Create preprocess_marc_ja.py
Browse files- preprocess_marc_ja.py +261 -0
preprocess_marc_ja.py
ADDED
|
@@ -0,0 +1,261 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
This code is licensed under CC-BY-4.0 from the original work by shunk031.
|
| 3 |
+
The code is adapted from https://huggingface.co/datasets/shunk031/JGLUE/blob/main/JGLUE.py
|
| 4 |
+
with minor modifications to the code structure.
|
| 5 |
+
This codebase provides pre-processing functionality for the MARC-ja dataset in the Japanese GLUE benchmark.
|
| 6 |
+
The original code can be found at https://github.com/yahoojapan/JGLUE/blob/main/preprocess/marc-ja/scripts/marc-ja.py.
|
| 7 |
+
"""
|
| 8 |
+
|
| 9 |
+
import random
|
| 10 |
+
import warnings
|
| 11 |
+
from typing import Dict, List, Optional, Union
|
| 12 |
+
import string
|
| 13 |
+
|
| 14 |
+
import datasets as ds
|
| 15 |
+
import pandas as pd
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
class MarcJaConfig(ds.BuilderConfig):
|
| 19 |
+
def __init__(
|
| 20 |
+
self,
|
| 21 |
+
name: str = "MARC-ja",
|
| 22 |
+
is_han_to_zen: bool = False,
|
| 23 |
+
max_instance_num: Optional[int] = None,
|
| 24 |
+
max_char_length: int = 500,
|
| 25 |
+
remove_netural: bool = True,
|
| 26 |
+
train_ratio: float = 0.94,
|
| 27 |
+
val_ratio: float = 0.03,
|
| 28 |
+
test_ratio: float = 0.03,
|
| 29 |
+
output_testset: bool = False,
|
| 30 |
+
filter_review_id_list_valid: bool = True,
|
| 31 |
+
label_conv_review_id_list_valid: bool = True,
|
| 32 |
+
version: Optional[Union[ds.utils.Version, str]] = ds.utils.Version("0.0.0"),
|
| 33 |
+
data_dir: Optional[str] = None,
|
| 34 |
+
data_files: Optional[ds.data_files.DataFilesDict] = None,
|
| 35 |
+
description: Optional[str] = None,
|
| 36 |
+
) -> None:
|
| 37 |
+
super().__init__(
|
| 38 |
+
name=name,
|
| 39 |
+
version=version,
|
| 40 |
+
data_dir=data_dir,
|
| 41 |
+
data_files=data_files,
|
| 42 |
+
description=description,
|
| 43 |
+
)
|
| 44 |
+
if train_ratio + val_ratio + test_ratio != 1.0:
|
| 45 |
+
raise ValueError(
|
| 46 |
+
"train_ratio + val_ratio + test_ratio should be 1.0, "
|
| 47 |
+
f"but got {train_ratio} + {val_ratio} + {test_ratio} = {train_ratio + val_ratio + test_ratio}"
|
| 48 |
+
)
|
| 49 |
+
|
| 50 |
+
self.train_ratio = train_ratio
|
| 51 |
+
self.val_ratio = val_ratio
|
| 52 |
+
self.test_ratio = test_ratio
|
| 53 |
+
|
| 54 |
+
self.is_han_to_zen = is_han_to_zen
|
| 55 |
+
self.max_instance_num = max_instance_num
|
| 56 |
+
self.max_char_length = max_char_length
|
| 57 |
+
self.remove_netural = remove_netural
|
| 58 |
+
self.output_testset = output_testset
|
| 59 |
+
|
| 60 |
+
self.filter_review_id_list_valid = filter_review_id_list_valid
|
| 61 |
+
self.label_conv_review_id_list_valid = label_conv_review_id_list_valid
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
def get_label(rating: int, remove_netural: bool = False) -> Optional[str]:
|
| 65 |
+
if rating >= 4:
|
| 66 |
+
return "positive"
|
| 67 |
+
elif rating <= 2:
|
| 68 |
+
return "negative"
|
| 69 |
+
else:
|
| 70 |
+
if remove_netural:
|
| 71 |
+
return None
|
| 72 |
+
else:
|
| 73 |
+
return "neutral"
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
def is_filtered_by_ascii_rate(text: str, threshold: float = 0.9) -> bool:
|
| 77 |
+
ascii_letters = set(string.printable)
|
| 78 |
+
rate = sum(c in ascii_letters for c in text) / len(text)
|
| 79 |
+
return rate >= threshold
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
def shuffle_dataframe(df: pd.DataFrame) -> pd.DataFrame:
|
| 83 |
+
instances = df.to_dict(orient="records")
|
| 84 |
+
random.seed(1)
|
| 85 |
+
random.shuffle(instances)
|
| 86 |
+
return pd.DataFrame(instances)
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
def get_filter_review_id_list(
|
| 90 |
+
filter_review_id_list_paths: Dict[str, str],
|
| 91 |
+
) -> Dict[str, List[str]]:
|
| 92 |
+
filter_review_id_list_valid = filter_review_id_list_paths.get("valid")
|
| 93 |
+
filter_review_id_list_test = filter_review_id_list_paths.get("test")
|
| 94 |
+
|
| 95 |
+
filter_review_id_list = {}
|
| 96 |
+
|
| 97 |
+
if filter_review_id_list_valid is not None:
|
| 98 |
+
with open(filter_review_id_list_valid, "r") as rf:
|
| 99 |
+
filter_review_id_list["valid"] = [line.rstrip() for line in rf]
|
| 100 |
+
|
| 101 |
+
if filter_review_id_list_test is not None:
|
| 102 |
+
with open(filter_review_id_list_test, "r") as rf:
|
| 103 |
+
filter_review_id_list["test"] = [line.rstrip() for line in rf]
|
| 104 |
+
|
| 105 |
+
return filter_review_id_list
|
| 106 |
+
|
| 107 |
+
|
| 108 |
+
def get_label_conv_review_id_list(
|
| 109 |
+
label_conv_review_id_list_paths: Dict[str, str],
|
| 110 |
+
) -> Dict[str, Dict[str, str]]:
|
| 111 |
+
import csv
|
| 112 |
+
|
| 113 |
+
label_conv_review_id_list_valid = label_conv_review_id_list_paths.get("valid")
|
| 114 |
+
label_conv_review_id_list_test = label_conv_review_id_list_paths.get("test")
|
| 115 |
+
|
| 116 |
+
label_conv_review_id_list: Dict[str, Dict[str, str]] = {}
|
| 117 |
+
|
| 118 |
+
if label_conv_review_id_list_valid is not None:
|
| 119 |
+
with open(label_conv_review_id_list_valid, "r", encoding="utf-8") as rf:
|
| 120 |
+
label_conv_review_id_list["valid"] = {row[0]: row[1] for row in csv.reader(rf)}
|
| 121 |
+
|
| 122 |
+
if label_conv_review_id_list_test is not None:
|
| 123 |
+
with open(label_conv_review_id_list_test, "r", encoding="utf-8") as rf:
|
| 124 |
+
label_conv_review_id_list["test"] = {row[0]: row[1] for row in csv.reader(rf)}
|
| 125 |
+
|
| 126 |
+
return label_conv_review_id_list
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
def output_data(
|
| 130 |
+
df: pd.DataFrame,
|
| 131 |
+
train_ratio: float,
|
| 132 |
+
val_ratio: float,
|
| 133 |
+
test_ratio: float,
|
| 134 |
+
output_testset: bool,
|
| 135 |
+
filter_review_id_list_paths: Dict[str, str],
|
| 136 |
+
label_conv_review_id_list_paths: Dict[str, str],
|
| 137 |
+
) -> Dict[str, pd.DataFrame]:
|
| 138 |
+
instance_num = len(df)
|
| 139 |
+
split_dfs: Dict[str, pd.DataFrame] = {}
|
| 140 |
+
length1 = int(instance_num * train_ratio)
|
| 141 |
+
split_dfs["train"] = df.iloc[:length1]
|
| 142 |
+
|
| 143 |
+
length2 = int(instance_num * (train_ratio + val_ratio))
|
| 144 |
+
split_dfs["valid"] = df.iloc[length1:length2]
|
| 145 |
+
split_dfs["test"] = df.iloc[length2:]
|
| 146 |
+
|
| 147 |
+
filter_review_id_list = get_filter_review_id_list(
|
| 148 |
+
filter_review_id_list_paths=filter_review_id_list_paths,
|
| 149 |
+
)
|
| 150 |
+
label_conv_review_id_list = get_label_conv_review_id_list(
|
| 151 |
+
label_conv_review_id_list_paths=label_conv_review_id_list_paths,
|
| 152 |
+
)
|
| 153 |
+
|
| 154 |
+
for eval_type in ("valid", "test"):
|
| 155 |
+
if filter_review_id_list.get(eval_type):
|
| 156 |
+
df = split_dfs[eval_type]
|
| 157 |
+
df = df[~df["review_id"].isin(filter_review_id_list[eval_type])]
|
| 158 |
+
split_dfs[eval_type] = df
|
| 159 |
+
|
| 160 |
+
for eval_type in ("valid", "test"):
|
| 161 |
+
if label_conv_review_id_list.get(eval_type):
|
| 162 |
+
df = split_dfs[eval_type]
|
| 163 |
+
df = df.assign(converted_label=df["review_id"].map(label_conv_review_id_list["valid"]))
|
| 164 |
+
df = df.assign(
|
| 165 |
+
label=df[["label", "converted_label"]].apply(
|
| 166 |
+
lambda xs: xs["label"] if pd.isnull(xs["converted_label"]) else xs["converted_label"],
|
| 167 |
+
axis=1,
|
| 168 |
+
)
|
| 169 |
+
)
|
| 170 |
+
df = df.drop(columns=["converted_label"])
|
| 171 |
+
split_dfs[eval_type] = df
|
| 172 |
+
|
| 173 |
+
return {
|
| 174 |
+
"train": split_dfs["train"],
|
| 175 |
+
"valid": split_dfs["valid"],
|
| 176 |
+
}
|
| 177 |
+
|
| 178 |
+
|
| 179 |
+
def preprocess_marc_ja(
|
| 180 |
+
config: MarcJaConfig,
|
| 181 |
+
data_file_path: str,
|
| 182 |
+
filter_review_id_list_paths: Dict[str, str],
|
| 183 |
+
label_conv_review_id_list_paths: Dict[str, str],
|
| 184 |
+
) -> Dict[str, pd.DataFrame]:
|
| 185 |
+
try:
|
| 186 |
+
import mojimoji
|
| 187 |
+
|
| 188 |
+
def han_to_zen(text: str) -> str:
|
| 189 |
+
return mojimoji.han_to_zen(text)
|
| 190 |
+
|
| 191 |
+
except ImportError:
|
| 192 |
+
warnings.warn(
|
| 193 |
+
"can't import `mojimoji`, failing back to method that do nothing. "
|
| 194 |
+
"We recommend running `pip install mojimoji` to reproduce the original preprocessing.",
|
| 195 |
+
UserWarning,
|
| 196 |
+
)
|
| 197 |
+
|
| 198 |
+
def han_to_zen(text: str) -> str:
|
| 199 |
+
return text
|
| 200 |
+
|
| 201 |
+
try:
|
| 202 |
+
from bs4 import BeautifulSoup
|
| 203 |
+
|
| 204 |
+
def cleanup_text(text: str) -> str:
|
| 205 |
+
return BeautifulSoup(text, "html.parser").get_text()
|
| 206 |
+
|
| 207 |
+
except ImportError:
|
| 208 |
+
warnings.warn(
|
| 209 |
+
"can't import `beautifulsoup4`, failing back to method that do nothing."
|
| 210 |
+
"We recommend running `pip install beautifulsoup4` to reproduce the original preprocessing.",
|
| 211 |
+
UserWarning,
|
| 212 |
+
)
|
| 213 |
+
|
| 214 |
+
def cleanup_text(text: str) -> str:
|
| 215 |
+
return text
|
| 216 |
+
|
| 217 |
+
from tqdm import tqdm
|
| 218 |
+
|
| 219 |
+
df = pd.read_csv(data_file_path, delimiter="\t")
|
| 220 |
+
df = df[["review_body", "star_rating", "review_id"]]
|
| 221 |
+
|
| 222 |
+
# rename columns
|
| 223 |
+
df = df.rename(columns={"review_body": "text", "star_rating": "rating"})
|
| 224 |
+
|
| 225 |
+
# convert the rating to label
|
| 226 |
+
tqdm.pandas(dynamic_ncols=True, desc="Convert the rating to the label")
|
| 227 |
+
df = df.assign(label=df["rating"].progress_apply(lambda rating: get_label(rating, config.remove_netural)))
|
| 228 |
+
|
| 229 |
+
# remove rows where the label is None
|
| 230 |
+
df = df[~df["label"].isnull()]
|
| 231 |
+
|
| 232 |
+
# remove html tags from the text
|
| 233 |
+
tqdm.pandas(dynamic_ncols=True, desc="Remove html tags from the text")
|
| 234 |
+
df = df.assign(text=df["text"].progress_apply(cleanup_text))
|
| 235 |
+
|
| 236 |
+
# filter by ascii rate
|
| 237 |
+
tqdm.pandas(dynamic_ncols=True, desc="Filter by ascii rate")
|
| 238 |
+
df = df[~df["text"].progress_apply(is_filtered_by_ascii_rate)]
|
| 239 |
+
|
| 240 |
+
if config.max_char_length is not None:
|
| 241 |
+
df = df[df["text"].str.len() <= config.max_char_length]
|
| 242 |
+
|
| 243 |
+
if config.is_han_to_zen:
|
| 244 |
+
df = df.assign(text=df["text"].apply(han_to_zen))
|
| 245 |
+
|
| 246 |
+
df = df[["text", "label", "review_id"]]
|
| 247 |
+
df = df.rename(columns={"text": "sentence"})
|
| 248 |
+
|
| 249 |
+
# shuffle dataset
|
| 250 |
+
df = shuffle_dataframe(df)
|
| 251 |
+
|
| 252 |
+
split_dfs = output_data(
|
| 253 |
+
df=df,
|
| 254 |
+
train_ratio=config.train_ratio,
|
| 255 |
+
val_ratio=config.val_ratio,
|
| 256 |
+
test_ratio=config.test_ratio,
|
| 257 |
+
output_testset=config.output_testset,
|
| 258 |
+
filter_review_id_list_paths=filter_review_id_list_paths,
|
| 259 |
+
label_conv_review_id_list_paths=label_conv_review_id_list_paths,
|
| 260 |
+
)
|
| 261 |
+
return split_dfs
|