init

configs.py

@@ -0,0 +1,89 @@
+import datasets
+from datasets import Value, Sequence
+
+PROCESS_FILES = [
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000001.parquet",
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000002.parquet",
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000003.parquet",
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000004.parquet",
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000005.parquet",
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000006.parquet",
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000007.parquet",
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000008.parquet",
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000009.parquet",
+    "https://huggingface.co/datasets/korean-corpus/namu/resolve/main/processed/namuwiki210301-processed-000010.parquet",
+]
+
+class NamuWikiConfig(datasets.BuilderConfig):
+    """BuilderConfig for NamuWiki"""
+
+    def __init__(self, features, data_url, citation, url, **kwargs):
+        """BuilderConfig for NamuWiki.
+        Args:
+
+          **kwargs: keyword arguments forwarded to super.
+        """
+        super().__init__(version=datasets.Version("1.0.2"), **kwargs)
+        self.features = features
+        self.data_url = data_url
+        self.citation = citation
+        self.url = url
+
+
+
+SUB_DATASETS = [
+        NamuWikiConfig(
+            name="raw",
+            features=datasets.Features(
+                {
+                    "title": Value(dtype='string', id=None),
+                    "text": Value(dtype='string', id=None),
+                    "contributors": Sequence(feature=Value(dtype='string', id=None), length=-1, id=None)
+                }
+            ),
+            data_url="https://www.dropbox.com/s/03b49jdrx7xi712/namuwiki210301.7z?dl=1",
+            citation="",
+            url="",
+        ),
+        NamuWikiConfig(
+            name="processed",
+            features=datasets.Features(
+                {
+                    "id": Value(dtype='string', id=None),
+                    "title": Value(dtype='string', id=None),
+                    "text": Value(dtype='string', id=None),
+                }
+            ),
+            data_url=PROCESS_FILES,
+            citation="",
+            url="",
+        ),
+]+[
+        NamuWikiConfig(
+            name=f"char-{length}",
+            features=datasets.Features(
+                {
+                    "id": Value(dtype='string', id=None),
+                    "title": Value(dtype='string', id=None),
+                    "text": Value(dtype='string', id=None),
+                }
+            ),
+            data_url=PROCESS_FILES,
+            citation="",
+            url="",
+        ) for length in [32, 64, 128, 256, 512]
+]+[
+        NamuWikiConfig(
+            name=f"word-{length}",
+            features=datasets.Features({
+                    "id": Value(dtype='string', id=None),
+                    "title": Value(dtype='string', id=None),
+                    "text": Value(dtype='string', id=None),
+
+            }),
+            data_url=PROCESS_FILES,
+            citation="",
+            url="",
+        ) for length in [32, 64, 128, 256, 512]
+]
+

namu.py

@@ -0,0 +1,94 @@
+import os
+from collections import defaultdict
+from typing import List
+
+import datasets
+from datasets import Sequence, Value, load_dataset
+
+from .process import process_text, get_structured_data
+from typing import List
+from math import ceil
+from .configs import SUB_DATASETS
+
+def processing(data, name):
+    if name == "processed":
+        data['text'] = [process_text(text) for text in data['text']]
+    elif name == "structured":
+        data['text'] = [process_text(text) for text in data['text']]
+        data['structured_text'] = [
+            get_structured_data(text, default_value={"item": [], "content": []}) for text in data['text']
+        ]
+    return data
+
+
+def sliding(texts: List[str], window_size: int=5, stride:int=3) -> List[str]:
+    n_iter = ceil((len(texts)-window_size)/stride)+1
+    return [texts[i*stride:i*stride+window_size] for i in range(n_iter)]
+
+class NamuWiki(datasets.GeneratorBasedBuilder):
+    BUILDER_CONFIGS = SUB_DATASETS
+
+    def _info(self):
+        return datasets.DatasetInfo(
+            description="",
+            features=self.config.features,
+            homepage=self.config.url,
+            citation=self.config.citation + "\n" + "",
+        )
+
+    def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
+        if self.config.name == "processed":
+            data_file = dl_manager.download(self.config.data_url)
+            return [
+                datasets.SplitGenerator(
+                    name=datasets.Split.TRAIN,
+                    gen_kwargs={
+                        "data_file": data_file,
+                        "split": "train"
+                    }
+                ),
+            ]
+
+        elif self.config.name.startswith(("char", "word")):
+            _, length = self.config.name.split("-")
+            length = int(length)
+            data_file = dl_manager.download(self.config.data_url)
+            return [
+                datasets.SplitGenerator(
+                    name=datasets.Split.TRAIN,
+                    gen_kwargs={
+                        "data_file": data_file,
+                        "split": "train",
+                        "length": length
+                    }
+                ),
+            ]
+
+        elif self.config.name == "raw":
+            data_file = dl_manager.download_and_extract(self.config.data_url)
+            return [
+                datasets.SplitGenerator(
+                    name=datasets.Split.TRAIN,
+                    gen_kwargs={
+                        "data_file": os.path.join(data_file, "namuwiki_20210301.json"),
+                        "split": "train"
+                    }
+                ),
+            ]
+
+    def _generate_examples(self, data_file, split, length=None):
+        os.system("pip install ijson")
+        import ijson
+        """Generate NamuWiki examples."""
+        _TARGET = {"title", "text", "contributors.item"}
+        n, output = 0, defaultdict(list)
+        with open(data_file) as f:
+            for key, dtype, value in ijson.parse(f):
+                key = key.replace("item.", "")
+                if key == "namespace" and len(output):
+                    output = {k: (v[0] if k != "contributors" else v) for k, v in output.items()}
+                    yield n, processing(output, self.config.name)
+                    output = defaultdict(list)
+                    n += 1
+                elif key in _TARGET:
+                    output[key.replace(".item", "")].append(value)

process.py

@@ -0,0 +1,76 @@
+import re
+from collections import defaultdict
+
+
+def remove_double_linked_text(text):
+    PATTERN = "\[\[[^\[\]\|]+\|([^\]]+)\]\]"
+    result = re.search(PATTERN,text)
+    while result is not None:
+        s,e = result.span()
+        text = text[:s]+result.group(1)+text[e:]
+        result = re.search(PATTERN, text)
+    return text
+
+
+def remove_linked_text(text):
+    PATTERN = "\[\[([^\[\]]+)\]\]"
+    result = re.search(PATTERN,text)
+    while result is not None:
+        s,e = result.span()
+        text = text[:s]+result.group(1)+text[e:]
+        result = re.search(PATTERN, text)
+    return text
+
+
+def remove_attribute_in_table(text):
+    PATTERN = "{{{[^}]+ ([^\}]+)}}}"
+    result = re.search(PATTERN,text)
+    while result is not None:
+        s,e = result.span()
+        text = text[:s]+result.group(1)+text[e:]
+        result = re.search(PATTERN, text)
+
+    text = re.sub("<bgcolor=#[^>]+>", "", text)
+    text = re.sub("<-[0-9]>", "", text)
+    text = re.sub("\|\|<table[^\n]+\n", "", text)
+    text = re.sub("<tablewidth\=[^>]+>", "", text)
+    text = re.sub("<width\=[^>]+>", "", text)
+    text = re.sub("(?<=코멘트\-)\|\|(?=\n)", "", text)
+
+    return text
+
+
+def replace_link(text):
+    text = re.sub("\[youtube\([^\]]+\)\]", "[YOUTUBE LINK]", text)
+    return text
+
+
+def process_text(text: str):
+    text = text.strip()
+    text = re.sub("\[\[파일:[^\]]+\]\]", "", text)
+    text = remove_double_linked_text(text)
+    text = remove_linked_text(text)
+    text = re.sub("'''", "", text)
+    text = replace_link(text)
+    text = remove_attribute_in_table(text)
+    return text
+
+
+def get_structured_data(text: str, pattern="\n\=\= ([^\=\n]+) \=\=\n", default_value=None) -> dict:
+    outputs = defaultdict(list)
+    matched = re.search(pattern, text)
+    is_first = True
+    while matched is not None:
+        b,s = matched.span()
+        if is_first:
+            outputs['item'].append("meta")
+            outputs["content"].append(text[:b])
+            is_first = False
+        outputs["item"].append(matched.group(1))
+        text = text[s:]
+        matched = re.search(pattern, text)
+        e = matched.start() if matched is not None else None
+        outputs["content"].append(text[:e].strip())
+    if not outputs and default_value is not None:
+        return default_value
+    return dict(outputs)