initial commit

README.md

@@ -3,6 +3,32 @@ license: bsd
 language:
 - nl
 - fr
+configs:
+- config_name: line_by_line
+  data_files: 
+    - split: train
+      path : "data/line_by_line/train.csv"
+    - split: test
+      path : "data/line_by_line/test.csv"
+    - split: valid
+      path : "data/line_by_line/valid.csv"
+  default: true
+- config_name: page_by_page
+  data_files: 
+    - split: train
+      path : "data/page_by_page/train.csv"
+    - split: test
+      path : "data/page_by_page/test.csv"
+    - split: valid
+      path : "data/page_by_page/valid.csv"
+- config_name: doc_by_doc
+  data_files: 
+    - split: train
+      path : "data/doc_by_doc/train.csv"
+    - split: test
+      path : "data/doc_by_doc/test.csv"
+    - split: valid
+      path : "data/doc_by_doc/valid.csv"
 ---
 # Brabant XVII
 

todataset.py

@@ -0,0 +1,179 @@
+import os
+import os.path as path
+import unittest
+import pandas as pd
+
+from tokenizers             import Regex, NormalizedString
+from tokenizers.normalizers import Sequence, NFKC, Replace, BertNormalizer, Strip
+
+from datasets               import Dataset
+
+class CustomNormalizer:
+    def __init__(self):
+        self.normalization = Sequence([
+            NFKC(),
+            Replace("[",    ""), # les annotations sont tellement inconsistantes pour les [] et le () que
+            Replace("(",    ""), # que je pense qu'il vaut mieux tout virer et ls considérer tous pareil
+            Replace("]",    ""),
+            Replace(")",    ""),
+            Replace("\r\n", ""),
+            Replace(Regex(r"\s{2,}"), " "),
+            BertNormalizer(clean_text=True, handle_chinese_chars=True, strip_accents=True, lowercase=True),
+            Strip()
+        ])
+
+    def normalize(self, text: NormalizedString):
+        return self.normalization.normalize(text)
+    
+    def normalize_str(self, sequence: str):
+        return self.normalization.normalize_str(sequence)
+
+
+class TestNormalizer(unittest.TestCase):
+    def test_it_normalizes_in_place(self):
+        norm     = CustomNormalizer()
+        actual   = NormalizedString("coucou[[[xavier")
+        norm.normalize(actual)
+        expected = "coucouxavier"
+        self.assertEqual(actual.normalized, expected)
+
+    def test_it_removes_opening_bracket(self):
+        norm     = CustomNormalizer()
+        actual   = norm.normalize_str("coucou[[[xavier")
+        expected = "coucouxavier"
+        self.assertEqual(actual, expected)
+    
+    def test_it_removes_closing_bracket(self):
+        norm     = CustomNormalizer()
+        actual   = norm.normalize_str("coucou]]]xavier")
+        expected = "coucouxavier"
+        self.assertEqual(actual, expected)
+    
+    def test_it_removes_both_brackets(self):
+        norm     = CustomNormalizer()
+        actual   = norm.normalize_str("coucou[[[-]]]xavier")
+        expected = "coucou-xavier"
+        self.assertEqual(actual, expected)
+    
+    def test_it_lowertcases_text(self):
+        norm     = CustomNormalizer()
+        actual   = norm.normalize_str("CouCou Xavier")
+        expected = "coucou xavier"
+        self.assertEqual(actual, expected)
+
+    def test_it_replaces_long_sequences_of_whitespaces(self):
+        norm     = CustomNormalizer()
+        actual   = norm.normalize_str("coucou    xavier")
+        expected = "coucou xavier"
+        self.assertEqual(actual, expected)
+        #
+        actual   = norm.normalize_str("coucou  xavier")
+        expected = "coucou xavier"
+        self.assertEqual(actual, expected)
+        #
+        actual   = norm.normalize_str("coucou \txavier")
+        expected = "coucou xavier"
+        self.assertEqual(actual, expected)
+    
+    def test_it_normalizes_funky_chars(self):
+        norm     = CustomNormalizer()
+        actual   = norm.normalize_str("coucou xavier")
+        expected = "coucou xavier"
+        self.assertEqual(actual, expected)
+        #
+        actual   = norm.normalize_str("coucou  xavier")
+        expected = "coucou xavier"
+        self.assertEqual(actual, expected)
+        #
+        actual   = norm.normalize_str("coucou \txavier")
+        expected = "coucou xavier"
+        self.assertEqual(actual, expected)
+
+
+def file_to_pandas(project, fname):
+    norm    = CustomNormalizer()
+    file_id = path.basename(fname)
+    page_id = 0
+    line_id = 0
+    df_dict = {
+        "project": [],
+        "file_id": [],
+        "page_id": [],
+        "line_id": [],
+        "text"   : []
+    }
+    #
+    with open(fname, encoding="utf8") as f:
+        for line in f.readlines():
+            line = NormalizedString(line)
+            norm.normalize(line)
+            line = line.normalized
+            #
+            if not line: # emptystring is seen as false
+                pass
+            elif line == "###":
+                page_id += 1
+            else:
+                line_id += 1
+                df_dict["project"].append(project)
+                df_dict["file_id"].append(file_id)
+                df_dict["page_id"].append(page_id)
+                df_dict["line_id"].append(line_id)
+                df_dict["text"]   .append(line)
+    #
+    return pd.DataFrame.from_dict(df_dict)
+
+def text(df):
+    return "\n".join(df["text"].to_list())
+
+def documents_text(df: pd.DataFrame):
+    '''
+    This function returns a dataframe equivalent to the original 'df'. However, the returned dataframe
+    loses the page and line information so as to retrieve the complete text of each documents pertaining
+    to the dataset
+    
+    # Params
+    - df is a pandas dataframe
+    '''
+    groups = df.groupby(["project", "file_id"])
+    data   = [(project, file_id, text(groups.get_group((project, file_id)).sort_values(["page_id", "line_id"]))) for (project, file_id),_v in groups ]
+    return pd.DataFrame.from_records(data, columns=["project", "file_id", "text"])
+
+def pages_text(df: pd.DataFrame):
+    '''
+    This function returns a dataframe equivalent to the original 'df'. However, the returned dataframe
+    loses the line information so as to retrieve the complete text of each page of processed data.s
+    
+    # Params
+    - df is a pandas dataframe
+    '''
+    groups = df.groupby(["project", "file_id", "page_id"])
+    data   = [(project, file_id, page_id, text(groups.get_group((project, file_id, page_id)).sort_values("line_id"))) for (project, file_id, page_id),_v in groups ]
+    return pd.DataFrame.from_records(data, columns=["project", "file_id","page_id", "text"])
+
+def prepare(df: pd.DataFrame, config_name: str):
+    ds     = Dataset.from_pandas(df)
+    splits = ds.train_test_split(train_size=0.8, shuffle=True)
+    train  = splits["train"]
+    test   = splits["test"]
+    splits = train.train_test_split(train_size=0.8)
+    train  = splits["train"]
+    valid  = splits["test"]
+    train.to_csv("data/{config}/train.csv".format(config=config_name))
+    test .to_csv("data/{config}/test.csv".format(config=config_name))
+    valid.to_csv("data/{config}/valid.csv".format(config=config_name))
+
+if __name__ == '__main__':
+    #pth= r'C:\Users\xavier.gillard\Documents\ARKEY\data'
+    #prj= "pardons"
+    #df = pd.concat([file_to_pandas(prj, path.join(pth, f)) for f in os.listdir(path=pth)], axis=0, ignore_index=True)
+    #df.to_csv("data/full_corpus.csv", index=False)
+    
+    df = pd.read_csv("data/full_corpus.csv")
+    lines = df[["text"]]
+    docs  = documents_text(df)[["text"]]
+    pages = pages_text(df)[["text"]]
+
+    prepare(lines, "line_by_line")
+    prepare(pages, "page_by_page")
+    prepare(docs,  "doc_by_doc")