The dataset and the cleaning script.

cleanup.py

@@ -0,0 +1,107 @@
+#!/usr/bin/python3
+
+from nltk.tokenize import word_tokenize
+from spellchecker import SpellChecker
+import pyarrow as pa
+import pyarrow.parquet as pq
+import pandas as pd
+import sys
+
+# Line acceptance threshold [0,1]
+THRESHOLD = 0.4
+
+#Levenshtein Distance weights
+LD1_WEIGHT = 0.5
+LD2_WEIGHT = 0.2
+
+# Minimum word length to be accepted at specified Levenshtein Distance
+LD1_MIN_LENGTH = 5
+LD2_MIN_LENGTH = 10
+
+# Tokens not to be penalized
+COMMON = set([',','.','|','\'','-','_','"','“','(',')','[',']'])
+def accept_token(token) :
+    """
+    If a word has already passed the spell checker, accept it under 
+    the conditions below.
+    """
+    if token.isdigit() :
+        return True
+    if token.isupper() :
+        return True
+    if token.istitle() :
+        return True
+    if len(token) == 1 and token in COMMON :
+        return True
+    return False
+    
+spell = SpellChecker()    
+def line_score(line) :
+    """
+    Assign each line a score based on how many words are spelled correctly
+    and how many have small mistakes.  Aome tokens are accepted automatically.
+    """
+    tokens = word_tokenize(line)
+    if len(tokens) == 0 :
+        return 0
+    known = spell.known(tokens)
+    spelled = 0
+    accepted = 0
+    ldist1 , ldist2 = 0, 0
+    for token in tokens :
+        l_token = token.lower()
+        if l_token in known :
+            spelled += 1
+        elif len(l_token) > LD1_MIN_LENGTH and len(spell.edit_distance_1(l_token)) > 0 :
+            ldist1 += 1
+        elif len(l_token) > LD2_MIN_LENGTH and len(spell.edit_distance_2(l_token)) > 0 :
+            ldist2 += 1
+        elif accept_token(token) :
+            accepted += 1
+    return ( spelled + accepted + (ldist1 * LD1_WEIGHT) + (ldist2 * LD2_WEIGHT) ) * 1.0 / len(tokens) 
+    
+def filter_content(content) :
+    """
+    Evaluate each line of the content, assign a line number to each line
+    and split the content in 'Accepted' and 'Rejected' catenations.
+    """
+    content = content.split("\n")    
+    line_no = 0
+    accepted, rejected = [], []
+    for line in content:
+        line_no += 1
+        score = line_score(line)
+        line = f"[{line_no}]{line}"
+        if score >= THRESHOLD :
+            accepted.append(line)
+        else :
+            rejected.append(line)
+    accepted = "\n".join(accepted)
+    rejected = "\n".join(rejected)
+    return accepted, rejected
+
+source = pq.ParquetFile(sys.argv[1])
+dest = None
+
+for rgi in range(source.num_row_groups) :
+    rg = source.read_row_group(rgi).to_pylist()
+    d = {}
+    d['File'] = []
+    d['Accepted'] = []
+    d['Rejected'] = []
+    for row in rg :
+        file_name = row['File']
+        content = row['Content']
+        accepted, rejected = filter_content(content)
+        d['File'].append(file_name)
+        d['Accepted'].append(accepted)
+        d['Rejected'].append(rejected)
+    df = pd.DataFrame(d, dtype=str)
+    table = pa.Table.from_pandas(df)
+    if dest is None :
+        dest = pq.ParquetWriter(sys.argv[2], table.schema)
+    dest.write_table(table)
+
+dest.close()
+
+

jfk_2025_cleaned.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edd7b2bae46d4d84ea8bc69cd69aa56f70bc243041e328d80ec18ee68ee67927
+size 81216939