podcast-mindmap/scripts/analyse_arguments.py

#!/usr/bin/env python3
"""#13 Argumentketten-Tracker: Klassifiziere logische Relationen zwischen semantisch ähnlichen Absätzen.

Nimmt die Top-N semantic_links und lässt Qwen die Relation klassifizieren:
erweitert, widerspricht, belegt, relativiert, gleicher_punkt, kein_bezug.

Nutzung:
    DASHSCOPE_API_KEY=... python3 analyse_arguments.py [db-pfad] [limit]
"""

import json
import os
import sys
import time
import sqlite3

from openai import OpenAI

DB_PATH = sys.argv[1] if len(sys.argv) > 1 else "data/db.sqlite"
LIMIT = int(sys.argv[2]) if len(sys.argv) > 2 else 500

API_KEY = os.environ.get("DASHSCOPE_API_KEY", "")
BASE_URL = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
MODEL = "qwen-plus"

SYSTEM_PROMPT = """Du bist ein Diskursanalyst. Du erhältst zwei Textabschnitte aus Podcast-Transkripten.
Klassifiziere die logische Relation zwischen ihnen. Antworte NUR mit einem JSON-Objekt:

{"relation": "...", "confidence": 0.0-1.0, "explanation": "Ein Satz Begründung"}

Mögliche Relationen:
- "erweitert": B baut auf A auf, ergänzt, vertieft
- "widerspricht": B widerspricht A, nennt Gegenargument
- "belegt": B liefert Evidenz/Daten für A's These
- "relativiert": B schränkt A ein, nennt Ausnahmen/Bedingungen
- "gleicher_punkt": A und B sagen im Kern dasselbe
- "kein_bezug": Trotz thematischer Nähe kein logischer Bezug"""


def classify_pair(client, text_a, meta_a, text_b, meta_b):
    user_msg = f"""Absatz A ({meta_a}):
"{text_a}"

Absatz B ({meta_b}):
"{text_b}"

Welche logische Relation besteht von A zu B?"""

    try:
        resp = client.chat.completions.create(
            model=MODEL,
            messages=[
                {"role": "system", "content": SYSTEM_PROMPT},
                {"role": "user", "content": user_msg},
            ],
            temperature=0.1,
            max_tokens=150,
        )
        content = resp.choices[0].message.content.strip()
        # Parse JSON from response
        if content.startswith("```"):
            content = content.split("```")[1].strip()
            if content.startswith("json"):
                content = content[4:].strip()
        return json.loads(content)
    except Exception as e:
        return {"relation": "error", "confidence": 0, "explanation": str(e)}


def main():
    if not API_KEY:
        print("DASHSCOPE_API_KEY nicht gesetzt.")
        sys.exit(1)

    client = OpenAI(api_key=API_KEY, base_url=BASE_URL)
    db = sqlite3.connect(DB_PATH)
    db.row_factory = sqlite3.Row

    # Create output table
    db.executescript("""
    CREATE TABLE IF NOT EXISTS argument_links (
        id INTEGER PRIMARY KEY AUTOINCREMENT,
        source_podcast TEXT, source_episode TEXT, source_idx INTEGER,
        target_podcast TEXT, target_episode TEXT, target_idx INTEGER,
        relation TEXT, confidence REAL, explanation TEXT, score REAL
    );
    CREATE INDEX IF NOT EXISTS idx_arglinks ON argument_links(relation);
    """)

    # Get top semantic links (cross-episode, prefer cross-podcast)
    rows = db.execute("""
        SELECT sl.podcast_id, sl.source_episode, sl.source_idx,
               sl.target_podcast, sl.target_episode, sl.target_idx, sl.score,
               p1.text as source_text, p2.text as target_text,
               e1.title as source_title, e1.guest as source_guest,
               e2.title as target_title, e2.guest as target_guest
        FROM semantic_links sl
        JOIN paragraphs p1 ON sl.podcast_id = p1.podcast_id AND sl.source_episode = p1.episode_id AND sl.source_idx = p1.idx
        JOIN paragraphs p2 ON sl.target_podcast = p2.podcast_id AND sl.target_episode = p2.episode_id AND sl.target_idx = p2.idx
        JOIN episodes e1 ON sl.podcast_id = e1.podcast_id AND sl.source_episode = e1.id
        JOIN episodes e2 ON sl.target_podcast = e2.podcast_id AND sl.target_episode = e2.id
        WHERE sl.source_episode != sl.target_episode
        ORDER BY sl.score DESC
        LIMIT ?
    """, (LIMIT,)).fetchall()

    print(f"Klassifiziere {len(rows)} Paare mit {MODEL}…")

    # Check already processed
    existing = set()
    try:
        for r in db.execute("SELECT source_podcast||source_episode||source_idx||target_podcast||target_episode||target_idx as k FROM argument_links").fetchall():
            existing.add(r["k"])
    except Exception:
        pass

    processed = 0
    skipped = 0

    for i, row in enumerate(rows):
        key = f"{row['podcast_id']}{row['source_episode']}{row['source_idx']}{row['target_podcast']}{row['target_episode']}{row['target_idx']}"
        if key in existing:
            skipped += 1
            continue

        meta_a = f"{row['source_episode']}: {row['source_title']} — {row['source_guest']}"
        meta_b = f"{row['target_episode']}: {row['target_title']} — {row['target_guest']}"

        result = classify_pair(
            client,
            row["source_text"][:800], meta_a,
            row["target_text"][:800], meta_b
        )

        db.execute(
            "INSERT INTO argument_links (source_podcast, source_episode, source_idx, "
            "target_podcast, target_episode, target_idx, relation, confidence, explanation, score) "
            "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
            (row["podcast_id"], row["source_episode"], row["source_idx"],
             row["target_podcast"], row["target_episode"], row["target_idx"],
             result.get("relation", "error"), result.get("confidence", 0),
             result.get("explanation", ""), row["score"])
        )

        processed += 1
        if processed % 10 == 0:
            db.commit()
            print(f"  {processed}/{len(rows) - skipped} klassifiziert…")

        # Rate limiting
        time.sleep(0.3)

    db.commit()

    # Stats
    stats = db.execute("SELECT relation, COUNT(*) as c FROM argument_links GROUP BY relation ORDER BY c DESC").fetchall()
    print(f"\nFertig: {processed} neue, {skipped} übersprungen.")
    print("Verteilung:")
    for s in stats:
        print(f"  {s['relation']}: {s['c']}")

    db.close()


if __name__ == "__main__":
    main()