Add support for NoiseBench #3512
Conversation
| **corpusargs: The arguments propagated to :meth:'flair.datasets.ColumnCorpus.__init__'. | ||
| """ | ||
| if noise not in ["clean", "crowd", "crowdbest", "expert", "distant", "weak", "llm"]: | ||
| raise Exception("Please choose a valid version") |
There was a problem hiding this comment.
It is recommended to raise ValueError for wrong and unexpected values passed to the argument.
Also, an error message would be more helpful if it listed valid values. Something like:
if noise not in VALID_NOISE_VALUES:
raise ValueError(f"Unsupported value for noise argument. Got {noise}, expected one of {VALUE_NOISE_VALUES}!")
| if noise not in ["clean", "crowd", "crowdbest", "expert", "distant", "weak", "llm"]: | ||
| raise Exception("Please choose a valid version") | ||
|
|
||
| self._set_path(base_path) |
There was a problem hiding this comment.
I would remove this function and just set the base_path here in __init__:
if base_path:
self.base_path = Path(base_path)
else:
self.base_path = flair.cache_root / "datasets" / "noisebench"
| self.base_path = flair.cache_root / "datasets" / "noisebench" if not base_path else Path(base_path) | ||
|
|
||
| @staticmethod | ||
| def read_column_file(filename): |
There was a problem hiding this comment.
You are adding a new public method to NER_NOISEBENCH. I would avoid this. Also, you are relying on this method only in non-public methods in the class (except in generate_data_files, which I would also put as non-public.)
My suggestion is to rename the method to _read_column_file and add type annotations for filename argument.
| return all_x | ||
|
|
||
| @staticmethod | ||
| def save_to_column_file(filename, list): |
There was a problem hiding this comment.
- Rename to
_save_to_column_file - Add type annotations to arguments
- Avoid using the name of built-in modules/types/etc. for argument and variable names. Here, you are using
listas the name for the argument. Seems like a more appropriate name issentences(which is of typelist).
|
|
||
| def generate_data_files(self, filename): | ||
|
|
||
| with open(os.path.join(self.base_path, "annotations_only", "index.txt")) as index_file: |
There was a problem hiding this comment.
You can avoid this os.path.join since base_path is of type Path. Just use / to concatenate the path. Also, we should avoid relying on implicit default values, it is better to explicitly mention that we want to read the file with utf-8:
with open(self.base_path / "annotations_only" / "index.txt", "r", encoding="utf-8") as fp:
# ...
| with open(os.path.join(self.base_path, "annotations_only", "index.txt")) as index_file: | ||
| token_indices = index_file.readlines() | ||
|
|
||
| all_clean_sentences = self.read_column_file(os.path.join(self.cleanconll_base_path, "cleanconll.train")) |
There was a problem hiding this comment.
Same here, use self.cleanconll_base_path / "cleanconll.train" instead of os.path.join.
| ) | ||
|
|
||
| # copy test set | ||
| all_clean_test_sentences = self.read_column_file(os.path.join(self.cleanconll_base_path, "cleanconll.test")) |
There was a problem hiding this comment.
Same here, use self.cleanconll_base_path / "cleanconll.test" instead of os.path.join.
| new_s.append([token[0], token[4]]) | ||
| test_sentences.append(new_s) | ||
|
|
||
| self.save_to_column_file(os.path.join(self.base_path, "clean.test"), test_sentences) |
There was a problem hiding this comment.
Same here, use self.base_path / "clean.test" instead of os.path.join.
| new_s = [] | ||
| for token in s: | ||
| new_s.append([token[0], token[4]]) | ||
| test_sentences.append(new_s) |
There was a problem hiding this comment.
This can be refactored to a more idiomatic approach with list comprehensions:
for sentence in all_clean_test_sentences:
new_sentence = [[tokens[0], tokens[4]] for tokens in sentence]
test_sentences.append(new_sentence)
| for token, label in zip(clean_sentence, sentence): | ||
| label[0] = token[0] # token[0] -> text, token[1] -> BIO label | ||
| if self.SAVE_TRAINDEV_FILE: | ||
| self.save_to_column_file(os.path.join(self.base_path, f"{corpus}.traindev"), noisy_labels) |
There was a problem hiding this comment.
Same here, use self.base_path / f"{corpus}.traindev" instead of os.path.join.
|
|
||
| noisy_labels = self.read_column_file(os.path.join(self.base_path, "annotations_only", f"{corpus}.traindev")) | ||
| # print(noisy_labels) | ||
| # print(token_indices) |
| def _merge_tokens_labels(self, corpus, all_clean_sentences, token_indices): | ||
| # generate NoiseBench dataset variants, given CleanCoNLL, noisy label files and index file | ||
|
|
||
| noisy_labels = self.read_column_file(os.path.join(self.base_path, "annotations_only", f"{corpus}.traindev")) |
There was a problem hiding this comment.
Same here, use self.base_path / "annotations_only" / f"{corpus}.traindev" instead of os.path.join.
| point = [] | ||
| for line in lines: | ||
| if "\t" in line.strip(): | ||
| stripped_line = line.strip().split("\t") if "\t" in line.strip() else line.strip().split(" ") |
There was a problem hiding this comment.
Looks like stripped_line = line.strip().split("\t") if "\t" in line.strip() will always be set due to the condition if "\t" in line.strip() in line 5076. Refactor to:
if "\t" in line.strip():
stripped_line = line.strip().split("\t")
else:
stripped_line = line.strip().split(" ")
| def save_to_column_file(filename, list): | ||
| with open(filename, "w") as f: | ||
| for sentence in list: | ||
| for token in sentence: |
There was a problem hiding this comment.
Rename the iteration variable to tokens to convert the meaning that this is a container of elements and not a single token.
|
@fkdosilovic thanks a lot for your review, this is very helpful! Please note that this is still a draft PR and one major change is still planned. |
This PR adds support for NoiseBench, a subset of CoNLL-03 for English with 7 different label sets, including 6 different types of label noise.
Example usage (intialize dataset with a crowdsourced version of the NER labels):
Potential issues:
To create the dataset, we need the CleanCoNLL labels as gold labels. This required running an external script in this implementation: https://github.com/flairNLP/CleanCoNLL/blob/main/create_cleanconll_from_conll03.sh