) and concatenating the items in the rows of that column. |
| 7 | Predicate lemma | The predicate lemma is mentioned for the rows for which we have semantic role information. All other rows are marked with a "-". |
| 8 | Predicate Frameset ID | This is the PropBank frameset ID of the predicate in Column 7. |
| 9 | Word sense | This is the word sense of the word in Column 3. |
| 10 | Speaker/Author | This is the speaker or author name where available. Mostly in Broadcast Conversation and Web Log data. |
| 11 | Named Entities | These columns identifies the spans representing various named entities. |
| 12:N | Predicate Arguments | There is one column each of predicate argument structure information for the predicate mentioned in Column 7. |
| N | Coreference | Coreference chain information encoded in a parenthesis structure. |
## Citations
```bibtex
@InProceedings{moosavi2019minimum,
author = { Nafise Sadat Moosavi, Leo Born, Massimo Poesio and Michael Strube},
title = {Using Automatically Extracted Minimum Spans to Disentangle Coreference Evaluation from Boundary Detection},
year = {2019},
booktitle = {Proceedings of the 57th Annual Meeting of
the Association for Computational Linguistics (Volume 1: Long Papers)},
publisher = {Association for Computational Linguistics},
address = {Florence, Italy},
}
```
```bibtex
@inproceedings{10.3115/1072399.1072405,
author = {Vilain, Marc and Burger, John and Aberdeen, John and Connolly, Dennis and Hirschman, Lynette},
title = {A Model-Theoretic Coreference Scoring Scheme},
year = {1995},
isbn = {1558604022},
publisher = {Association for Computational Linguistics},
address = {USA},
url = {https://doi.org/10.3115/1072399.1072405},
doi = {10.3115/1072399.1072405},
booktitle = {Proceedings of the 6th Conference on Message Understanding},
pages = {45–52},
numpages = {8},
location = {Columbia, Maryland},
series = {MUC6 ’95}
}
```
```bibtex
@INPROCEEDINGS{Bagga98algorithmsfor,
author = {Amit Bagga and Breck Baldwin},
title = {Algorithms for Scoring Coreference Chains},
booktitle = {In The First International Conference on Language Resources and Evaluation Workshop on Linguistics Coreference},
year = {1998},
pages = {563--566}
}
```
```bibtex
@INPROCEEDINGS{Luo05oncoreference,
author = {Xiaoqiang Luo},
title = {On coreference resolution performance metrics},
booktitle = {In Proc. of HLT/EMNLP},
year = {2005},
pages = {25--32},
publisher = {URL}
}
```
```bibtex
@inproceedings{moosavi-strube-2016-coreference,
title = "Which Coreference Evaluation Metric Do You Trust? A Proposal for a Link-based Entity Aware Metric",
author = "Moosavi, Nafise Sadat and
Strube, Michael",
booktitle = "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = aug,
year = "2016",
address = "Berlin, Germany",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/P16-1060",
doi = "10.18653/v1/P16-1060",
pages = "632--642",
}
```
## Further References
- [CoNLL 2012 Task Description](http://www.conll.cemantix.org/2012/data.html): for information on the format (section "*_conll File Format")
- [CoNLL Evaluation details](https://github.com/ns-moosavi/coval/blob/master/conll/README.md)
- [Hugging Face - Neural Coreference Resolution (Neuralcoref)](https://huggingface.co/coref/)
evaluate-0.4.2/metrics/coval/app.py 0 → 100644
View file @ 25991f98
import sys
import evaluate
from evaluate.utils import launch_gradio_widget
sys.path = [p for p in sys.path if p != "/home/user/app"]
module = evaluate.load("coval")
sys.path = ["/home/user/app"] + sys.path
launch_gradio_widget(module)
evaluate-0.4.2/metrics/coval/coval.py 0 → 100644
View file @ 25991f98
# Copyright 2020 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.
""" CoVal metric. """
import coval # From: git+https://github.com/ns-moosavi/coval.git noqa: F401
import datasets
from coval.conll import reader, util
from coval.eval import evaluator
import evaluate
logger = evaluate.logging.get_logger(__name__)
_CITATION = """\
@InProceedings{moosavi2019minimum,
author = { Nafise Sadat Moosavi, Leo Born, Massimo Poesio and Michael Strube},
title = {Using Automatically Extracted Minimum Spans to Disentangle Coreference Evaluation from Boundary Detection},
year = {2019},
booktitle = {Proceedings of the 57th Annual Meeting of
the Association for Computational Linguistics (Volume 1: Long Papers)},
publisher = {Association for Computational Linguistics},
address = {Florence, Italy},
}
@inproceedings{10.3115/1072399.1072405,
author = {Vilain, Marc and Burger, John and Aberdeen, John and Connolly, Dennis and Hirschman, Lynette},
title = {A Model-Theoretic Coreference Scoring Scheme},
year = {1995},
isbn = {1558604022},
publisher = {Association for Computational Linguistics},
address = {USA},
url = {https://doi.org/10.3115/1072399.1072405},
doi = {10.3115/1072399.1072405},
booktitle = {Proceedings of the 6th Conference on Message Understanding},
pages = {45–52},
numpages = {8},
location = {Columbia, Maryland},
series = {MUC6 ’95}
}
@INPROCEEDINGS{Bagga98algorithmsfor,
author = {Amit Bagga and Breck Baldwin},
title = {Algorithms for Scoring Coreference Chains},
booktitle = {In The First International Conference on Language Resources and Evaluation Workshop on Linguistics Coreference},
year = {1998},
pages = {563--566}
}
@INPROCEEDINGS{Luo05oncoreference,
author = {Xiaoqiang Luo},
title = {On coreference resolution performance metrics},
booktitle = {In Proc. of HLT/EMNLP},
year = {2005},
pages = {25--32},
publisher = {URL}
}
@inproceedings{moosavi-strube-2016-coreference,
title = "Which Coreference Evaluation Metric Do You Trust? A Proposal for a Link-based Entity Aware Metric",
author = "Moosavi, Nafise Sadat and
Strube, Michael",
booktitle = "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = aug,
year = "2016",
address = "Berlin, Germany",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/P16-1060",
doi = "10.18653/v1/P16-1060",
pages = "632--642",
}
"""
_DESCRIPTION = """\
CoVal is a coreference evaluation tool for the CoNLL and ARRAU datasets which
implements of the common evaluation metrics including MUC [Vilain et al, 1995],
B-cubed [Bagga and Baldwin, 1998], CEAFe [Luo et al., 2005],
LEA [Moosavi and Strube, 2016] and the averaged CoNLL score
(the average of the F1 values of MUC, B-cubed and CEAFe)
[Denis and Baldridge, 2009a; Pradhan et al., 2011].
This wrapper of CoVal currently only work with CoNLL line format:
The CoNLL format has one word per line with all the annotation for this word in column separated by spaces:
Column Type Description
1 Document ID This is a variation on the document filename
2 Part number Some files are divided into multiple parts numbered as 000, 001, 002, ... etc.
3 Word number
4 Word itself This is the token as segmented/tokenized in the Treebank. Initially the *_skel file contain the placeholder [WORD] which gets replaced by the actual token from the Treebank which is part of the OntoNotes release.
5 Part-of-Speech
6 Parse bit This is the bracketed structure broken before the first open parenthesis in the parse, and the word/part-of-speech leaf replaced with a *. The full parse can be created by substituting the asterix with the "([pos] [word])" string (or leaf) and concatenating the items in the rows of that column.
7 Predicate lemma The predicate lemma is mentioned for the rows for which we have semantic role information. All other rows are marked with a "-"
8 Predicate Frameset ID This is the PropBank frameset ID of the predicate in Column 7.
9 Word sense This is the word sense of the word in Column 3.
10 Speaker/Author This is the speaker or author name where available. Mostly in Broadcast Conversation and Web Log data.
11 Named Entities These columns identifies the spans representing various named entities.
12:N Predicate Arguments There is one column each of predicate argument structure information for the predicate mentioned in Column 7.
N Coreference Coreference chain information encoded in a parenthesis structure.
More informations on the format can be found here (section "*_conll File Format"): http://www.conll.cemantix.org/2012/data.html
Details on the evaluation on CoNLL can be found here: https://github.com/ns-moosavi/coval/blob/master/conll/README.md
CoVal code was written by @ns-moosavi.
Some parts are borrowed from https://github.com/clarkkev/deep-coref/blob/master/evaluation.py
The test suite is taken from https://github.com/conll/reference-coreference-scorers/
Mention evaluation and the test suite are added by @andreasvc.
Parsing CoNLL files is developed by Leo Born.
"""
_KWARGS_DESCRIPTION = """
Calculates coreference evaluation metrics.
Args:
predictions: list of sentences. Each sentence is a list of word predictions to score in the CoNLL format.
Each prediction is a word with its annotations as a string made of columns joined with spaces.
Only columns 4, 5, 6 and the last column are used (word, POS, Pars and coreference annotation)
See the details on the format in the description of the metric.
references: list of sentences. Each sentence is a list of word reference to score in the CoNLL format.
Each reference is a word with its annotations as a string made of columns joined with spaces.
Only columns 4, 5, 6 and the last column are used (word, POS, Pars and coreference annotation)
See the details on the format in the description of the metric.
keep_singletons: After extracting all mentions of key or system files,
mentions whose corresponding coreference chain is of size one,
are considered as singletons. The default evaluation mode will include
singletons in evaluations if they are included in the key or the system files.
By setting 'keep_singletons=False', all singletons in the key and system files
will be excluded from the evaluation.
NP_only: Most of the recent coreference resolvers only resolve NP mentions and
leave out the resolution of VPs. By setting the 'NP_only' option, the scorer will only evaluate the resolution of NPs.
min_span: By setting 'min_span', the scorer reports the results based on automatically detected minimum spans.
Minimum spans are determined using the MINA algorithm.
Returns:
'mentions': mentions
'muc': MUC metric [Vilain et al, 1995]
'bcub': B-cubed [Bagga and Baldwin, 1998]
'ceafe': CEAFe [Luo et al., 2005]
'lea': LEA [Moosavi and Strube, 2016]
'conll_score': averaged CoNLL score (the average of the F1 values of MUC, B-cubed and CEAFe)
Examples:
>>> coval = evaluate.load('coval')
>>> words = ['bc/cctv/00/cctv_0005 0 0 Thank VBP (TOP(S(VP* thank 01 1 Xu_li * (V*) * -',
... 'bc/cctv/00/cctv_0005 0 1 you PRP (NP*) - - - Xu_li * (ARG1*) (ARG0*) (116)',
... 'bc/cctv/00/cctv_0005 0 2 everyone NN (NP*) - - - Xu_li * (ARGM-DIS*) * (116)',
... 'bc/cctv/00/cctv_0005 0 3 for IN (PP* - - - Xu_li * (ARG2* * -',
... 'bc/cctv/00/cctv_0005 0 4 watching VBG (S(VP*)))) watch 01 1 Xu_li * *) (V*) -',
... 'bc/cctv/00/cctv_0005 0 5 . . *)) - - - Xu_li * * * -']
>>> references = [words]
>>> predictions = [words]
>>> results = coval.compute(predictions=predictions, references=references)
>>> print(results) # doctest:+ELLIPSIS
{'mentions/recall': 1.0,[...] 'conll_score': 100.0}
"""
def get_coref_infos(
key_lines, sys_lines, NP_only=False, remove_nested=False, keep_singletons=True, min_span=False, doc="dummy_doc"
):
key_doc_lines = {doc: key_lines}
sys_doc_lines = {doc: sys_lines}
doc_coref_infos = {}
key_nested_coref_num = 0
sys_nested_coref_num = 0
key_removed_nested_clusters = 0
sys_removed_nested_clusters = 0
key_singletons_num = 0
sys_singletons_num = 0
key_clusters, singletons_num = reader.get_doc_mentions(doc, key_doc_lines[doc], keep_singletons)
key_singletons_num += singletons_num
if NP_only or min_span:
key_clusters = reader.set_annotated_parse_trees(key_clusters, key_doc_lines[doc], NP_only, min_span)
sys_clusters, singletons_num = reader.get_doc_mentions(doc, sys_doc_lines[doc], keep_singletons)
sys_singletons_num += singletons_num
if NP_only or min_span:
sys_clusters = reader.set_annotated_parse_trees(sys_clusters, key_doc_lines[doc], NP_only, min_span)
if remove_nested:
nested_mentions, removed_clusters = reader.remove_nested_coref_mentions(key_clusters, keep_singletons)
key_nested_coref_num += nested_mentions
key_removed_nested_clusters += removed_clusters
nested_mentions, removed_clusters = reader.remove_nested_coref_mentions(sys_clusters, keep_singletons)
sys_nested_coref_num += nested_mentions
sys_removed_nested_clusters += removed_clusters
sys_mention_key_cluster = reader.get_mention_assignments(sys_clusters, key_clusters)
key_mention_sys_cluster = reader.get_mention_assignments(key_clusters, sys_clusters)
doc_coref_infos[doc] = (key_clusters, sys_clusters, key_mention_sys_cluster, sys_mention_key_cluster)
if remove_nested:
logger.info(
"Number of removed nested coreferring mentions in the key "
f"annotation: {key_nested_coref_num}; and system annotation: {sys_nested_coref_num}"
)
logger.info(
"Number of resulting singleton clusters in the key "
f"annotation: {key_removed_nested_clusters}; and system annotation: {sys_removed_nested_clusters}"
)
if not keep_singletons:
logger.info(
f"{key_singletons_num:d} and {sys_singletons_num:d} singletons are removed from the key and system "
"files, respectively"
)
return doc_coref_infos
def compute_score(key_lines, sys_lines, metrics, NP_only, remove_nested, keep_singletons, min_span):
doc_coref_infos = get_coref_infos(key_lines, sys_lines, NP_only, remove_nested, keep_singletons, min_span)
output_scores = {}
conll = 0
conll_subparts_num = 0
for name, metric in metrics:
recall, precision, f1 = evaluator.evaluate_documents(doc_coref_infos, metric, beta=1)
if name in ["muc", "bcub", "ceafe"]:
conll += f1
conll_subparts_num += 1
output_scores.update({f"{name}/recall": recall, f"{name}/precision": precision, f"{name}/f1": f1})
logger.info(
name.ljust(10),
f"Recall: {recall * 100:.2f}",
f" Precision: {precision * 100:.2f}",
f" F1: {f1 * 100:.2f}",
)
if conll_subparts_num == 3:
conll = (conll / 3) * 100
logger.info(f"CoNLL score: {conll:.2f}")
output_scores.update({"conll_score": conll})
return output_scores
def check_gold_parse_annotation(key_lines):
has_gold_parse = False
for line in key_lines:
if not line.startswith("#"):
if len(line.split()) > 6:
parse_col = line.split()[5]
if not parse_col == "-":
has_gold_parse = True
break
else:
break
return has_gold_parse
@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class Coval(evaluate.Metric):
def _info(self):
return evaluate.MetricInfo(
description=_DESCRIPTION,
citation=_CITATION,
inputs_description=_KWARGS_DESCRIPTION,
features=datasets.Features(
{
"predictions": datasets.Sequence(datasets.Value("string")),
"references": datasets.Sequence(datasets.Value("string")),
}
),
codebase_urls=["https://github.com/ns-moosavi/coval"],
reference_urls=[
"https://github.com/ns-moosavi/coval",
"https://www.aclweb.org/anthology/P16-1060",
"http://www.conll.cemantix.org/2012/data.html",
],
)
def _compute(
self, predictions, references, keep_singletons=True, NP_only=False, min_span=False, remove_nested=False
):
allmetrics = [
("mentions", evaluator.mentions),
("muc", evaluator.muc),
("bcub", evaluator.b_cubed),
("ceafe", evaluator.ceafe),
("lea", evaluator.lea),
]
if min_span:
has_gold_parse = util.check_gold_parse_annotation(references)
if not has_gold_parse:
raise NotImplementedError("References should have gold parse annotation to use 'min_span'.")
# util.parse_key_file(key_file)
# key_file = key_file + ".parsed"
score = compute_score(
key_lines=references,
sys_lines=predictions,
metrics=allmetrics,
NP_only=NP_only,
remove_nested=remove_nested,
keep_singletons=keep_singletons,
min_span=min_span,
)
return score
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment