Data splits (compatible with LEXTREME benchmark) are uploaded

convert_to_hf_dataset.py

@@ -0,0 +1,221 @@
+import os
+
+import numpy as np
+import pandas as pd
+
+import pickle
+import math
+import random
+
+"""
+Dataset url: https://github.com/koc-lab/law-turk/tree/main/data/constitutional/deep
+Paper url: https://www.sciencedirect.com/science/article/abs/pii/S0306457321001692
+"""
+
+def prepare_data():
+    label_dict = {' İhlal' : 0, ' İhlal Olmadığı' : 1} 
+
+    ### Load data
+    with open('old_dataset/constitutional_tokenized.law' ,'rb') as pickle_file:
+        tokenized = pickle.load(pickle_file)
+    with open('old_dataset/constitutional_labels.law' ,'rb') as pickle_file:
+        labels = pickle.load(pickle_file)
+
+    tokenized_lower = []
+
+    for text in tokenized:
+        temp_text = ""
+        for word in text:
+            temp_text += word.lower() + " "
+
+        tokenized_lower.append(temp_text)
+
+    ### Split the data
+
+    train_ratio = 0.70
+    val_ratio = 0.15
+
+    list_indices = []
+
+    for i, lbl in enumerate(labels):
+        if lbl in label_dict:
+            list_indices.append(i)
+
+    random.Random(13).shuffle(list_indices)
+
+    new_length = len(list_indices)
+
+    train_idx = math.floor(new_length * train_ratio)
+    val_idx = math.floor(new_length * (train_ratio + val_ratio))
+
+    train_indices = list_indices[0:train_idx]
+    val_indices = list_indices[train_idx : val_idx]
+    test_indices = list_indices[val_idx:]
+
+    train_list = []
+    val_list = []
+    test_list = []
+
+    for ind in train_indices:
+        train_list.append(tokenized_lower[ind])
+
+    for ind in val_indices:
+        val_list.append(tokenized_lower[ind])
+
+    for ind in test_indices:
+        test_list.append(tokenized_lower[ind])
+
+    train_labels = []
+    val_labels = []
+    test_labels = []
+    
+    for ind in train_indices:
+        if label_dict[labels[ind]] == " İhlal":
+            train_labels.append("Violation")
+        else:
+            train_labels.append("No violation")
+
+    for ind in val_indices:
+        if label_dict[labels[ind]] == " İhlal":
+            val_labels.append("Violation")
+        else:
+            val_labels.append("No violation")
+
+    for ind in test_indices:
+        if label_dict[labels[ind]] == " İhlal":
+            test_labels.append("Violation")
+        else:
+            test_labels.append("No violation")
+    
+    train_split = np.concatenate((np.expand_dims(train_list, axis=1), np.expand_dims(train_labels, axis=1)), axis=1)
+    val_split = np.concatenate((np.expand_dims(val_list, axis=1), np.expand_dims(val_labels, axis=1)), axis=1)
+    test_split = np.concatenate((np.expand_dims(test_list, axis=1), np.expand_dims(test_labels, axis=1)), axis=1)
+
+    return pd.DataFrame(train_split, columns=["Text", "Label"]), pd.DataFrame(val_split, columns=["Text", "Label"]), pd.DataFrame(test_split, columns=["Text", "Label"])
+
+train_split, val_split, test_split = prepare_data()
+train_split.to_json(os.path.join("train.jsonl"), lines=True, orient="records", force_ascii=False)
+val_split.to_json(os.path.join("validation.jsonl"), lines=True, orient="records", force_ascii=False)
+test_split.to_json(os.path.join("test.jsonl"), lines=True, orient="records", force_ascii=False)
+
+'''
+pd.set_option('display.max_colwidth', None)
+pd.set_option('display.max_columns', None)
+
+
+def perform_original_preprocessing():
+    # Original Preprocessing from: https://github.com/lagefreitas/predicting-brazilian-court-decisions/blob/main/predicting-brazilian-court-decisions.py#L81
+    # Loading the labeled decisions
+    data = pd.read_csv("dataset.csv", sep='<=>', header=0)
+    print('data.shape=' + str(data.shape) + ' full data set')
+    # Removing NA values
+    data = data.dropna(subset=[data.columns[9]])  # decision_description
+    data = data.dropna(subset=[data.columns[11]])  # decision_label
+    print('data.shape=' + str(data.shape) + ' dropna')
+    # Removing duplicated samples
+    data = data.drop_duplicates(subset=[data.columns[1]])  # process_number
+    print('data.shape=' + str(data.shape) + ' removed duplicated samples by process_number')
+    data = data.drop_duplicates(subset=[data.columns[9]])  # decision_description
+    print('data.shape=' + str(data.shape) + ' removed duplicated samples by decision_description')
+    # Removing not relevant decision labels and decision not properly labeled
+    data = data.query('decision_label != "conflito-competencia"')
+    print('data.shape=' + str(data.shape) + ' removed decisions labeled as conflito-competencia')
+    data = data.query('decision_label != "prejudicada"')
+    print('data.shape=' + str(data.shape) + ' removed decisions labeled as prejudicada')
+    data = data.query('decision_label != "not-cognized"')
+    print('data.shape=' + str(data.shape) + ' removed decisions labeled as not-cognized')
+    data_no = data.query('decision_label == "no"')
+    print('data_no.shape=' + str(data_no.shape))
+    data_yes = data.query('decision_label == "yes"')
+    print('data_yes.shape=' + str(data_yes.shape))
+    data_partial = data.query('decision_label == "partial"')
+    print('data_partial.shape=' + str(data_partial.shape))
+    # Merging decisions whose labels are yes, no, and partial to build the final data set
+    data_merged = data_no.merge(data_yes, how='outer')
+    data = data_merged.merge(data_partial, how='outer')
+    print('data.shape=' + str(data.shape) + ' merged decisions whose labels are yes, no, and partial')
+    # Removing decision_description and decision_labels whose values are -1 and -2
+    indexNames = data[(data['decision_description'] == str(-1)) | (data['decision_description'] == str(-2)) | (
+            data['decision_label'] == str(-1)) | (data['decision_label'] == str(-2))].index
+    data.drop(indexNames, inplace=True)
+    print('data.shape=' + str(data.shape) + ' removed -1 and -2 decision descriptions and labels')
+
+    data.to_csv("dataset_processed_original.csv", index=False)
+
+
+def perform_additional_processing():
+    df = pd.read_csv("dataset_processed_original.csv")
+
+    # remove strange " characters sometimes occurring in the beginning and at the end of a line
+    df.ementa_filepath = df.ementa_filepath.str.replace('^"', '')
+    df.decision_unanimity = df.decision_unanimity.str.replace('"$', '')
+
+    # removing process_type and judgment_date, since they are the same everywhere (-)
+    # decisions only contains 'None', nan and '-2'
+    # ementa_filepath refers to the name of file in the filesystem that we created when we scraped the data from the Court. It is temporary data and can be removed
+    # decision_description = ementa_text - decision_text - decision_unanimity_text
+    df = df.drop(['process_type', 'judgment_date', 'decisions', 'ementa_filepath'], axis=1)
+
+    # some rows are somehow not read correctly. With this, we can filter them
+    df = df[df.decision_text.str.len() > 1]
+
+    # rename "-2" to more descriptive name ==> -2 means, that they were not able to determine it
+    df.decision_unanimity = df.decision_unanimity.replace('-2', 'not_determined')
+
+    # rename cols for more clarity
+    df = df.rename(columns={"decision_unanimity": "unanimity_label"})
+    df = df.rename(columns={"decision_unanimity_text": "unanimity_text"})
+    df = df.rename(columns={"decision_text": "judgment_text"})
+    df = df.rename(columns={"decision_label": "judgment_label"})
+
+    df.to_csv("dataset_processed_additional.csv", index=False)
+
+    return df
+
+
+perform_original_preprocessing()
+df = perform_additional_processing()
+
+# perform random split 80% train (3234), 10% validation (404), 10% test (405)
+train, validation, test = np.split(df.sample(frac=1, random_state=42), [int(.8 * len(df)), int(.9 * len(df))])
+
+
+def save_splits_to_jsonl(config_name):
+    # save to jsonl files for huggingface
+    if config_name: os.makedirs(config_name, exist_ok=True)
+    train.to_json(os.path.join(config_name, "train.jsonl"), lines=True, orient="records", force_ascii=False)
+    validation.to_json(os.path.join(config_name, "validation.jsonl"), lines=True, orient="records", force_ascii=False)
+    test.to_json(os.path.join(config_name, "test.jsonl"), lines=True, orient="records", force_ascii=False)
+
+
+def print_split_table_single_label(train, validation, test, label_name):
+    train_counts = train[label_name].value_counts().to_frame().rename(columns={label_name: "train"})
+    validation_counts = validation[label_name].value_counts().to_frame().rename(columns={label_name: "validation"})
+    test_counts = test[label_name].value_counts().to_frame().rename(columns={label_name: "test"})
+
+    table = train_counts.join(validation_counts)
+    table = table.join(test_counts)
+    table[label_name] = table.index
+    total_row = {label_name: "total",
+                 "train": len(train.index),
+                 "validation": len(validation.index),
+                 "test": len(test.index)}
+    table = table.append(total_row, ignore_index=True)
+    table = table[[label_name, "train", "validation", "test"]]  # reorder columns
+    print(table.to_markdown(index=False))
+
+
+save_splits_to_jsonl("")
+
+print_split_table_single_label(train, validation, test, "judgment_label")
+print_split_table_single_label(train, validation, test, "unanimity_label")
+
+# create second config by filtering out rows with unanimity label == not_determined, while keeping the same splits
+# train = train[train.unanimity_label != "not_determined"]
+# validation = validation[validation.unanimity_label != "not_determined"]
+# test = test[test.unanimity_label != "not_determined"]
+
+
+# it is a very small dataset and very imbalanced (only very few not-unanimity labels)
+# save_splits_to_jsonl("unanimity")
+'''