scripts/build_features.py

import os
import re
import sys
import yaml
import argparse
import pandas as pd
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
import joblib

# French stopwords list
STOPWORDS = {
    "le", "la", "les", "de", "des", "du", "d'", "l'", "un", "une", "et", "en", "que", "qui", 
    "pour", "dans", "ce", "ces", "se", "par", "sur", "ou", "a", "au", "aux", "est", "sont", 
    "ont", "aussi", "avec", "mais", "pas", "plus", "ne", "je", "tu", "il", "nous", "vous", 
    "ils", "elle", "elles", "on", "y", "en", "qu'", "ceux", "celles", "ceci", "cela", "c'"
}

PUNCTUATION_CHARS = ".,!?;:-()\"'«»"

def load_config(config_path):
    with open(config_path, "r", encoding="utf-8") as f:
        return yaml.safe_load(f)

def tokenize(text):
    """Simple French word tokenizer."""
    # Replace punctuation with spaces except internal apostrophes
    text_clean = re.sub(r"[^\w\d'\-]", " ", text.lower())
    words = text_clean.split()
    return words

def get_sentences(text):
    """Simple sentence segmenter."""
    # Split sentences by period, exclamation, question mark, or newline
    sentences = re.split(r"(?<=[.!?])\s+|\n+", text)
    # Remove empty sentences
    sentences = [s.strip() for s in sentences if s.strip()]
    return sentences

def extract_stylometric_features(text, connecteurs_list):
    """Extracts stylometric, lexical, and syntactic features from a French text."""
    features = {}
    
    # Raw Text Stats
    features["num_chars"] = len(text)
    words = tokenize(text)
    features["num_words"] = len(words)
    
    sentences = get_sentences(text)
    features["num_sentences"] = len(sentences)
    
    # Sentence lengths (word counts)
    sent_lens = [len(tokenize(s)) for s in sentences if len(tokenize(s)) > 0]
    if sent_lens:
        features["avg_sentence_len"] = np.mean(sent_lens)
        features["std_sentence_len"] = np.std(sent_lens) if len(sent_lens) > 1 else 0.0
    else:
        features["avg_sentence_len"] = 0.0
        features["std_sentence_len"] = 0.0
        
    # Word lengths
    word_lens = [len(w) for w in words]
    if word_lens:
        features["avg_word_len"] = np.mean(word_lens)
        long_words = sum(1 for l in word_lens if l > 6)
        features["ratio_long_words"] = long_words / len(words)
    else:
        features["avg_word_len"] = 0.0
        features["ratio_long_words"] = 0.0
        
    # Punctuation and symbols
    punc_count = sum(1 for c in text if c in PUNCTUATION_CHARS)
    features["ratio_punctuation"] = punc_count / len(text) if len(text) > 0 else 0.0
    
    uppercase_count = sum(1 for c in text if c.isupper())
    features["freq_uppercase"] = uppercase_count / len(text) if len(text) > 0 else 0.0
    
    digit_count = sum(1 for c in text if c.isdigit())
    features["freq_digits"] = digit_count / len(text) if len(text) > 0 else 0.0
    
    # Non-alphanumeric, non-punctuation, non-space characters
    symbol_count = sum(1 for c in text if not c.isalnum() and c not in PUNCTUATION_CHARS and not c.isspace())
    features["freq_symbols"] = symbol_count / len(text) if len(text) > 0 else 0.0
    
    # Lexical Features
    if words:
        unique_words = set(words)
        features["vocabulary_diversity"] = len(unique_words) / len(words)
        
        # Hapax legomena
        word_counts = {}
        for w in words:
            word_counts[w] = word_counts.get(w, 0) + 1
        hapaxes = sum(1 for w, c in word_counts.items() if c == 1)
        features["hapax_ratio"] = hapaxes / len(words)
        
        # Stopwords
        stop_count = sum(1 for w in words if w in STOPWORDS)
        features["stopword_ratio"] = stop_count / len(words)
        
        # Connectors
        connector_count = 0
        text_lower = text.lower()
        for conn in connecteurs_list:
            # Match whole phrase/word using regex word boundaries (need to escape connecteurs)
            pattern = r'\b' + re.escape(conn) + r'\b'
            matches = re.findall(pattern, text_lower)
            connector_count += len(matches)
        features["connector_ratio"] = connector_count / len(words)
        
        # Repetition ratio of content words (non-stopwords)
        content_words = [w for w in words if w not in STOPWORDS]
        if content_words:
            content_counts = {}
            for w in content_words:
                content_counts[w] = content_counts.get(w, 0) + 1
            features["repetition_ratio"] = (len(content_words) - len(content_counts)) / len(content_words)
        else:
            features["repetition_ratio"] = 0.0
    else:
        features["vocabulary_diversity"] = 0.0
        features["hapax_ratio"] = 0.0
        features["stopword_ratio"] = 0.0
        features["connector_ratio"] = 0.0
        features["repetition_ratio"] = 0.0
        
    # Syntax-like Proxy Features
    # Count typical conjunctions and subordinators that indicate phrase complexity
    subordinators = ["que", "qui", "dont", "où", "lequel", "laquelle", "lesquels", "auxquels", "quand", "comme", "si", "car", "puisque", "lorsque"]
    sub_count = sum(1 for w in words if w in subordinators)
    features["syntactic_complexity_score"] = sub_count / len(sentences) if len(sentences) > 0 else 0.0
    
    # Sentences structure distribution
    features["ratio_interrogative"] = sum(1 for s in sentences if s.endswith("?")) / len(sentences) if len(sentences) > 0 else 0.0
    features["ratio_exclamative"] = sum(1 for s in sentences if s.endswith("!")) / len(sentences) if len(sentences) > 0 else 0.0
    features["ratio_declarative"] = 1.0 - (features["ratio_interrogative"] + features["ratio_exclamative"])
    
    # Verb tense approximations (simple counts of suffix occurrences)
    # imparfait endings: ais, ait, ions, iez, aient
    imparfait = len(re.findall(r'\b\w+(?:ais|ait|ions|iez|aient)\b', text_lower))
    # futur endings: rai, ras, ra, rons, rez, ront
    futur = len(re.findall(r'\b\w+(?:rai|ras|ra|rons|rez|ront)\b', text_lower))
    # conditional endings: rais, rait, rions, riez, raient
    conditional = len(re.findall(r'\b\w+(?:rais|rait|rions|riez|raient)\b', text_lower))
    
    total_verbs_est = imparfait + futur + conditional + 1 # avoid division by zero
    features["imparfait_ratio"] = imparfait / total_verbs_est
    features["futur_ratio"] = futur / total_verbs_est
    features["conditional_ratio"] = conditional / total_verbs_est
    
    return features

def main():
    parser = argparse.ArgumentParser(description="Feature engineering for AI text detection.")
    parser.add_argument("--config", default="configs/config.yaml", help="Path to config file")
    args = parser.parse_args()
    
    config = load_config(args.config)
    raw_dir = config["paths"]["raw_dir"]
    processed_dir = config["paths"]["processed_dir"]
    os.makedirs(processed_dir, exist_ok=True)
    os.makedirs(config["paths"]["models_dir"], exist_ok=True)
    
    connecteurs = config["features"]["connecteurs"]
    
    # 1. Process Training Data
    print("Loading raw training datasets...")
    human_path = os.path.join(raw_dir, "human_corpus.csv")
    ai_path = os.path.join(raw_dir, "ai_corpus.csv")
    
    if not os.path.exists(human_path) or not os.path.exists(ai_path):
        print("Error: Raw training data not found! Please run 'python scripts/collect_data.py' first.")
        sys.exit(1)
        
    df_human = pd.read_csv(human_path)
    df_ai = pd.read_csv(ai_path)
    
    # Concatenate training corpora
    df_train = pd.concat([df_human, df_ai], ignore_index=True)
    print(f"Total training data size: {len(df_train)} rows.")
    
    # Extract stylometric features
    print("Extracting stylometric, lexical, and syntactic features for training set...")
    stylometric_features = []
    for text in df_train["text"]:
        stylometric_features.append(extract_stylometric_features(text, connecteurs))
        
    df_sty = pd.DataFrame(stylometric_features)
    df_train_feats = pd.concat([df_train, df_sty], axis=1)
    
    # N-gram feature extraction using TF-IDF (word 1-2 and char 3-4 grams)
    print("Fitting TF-IDF Vectorizers for word and character n-grams...")
    # Word n-grams
    word_vectorizer = TfidfVectorizer(
        ngram_range=tuple(config["features"]["ngram_word_range"]),
        max_features=config["features"]["top_n_ngrams"] // 2,
        stop_words=None  # We want function words / stopwords in n-grams!
    )
    word_tfidf = word_vectorizer.fit_transform(df_train_feats["text"])
    word_cols = [f"ngram_word_{i}" for i in range(word_tfidf.shape[1])]
    df_word_ngrams = pd.DataFrame(word_tfidf.toarray(), columns=word_cols)
    
    # Character n-grams
    char_vectorizer = TfidfVectorizer(
        analyzer='char',
        ngram_range=tuple(config["features"]["ngram_char_range"]),
        max_features=config["features"]["top_n_ngrams"] // 2
    )
    char_tfidf = char_vectorizer.fit_transform(df_train_feats["text"])
    char_cols = [f"ngram_char_{i}" for i in range(char_tfidf.shape[1])]
    df_char_ngrams = pd.DataFrame(char_tfidf.toarray(), columns=char_cols)
    
    # Combine everything
    df_train_final = pd.concat([df_train_feats, df_word_ngrams, df_char_ngrams], axis=1)
    
    # Save vectorizers for inference
    joblib.dump(word_vectorizer, os.path.join(config["paths"]["models_dir"], "word_vectorizer.pkl"))
    joblib.dump(char_vectorizer, os.path.join(config["paths"]["models_dir"], "char_vectorizer.pkl"))
    print("Vectorizers saved.")
    
    # Save final processed training dataset
    train_output = os.path.join(processed_dir, "train_features.csv")
    df_train_final.to_csv(train_output, index=False)
    print(f"Processed training dataset saved to {train_output}")
    
    # 2. Process Recent Debates Data
    print("Loading recent debates for inference...")
    recent_path = os.path.join(raw_dir, "recent_debates.csv")
    if os.path.exists(recent_path):
        df_recent = pd.read_csv(recent_path)
        print("Extracting features for recent debates...")
        recent_sty_list = []
        for text in df_recent["text"]:
            recent_sty_list.append(extract_stylometric_features(text, connecteurs))
        df_recent_sty = pd.DataFrame(recent_sty_list)
        df_recent_feats = pd.concat([df_recent, df_recent_sty], axis=1)
        
        # Transform n-grams using fitted vectorizers
        recent_word_tfidf = word_vectorizer.transform(df_recent_feats["text"])
        df_recent_word_ngrams = pd.DataFrame(recent_word_tfidf.toarray(), columns=word_cols)
        
        recent_char_tfidf = char_vectorizer.transform(df_recent_feats["text"])
        df_recent_char_ngrams = pd.DataFrame(recent_char_tfidf.toarray(), columns=char_cols)
        
        # Combine
        df_recent_final = pd.concat([df_recent_feats, df_recent_word_ngrams, df_recent_char_ngrams], axis=1)
        
        # Save processed recent debates
        recent_output = os.path.join(processed_dir, "recent_features.csv")
        df_recent_final.to_csv(recent_output, index=False)
        print(f"Processed recent debates dataset saved to {recent_output}")
    else:
        print("No recent debates dataset found in raw directory to process.")
        
    print("Feature engineering completed successfully.")

if __name__ == "__main__":
    main()