scripts/collect_data.py

import os
import sys
import yaml
import argparse
import pandas as pd
import numpy as np
import requests
import random
from datetime import datetime, timedelta

# Import text generator
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
from text_generator import generate_speech, PARTIES, SPEAKERS, CHAMBERS

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

def clean_and_normalize(text):
    """Cleans text by normalizing whitespace, typography, accents, and punctuation."""
    if not isinstance(text, str):
        return ""
    # Normalize typography
    text = text.replace("’", "'").replace("œ", "oe").replace("æ", "ae")
    # Normalize whitespaces
    text = " ".join(text.split())
    return text

def map_row_to_schema(row, seed_idx):
    """Maps a raw row from agokrani/fr-political-speeches to the pipeline schema."""
    text_clean = clean_and_normalize(row["text"])
    
    # Derive chamber
    src_lower = str(row["source"]).lower()
    url_lower = str(row.get("source_url", "")).lower()
    role_lower = str(row.get("speaker_role", "")).lower()
    
    if "senat" in src_lower or "senat" in url_lower or "sénat" in role_lower:
        chamber = "Sénat"
    else:
        chamber = "Assemblée nationale"
        
    # Derive party based on role or name, otherwise fallback
    party = "Non spécifié"
    role_check = role_lower if pd.notna(role_lower) else ""
    spk_check = str(row["speaker"]).lower() if pd.notna(row["speaker"]) else ""
    
    if "la france insoumise" in role_check or "lfi" in role_check or "mélenchon" in spk_check or "ruffin" in spk_check:
        party = "La France Insoumise"
    elif "socialiste" in role_check or "faure" in spk_check:
        party = "Socialistes"
    elif "écologiste" in role_check or "vert" in role_check:
        party = "Écologistes"
    elif "ump" in role_check or "républicains" in role_check or "wauquiez" in spk_check:
        party = "Les Républicains"
    elif "rassemblement national" in role_check or "rn" in role_check or "le pen" in spk_check:
        party = "Rassemblement National"
    elif "renaissance" in role_check or "lrem" in role_check or "majorité" in role_check or "darmanin" in spk_check:
        party = "Renaissance"
    elif "démocrate" in role_check or "modem" in role_check:
        party = "Démocrates"
    else:
        random.seed(seed_idx)
        party = random.choice(PARTIES)
        
    # Calculate legislature
    date_str = str(row["date"])
    try:
        dt = pd.to_datetime(date_str)
        year = dt.year
        date_formatted = dt.strftime("%Y-%m-%d")
    except:
        year = 2000
        date_formatted = "2000-01-01"
        
    if year >= 2022:
        leg = "16"
    elif year >= 2017:
        leg = "15"
    elif year >= 2012:
        leg = "14"
    elif year >= 2007:
        leg = "13"
    elif year >= 2002:
        leg = "12"
    elif year >= 1997:
        leg = "11"
    else:
        leg = str(random.randint(1, 10))
        
    # Map document type
    doc_type = "intervention_seance"
    st_lower = str(row.get("speech_type", "")).lower() if pd.notna(row.get("speech_type", "")) else ""
    title_check = str(row.get("title", "")).lower() if pd.notna(row.get("title", "")) else ""
    
    if "amendement" in st_lower or "amendement" in title_check:
        doc_type = "amendement"
    elif "déclaration" in st_lower or "prise de position" in st_lower:
        doc_type = "prise_position"
    elif "explication" in st_lower:
        doc_type = "explication_vote"
    elif "réponse" in st_lower:
        doc_type = "reponse_debat"
    elif "groupe" in st_lower:
        doc_type = "discours_groupe"
        
    return {
        "text": text_clean,
        "label_human_ai": 0,
        "source": row["source"] if pd.notna(row["source"]) else "Assemblée/Sénat",
        "speaker": row["speaker"] if pd.notna(row["speaker"]) else "Orateur inconnu",
        "party": party,
        "date": date_formatted,
        "chamber": chamber,
        "document_type": doc_type,
        "legislature": leg
    }

def generate_mock_human_corpus(sample_size, seed=42):
    """Generates a realistic mock corpus of human political speeches."""
    print(f"Generating mock human political corpus ({sample_size} samples)...")
    random.seed(seed)
    np.random.seed(seed)
    
    data = []
    start_date = datetime(1958, 1, 1)
    end_date = datetime(2022, 12, 31)
    date_range_days = (end_date - start_date).days
    
    doc_types = ["intervention_seance", "prise_position", "explication_vote", "amendement", "reponse_debat", "discours_groupe"]
    sources = ["french-political-speeches-1958-2022", "French Political Speeches (2000-2010)", "Sénat XML", "Assemblée nationale open-data"]
    
    for i in range(sample_size):
        date_speech = start_date + timedelta(days=random.randint(0, date_range_days))
        speaker = random.choice(SPEAKERS)
        party = random.choice(PARTIES)
        chamber = random.choice(CHAMBERS)
        doc_type = random.choice(doc_types)
        source = random.choice(sources)
        
        # Calculate legislature
        year = date_speech.year
        if year >= 2022:
            leg = "16"
        elif year >= 2017:
            leg = "15"
        elif year >= 2012:
            leg = "14"
        elif year >= 2007:
            leg = "13"
        elif year >= 2002:
            leg = "12"
        else:
            leg = str(random.randint(1, 11))
            
        speech_text = generate_speech(is_ai=False, seed=i)
        cleaned_text = clean_and_normalize(speech_text)
        
        data.append({
            "text": cleaned_text,
            "label_human_ai": 0,
            "source": source,
            "speaker": speaker,
            "party": party,
            "date": date_speech.strftime("%Y-%m-%d"),
            "chamber": chamber,
            "document_type": doc_type,
            "legislature": leg
        })
        
    return pd.DataFrame(data)

def generate_recent_debates_from_real(df_hf, sample_size=3000, seed=42):
    """Creates recent debates using real post-2003 human texts from HF dataset and injects post-2022 AI texts."""
    print(f"Creating recent debates from Hugging Face dataset (2003-2026) ({sample_size} samples)...")
    random.seed(seed + 100)
    np.random.seed(seed + 100)
    
    # Filter real speeches post 2003
    df_hf["parsed_date"] = pd.to_datetime(df_hf["date"], errors="coerce")
    df_recent_human = df_hf[df_hf["parsed_date"] >= "2003-01-01"].copy()
    
    if len(df_recent_human) < sample_size:
        print("Warning: not enough post-2003 real speeches, utilizing all of them and falling back for remaining.")
        df_sampled = df_recent_human
    else:
        df_sampled = df_recent_human.sample(n=sample_size, random_state=seed + 200)
        
    start_date_modern = datetime(2023, 1, 1)
    end_date_modern = datetime(2026, 5, 1)
    date_range_days_modern = (end_date_modern - start_date_modern).days
    
    start_date_hist = datetime(2004, 1, 1)
    end_date_hist = datetime(2022, 12, 31)
    date_range_days_hist = (end_date_hist - start_date_hist).days
    
    data = []

    for idx, (original_idx, row) in enumerate(df_sampled.iterrows()):
        mapped = map_row_to_schema(row, idx)
        
        # Spreading: project dates to cover the entire timeline continuously
        if random.random() < 0.20:
            # Project to modern era (2023-2026) where AI injection can happen
            random_days = random.randint(0, date_range_days_modern)
            dt_modern = start_date_modern + timedelta(days=random_days)
            mapped["date"] = dt_modern.strftime("%Y-%m-%d")
            mapped["legislature"] = "16"
        else:
            # Spread historical human debates uniformly across the 2004-2022 range
            random_days = random.randint(0, date_range_days_hist)
            dt_hist = start_date_hist + timedelta(days=random_days)
            mapped["date"] = dt_hist.strftime("%Y-%m-%d")
            
            # Map legislature based on year
            year = dt_hist.year
            if year >= 2022:
                mapped["legislature"] = "16"
            elif year >= 2017:
                mapped["legislature"] = "15"
            elif year >= 2012:
                mapped["legislature"] = "14"
            elif year >= 2007:
                mapped["legislature"] = "13"
            else:
                mapped["legislature"] = "12"
        
        dt = pd.to_datetime(mapped["date"])
        is_ai = False
        ai_model = "human"
        
        if dt.year >= 2023:
            # 15% probability of AI post 2022
            prob = 0.15
            if mapped["speaker"] == "Jean Dupuis" or "dupuis" in str(mapped["speaker"]).lower():
                prob = 0.40
            if mapped["party"] == "Démocrates":
                prob = 0.30
                
            if random.random() < prob:
                is_ai = True
                ai_model = random.choice(["gpt-4", "claude-3-opus", "qwen-72b", "gemma-7b"])
                
        if is_ai:
            speech_text = generate_speech(is_ai=True, ai_model=ai_model, doc_type=mapped["document_type"], seed=idx + 20000)
            mapped["text"] = clean_and_normalize(speech_text)
            mapped["actual_label"] = 1
            mapped["ai_model"] = ai_model
        else:
            mapped["actual_label"] = 0
            mapped["ai_model"] = "human"
            
        data.append(mapped)
        
    # Fill up to sample_size if needed
    current_len = len(data)
    if current_len < sample_size:
        print(f"Adding {sample_size - current_len} simulated recent speeches to reach sample size...")
        start_date = datetime(2003, 1, 1)
        end_date = datetime(2026, 5, 1)
        date_range_days = (end_date - start_date).days
        for i in range(sample_size - current_len):
            date_speech = start_date + timedelta(days=random.randint(0, date_range_days))
            speaker = random.choice(SPEAKERS)
            party = random.choice(PARTIES)
            chamber = random.choice(CHAMBERS)
            doc_type = random.choice(["intervention_seance", "prise_position", "explication_vote", "amendement"])
            
            is_ai = False
            ai_model = "human"
            if date_speech.year >= 2023:
                prob = 0.15
                if speaker == "Jean Dupuis":
                    prob = 0.40
                if party == "Démocrates":
                    prob = 0.30
                if random.random() < prob:
                    is_ai = True
                    ai_model = random.choice(["gpt-4", "claude-3-opus", "qwen-72b", "gemma-7b"])
                    
            leg = "16" if date_speech.year >= 2022 else ("15" if date_speech.year >= 2017 else ("14" if date_speech.year >= 2012 else "13"))
            speech_text = generate_speech(is_ai=is_ai, ai_model=ai_model, doc_type=doc_type, seed=i + 30000)
            cleaned = clean_and_normalize(speech_text)
            
            data.append({
                "text": cleaned,
                "label_human_ai": 0,
                "source": "simulated_debate",
                "speaker": speaker,
                "party": party,
                "date": date_speech.strftime("%Y-%m-%d"),
                "chamber": chamber,
                "document_type": doc_type,
                "legislature": leg,
                "actual_label": 1 if is_ai else 0,
                "ai_model": ai_model
            })
            
    return pd.DataFrame(data)

def generate_recent_debates_mock(sample_size=3000, seed=42):
    """Fallback generator for recent debates using mock logic."""
    print(f"Generating mock recent debates corpus for inference ({sample_size} samples)...")
    random.seed(seed + 100)
    np.random.seed(seed + 100)
    
    data = []
    start_date = datetime(2003, 1, 1)
    end_date = datetime(2026, 5, 1)
    date_range_days = (end_date - start_date).days
    doc_types = ["intervention_seance", "prise_position", "explication_vote", "amendement", "reponse_debat", "discours_groupe"]
    
    for i in range(sample_size):
        date_speech = start_date + timedelta(days=random.randint(0, date_range_days))
        speaker = random.choice(SPEAKERS)
        party = random.choice(PARTIES)
        chamber = random.choice(CHAMBERS)
        doc_type = random.choice(doc_types)
        
        is_ai = False
        ai_model = "human"
        
        if date_speech.year >= 2023:
            prob = 0.12
            if speaker == "Jean Dupuis":
                prob = 0.35
            if party == "Démocrates":
                prob = 0.25
            if random.random() < prob:
                is_ai = True
                ai_model = random.choice(["gpt-4", "claude-3-opus", "qwen-72b", "gemma-7b"])
                
        leg = "16" if date_speech.year >= 2022 else ("15" if date_speech.year >= 2017 else ("14" if date_speech.year >= 2012 else ("13" if date_speech.year >= 2007 else "12")))
        speech_text = generate_speech(is_ai=is_ai, ai_model=ai_model, doc_type=doc_type, seed=i + 5000)
        cleaned_text = clean_and_normalize(speech_text)
        
        data.append({
            "text": cleaned_text,
            "speaker": speaker,
            "party": party,
            "date": date_speech.strftime("%Y-%m-%d"),
            "chamber": chamber,
            "document_type": doc_type,
            "legislature": leg,
            "actual_label": 1 if is_ai else 0,
            "ai_model": ai_model
        })
        
    return pd.DataFrame(data)

def main():
    parser = argparse.ArgumentParser(description="Collect, ingest and clean French parliamentary speeches.")
    parser.add_argument("--config", default="configs/config.yaml", help="Path to config file")
    parser.add_argument("--mock", action="store_true", help="Force generating mock datasets")
    args = parser.parse_args()
    
    config = load_config(args.config)
    raw_dir = config["paths"]["raw_dir"]
    os.makedirs(raw_dir, exist_ok=True)
    
    use_mock = config["data_collection"]["use_mock"] or args.mock
    seed = config["data_collection"]["seed"]
    sample_size = config["data_collection"]["sample_size_human"]
    
    df_hf = None
    
    if not use_mock:
        try:
            print("Loading real human political speeches from Hugging Face (agokrani/fr-political-speeches)...")
            from datasets import load_dataset
            dataset = load_dataset("agokrani/fr-political-speeches", split="train")
            df_hf = dataset.to_pandas()
            print(f"Successfully downloaded {len(df_hf)} real speeches from Hugging Face.")
            
            # Map a sample of real speeches to training human corpus
            df_sampled = df_hf.sample(n=min(sample_size, len(df_hf)), random_state=seed)
            
            print("Mapping real dataset to unified schema...")
            data_mapped = []
            for i, (_, row) in enumerate(df_sampled.iterrows()):
                data_mapped.append(map_row_to_schema(row, i))
                
            df_human = pd.DataFrame(data_mapped)
            df_human.to_csv(os.path.join(raw_dir, "human_corpus.csv"), index=False)
            print(f"Saved real human training corpus to {os.path.join(raw_dir, 'human_corpus.csv')}")
            
        except Exception as e:
            print(f"Error downloading or processing from Hugging Face: {e}. Falling back to mock generation.")
            use_mock = True
            
    if use_mock:
        df_human = generate_mock_human_corpus(sample_size, seed)
        df_human.to_csv(os.path.join(raw_dir, "human_corpus.csv"), index=False)
        print(f"Saved mock human corpus to {os.path.join(raw_dir, 'human_corpus.csv')}")
        
    # Generate the AI synthetic corpus
    import subprocess
    script_dir = os.path.dirname(os.path.abspath(__file__))
    synthetic_script = os.path.join(script_dir, "generate_synthetic.py")
    print(f"Executing AI generation script: {synthetic_script}...")
    subprocess.run([sys.executable, synthetic_script, "--config", args.config], check=True)
        
    # Generate the recent debates corpus for inference
    if not use_mock and df_hf is not None:
        df_recent = generate_recent_debates_from_real(df_hf, sample_size=3000, seed=seed)
    else:
        df_recent = generate_recent_debates_mock(sample_size=3000, seed=seed)
        
    df_recent.to_csv(os.path.join(raw_dir, "recent_debates.csv"), index=False)
    print(f"Saved recent debates for inference to {os.path.join(raw_dir, 'recent_debates.csv')}")
    print("Ingestion step completed successfully.")

if __name__ == "__main__":
    main()