feat: training english

dataset.py

@@ -5,8 +5,6 @@ from read_data import TagIdConverter
 from preprocessing import readPreporcssedDataAll
 from transformers import PreTrainedTokenizer
 
-tagIdConverter = TagIdConverter()
-
 class DatasetArray(Dataset):
     def __init__(self, data):
         self.x = data
@@ -39,7 +37,7 @@ def wrap_sentence(tokenizer: PreTrainedTokenizer, sentence):
 def wrap_entities(tagIdConverter: TagIdConverter, entities):
     return [tagIdConverter.O_id] + entities + [tagIdConverter.O_id]
 
-def make_collate_fn(tokenizer: PreTrainedTokenizer):
+def make_collate_fn(tokenizer: PreTrainedTokenizer, tagIdConverter: TagIdConverter):
     def ret_fn(batch):
         words = [wrap_sentence(tokenizer,item["ids"]) for item in batch]
         entities = [wrap_entities(tagIdConverter,item["entity_ids"]) for item in batch]

eng_tags.json

@@ -0,0 +1,42 @@
+[
+  {
+    "name": "[PAD]",
+    "index": 0
+  },
+  {
+    "name": "B-LOC",
+    "index": 1
+  },
+  {
+    "name": "B-MISC",
+    "index": 2
+  },
+  {
+    "name": "B-ORG",
+    "index": 3
+  },
+  {
+    "name": "B-PER",
+    "index": 4
+  },
+  {
+    "name": "I-LOC",
+    "index": 5
+  },
+  {
+    "name": "I-MISC",
+    "index": 6
+  },
+  {
+    "name": "I-ORG",
+    "index": 7
+  },
+  {
+    "name": "I-PER",
+    "index": 8
+  },
+  {
+    "name": "O",
+    "index": 9
+  }
+]

groupby_index.py

@@ -29,6 +29,7 @@ if __name__ == "__main__":
     test_list = [*range(20,-10,-1)]
     for g in groupby_index(test_list,4):
         print([*g])
+    print("===")
     print([*map(lambda x:[*x],groupby_index([1,2,3,4],2))])
     for g in groupby_index([1,2,3,4],2):
         print([*g])

preprocessing.py

@@ -1,16 +1,17 @@
 
 
-from read_data import TagIdConverter, make_long_namedEntity, readKoreanDataAll, Sentence
+import argparse
-from typing import Any, NamedTuple, List, Sequence, TypeVar
+import os
+import sys
+from read_data import TagIdConverter, make_long_namedEntity, readEnglishDataAll, readKoreanDataAll, Sentence
+from typing import Any, List
 import json
-import os.path as path
 import tqdm
 from transformers import PreTrainedTokenizer
 
-PREPROCESSING_BASE_PATH = 'prepro'
+PRE_BASE_PATH = 'prepro'
-converter = TagIdConverter()
 
-def preprocessing(tokenizer : PreTrainedTokenizer,dataset: List[Sentence]):
+def preprocessing(tokenizer : PreTrainedTokenizer, converter :TagIdConverter,dataset: List[Sentence]):
     ret = []
     for item in tqdm.tqdm(dataset):
         assert len(item.word) == len(item.detail)
@@ -35,26 +36,44 @@ def readPreprocessedData(path: str):
     with open(path,"r", encoding="utf-8") as fp:
         return json.load(fp)
 
-def readPreporcssedDataAll():
+def readPreporcssedDataAll(path = PRE_BASE_PATH):
-    train = readPreprocessedData(path.join(PREPROCESSING_BASE_PATH,"train.json"))
+    train = readPreprocessedData(os.path.join(path,"train.json"))
-    dev = readPreprocessedData(path.join(PREPROCESSING_BASE_PATH,"dev.json"))
+    dev = readPreprocessedData(os.path.join(path,"dev.json"))
-    test = readPreprocessedData(path.join(PREPROCESSING_BASE_PATH,"test.json"))
+    test = readPreprocessedData(os.path.join(path,"test.json"))
     return train, dev, test
 
 if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--kind", default="korean")
+    parser.add_argument("path",default=PRE_BASE_PATH,help="directory path of processed data")
+    parser.add_argument("--tag", default="tags.json",help="path of tag description")
+
+    args = parser.parse_args()
+    dirPath = args.path
+    
+    if args.kind == "korean":
+        rawTrain, rawDev, rawTest = readKoreanDataAll()
+    elif args.kind == "english":
+        rawTrain, rawDev, rawTest = readEnglishDataAll()
+    else:
+        print("unknown language",file=sys.stderr)
+        exit(1)
+
+    converter = TagIdConverter(args.tag)
+    os.makedirs(dirPath)
+
     from transformers import BertTokenizer
     PRETAINED_MODEL_NAME = 'bert-base-multilingual-cased'
 
-    rawTrain, rawDev, rawTest = readKoreanDataAll()
+    print("load tokenzier...",file=sys.stderr)
-    print("load tokenzier...")
     tokenizer = BertTokenizer.from_pretrained(PRETAINED_MODEL_NAME)
 
-    print("process train...")
+    print("process train...",file=sys.stderr)
-    train = preprocessing(tokenizer,rawTrain)
+    train = preprocessing(tokenizer,converter,rawTrain)
-    saveObject(path.join(PREPROCESSING_BASE_PATH,"train.json"),train)
+    saveObject(path.join(dirPath,"train.json"),train)
-    print("process dev...")
+    print("process dev...",file=sys.stderr)
-    dev = preprocessing(tokenizer,rawDev)
+    dev = preprocessing(tokenizer,converter,rawDev)
-    saveObject(path.join(PREPROCESSING_BASE_PATH,"dev.json"),dev)
+    saveObject(path.join(dirPath,"dev.json"),dev)
-    print("process test...")
+    print("process test...",file=sys.stderr)
-    test = preprocessing(tokenizer,rawTest)
+    test = preprocessing(tokenizer,converter,rawTest)
-    saveObject(path.join(PREPROCESSING_BASE_PATH,"test.json"),test)
+    saveObject(path.join(dirPath,"test.json"),test)

read_data.py

@@ -1,6 +1,9 @@
 import enum
-from typing import NamedTuple, List, Sequence, TypeVar
+from io import TextIOWrapper
+import sys
+from typing import Iterable, NamedTuple, List, Sequence, TypeVar
 import json
+import argparse
 
 KoreanBase="[Ko, En] NER, POStag data/국문 NER, POS"
 EnglishBase="[Ko, En] NER, POStag data/영문 NER, POS"
@@ -25,48 +28,62 @@ class Sentence(NamedTuple):
         self.namedEntity.append(namedEntity)
         self.detail.append(detail)
 T = TypeVar('T')
-def readDataList(lst: List[T]):
+def readDataList(lst: Iterable[str], sep="\t"):
-    ret = []
+    ret:List[str] = []
     for l in lst:
-        if len(l) > 0:
+        l = l.strip()
-            ret.append(l)
+        if l == "":
-        else:
             yield ret
             ret.clear()
-
-def readKoreanData(path: str) -> List[Sentence]:
-    fp = open(path,encoding="utf-8")
-    data = []
-    for line in fp.readlines():
-        line = line.strip()
-        if line == "":
-            data.append([])
         else:
-            data.append(line.split("\t"))
+            ret.append(l.split(sep))
-    fp.close()
-    # Do not use csv reader.
-    ret = []
 
-    for lines in readDataList(data):
+def readKoreanData(fp: TextIOWrapper) -> List[Sentence]:
+    ret = []
+    # NOTE(monoid): Do not use csv reader.
+    for lines in readDataList(fp):
         sentence = Sentence([],[],[],[])
         for line in lines:
             word_pos:str = line[0]
             words = word_pos.split("/")
             sentence.append(words[0],line[1],line[2],line[3])
         ret.append(sentence)
-
+    
+    fp.close()
     return ret
 
+def readEnglishData(fp: TextIOWrapper) -> List[Sentence]:
+    ret = []
+    for lines in readDataList(fp,sep=" "):
+        if len(lines) == 1 and lines[0][0] == "-DOCSTART-":
+            continue
+        sentence = Sentence([],[],[],[])
+        for line in lines:
+            sentence.append(line[0],line[1],line[2],line[3])
+        ret.append(sentence)
+    return ret
 
 def readKoreanDataAll():
     """
-        @return train, dev, test tuple
         Each entry is structured as follows:
             POS, 
+        Return: train, dev, test tuple
     """
-    dev = readKoreanData(f"{KoreanBase}/dev.txt")
+    with open(f"{KoreanBase}/dev.txt", encoding="utf-8") as fp:
-    test = readKoreanData(f"{KoreanBase}/test.txt")
+        dev = readKoreanData(fp)
-    train = readKoreanData(f"{KoreanBase}/train.txt")
+    with open(f"{KoreanBase}/test.txt", encoding="utf-8") as fp:
+        test = readKoreanData(fp)
+    with open(f"{KoreanBase}/train.txt", encoding="utf-8") as fp:
+        train = readKoreanData(fp)
+    return train, dev, test
+
+def readEnglishDataAll():
+    with open(f"{EnglishBase}/valid.txt", encoding="utf-8") as fp:
+        dev = readEnglishData(fp)
+    with open(f"{EnglishBase}/test.txt", encoding="utf-8") as fp:
+        test = readEnglishData(fp)
+    with open(f"{EnglishBase}/train.txt", encoding="utf-8") as fp:
+        train = readEnglishData(fp)
     return train, dev, test
 
 class TagIdConverter:
@@ -86,7 +103,13 @@ class TagIdConverter:
     @property
     def pad_id(self):
         return self.vocab["[PAD]"]
+    @property
+    def size(self):
+        return len(self.vocab)
     
+    def __len__(self):
+        return self.size
+
     def convert_ids_to_tokens(self,ids: List[int]):
         return [self.ids_to_token[id] for id in ids]
     def convert_tokens_to_ids(self, tokens: List[str]):
@@ -151,29 +174,37 @@ def make_long_namedEntity(a,b,c):
             break
     return ret
 
-
-
 """
 extracts and stores tags set from the given data. 
 """
 if __name__ == "__main__":
-    from tqdm import tqdm
+    parser = argparse.ArgumentParser(description="create tags list")
-    t = TagIdConverter()
+    parser.add_argument("--kind","-k",default='korean', help='kind of language: korean or english')
+    parser.add_argument("--stdout",action='store_true',help='print tags data to stdout')
+    parser.add_argument("--path",default="tags.json", help="path of tags data")
     
-    train, dev, test = readKoreanDataAll()
+    args = parser.parse_args()
+
+    from tqdm import tqdm
+    if args.kind == "korean" or args.kind == "ko" or args.kind == "kor":
+        train, dev, test = readEnglishDataAll()
+    elif args.kind == "english" or args.kind == "en" or args.kind =="eng":
+        train, dev, test = readKoreanDataAll()
+    else:
+        print("unknown language",file=sys.stderr)
+        exit(1)
     vocab = set()
     def getTags(lst: List[Sentence]):
         for s in tqdm(lst):
             for e in s.detail:
                 vocab.add(e)
-    print("get tags from train...")
+    print("get tags from train...",file=sys.stderr)
     getTags(train)
-    print("get tags from dev...")
+    print("get tags from dev...",file=sys.stderr)
     getTags(dev)
-    print("get tags from test...")
+    print("get tags from test...",file=sys.stderr)
     getTags(test)
-    print(vocab)
+    print(vocab,file=sys.stderr)
-
     for v in vocab:
         if v == "O":
             continue
@@ -183,7 +214,6 @@ if __name__ == "__main__":
         if not v in vocab:
             print("could not found pair " ,v)
             vocab.add(v)
-
     tags = [{"name":"[PAD]","index":0}]
     i = 1
     vocab_list = [*vocab]
@@ -191,6 +221,10 @@ if __name__ == "__main__":
     for v in vocab_list:
         tags.append({"name":v,"index":i})
         i += 1
-    print(tags)
+    print(tags,file=sys.stderr)
-    with open("tags.json","w",encoding="utf-8") as fp:
+    if args.stdout:
-        json.dump(tags,fp,ensure_ascii=False, indent=2)
+        json.dump(tags,sys.stdout,ensure_ascii=False, indent=2)
+    else:
+        p = args.path
+        with open(p,"w",encoding="utf-8") as fp:
+            json.dump(tags,fp,ensure_ascii=False, indent=2)