Geoparsing: Nominatim-Plausibilitätsprüfung + regionale Stopwords

- SequenceMatcher prüft ob Nominatim-Ergebnis zum Suchbegriff passt (Schwelle 0.3)
- Verwirft Fehlzuordnungen wie 'Golf-Staaten' -> Uganda
- Regionale/vage Begriffe als Stopwords (Naher Osten, Golf-Staaten, Balkan etc.)
- Falscher DB-Eintrag (Botschafter-Residenz Uganda) bereinigt

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/agents/geoparsing.py

@@ -2,6 +2,7 @@
 import asyncio
 import logging
 import re
+from difflib import SequenceMatcher
 from typing import Optional
 
 logger = logging.getLogger("osint.geoparsing")
@@ -19,12 +20,70 @@ LOCATION_STOPWORDS = {
     "Twitter", "Facebook", "Telegram", "Signal", "WhatsApp",
     "Reuters", "AP", "AFP", "DPA", "dpa",
     "Internet", "Online", "Web",
+    # Regionale/vage Begriffe (kein einzelner Punkt auf der Karte)
+    "Naher Osten", "Mittlerer Osten", "Middle East", "Near East",
+    "Golf-Staaten", "Golfstaaten", "Golfregion", "Gulf States", "Persian Gulf",
+    "Nordafrika", "Subsahara", "Zentralasien", "Suedostasien",
+    "Westeuropa", "Osteuropa", "Suedeuropa", "Nordeuropa",
+    "Balkan", "Kaukasus", "Levante", "Maghreb", "Sahel",
+    "Arabische Welt", "Arab World",
 }
 
 # Maximale Textlaenge fuer NER-Verarbeitung
 MAX_TEXT_LENGTH = 10000
 
 
+# Marker-Kategorien fuer Karten-Klassifizierung
+CATEGORY_KEYWORDS = {
+    "target": [
+        "angriff", "angegriff", "bombardier", "luftschlag", "luftangriff",
+        "beschuss", "beschossen", "getroffen", "zerstoer", "einschlag",
+        "detonation", "explosion", "strike", "attack", "bombed", "hit",
+        "shelled", "destroyed", "targeted", "missile hit", "air strike",
+        "airstrike", "bombardment", "killed", "casualties", "dead",
+        "tote", "opfer", "getoetet",
+    ],
+    "retaliation": [
+        "gegenschlag", "vergeltung", "reaktion", "gegenangriff",
+        "abgefeuert", "retaliat", "counter-attack", "counterattack",
+        "counter-strike", "response", "fired back", "launched",
+        "rakete abgefeuert", "vergeltungsschlag", "abfangen",
+        "abgefangen", "intercepted", "eskalation", "escalat",
+    ],
+    "actor": [
+        "regierung", "praesident", "ministerium", "hauptquartier",
+        "kommando", "nato", "pentagon", "kongress", "senat", "parlament",
+        "government", "president", "ministry", "headquarters", "command",
+        "congress", "senate", "parliament", "white house", "weisses haus",
+        "verteidigungsminister", "aussenminister", "generalstab",
+        "defense secretary", "secretary of state", "general staff",
+        "un-sicherheitsrat", "security council", "summit", "gipfel",
+        "diplomati", "botschaft", "embassy",
+    ],
+}
+
+
+def _classify_location(source_text: str, article_text: str = "") -> str:
+    """Klassifiziert eine Location basierend auf dem Kontext.
+
+    Returns:
+        Kategorie: 'target', 'retaliation', 'actor', oder 'mentioned'
+    """
+    text = (source_text + " " + article_text[:500]).lower()
+
+    scores = {"target": 0, "retaliation": 0, "actor": 0}
+    for category, keywords in CATEGORY_KEYWORDS.items():
+        for kw in keywords:
+            if kw in text:
+                scores[category] += 1
+
+    best = max(scores, key=scores.get)
+    if scores[best] >= 1:
+        return best
+    return "mentioned"
+
+
+
 def _load_spacy_model(lang: str):
     """Laedt ein spaCy-Modell lazy (nur beim ersten Aufruf)."""
     global _nlp_de, _nlp_en
@@ -206,11 +265,19 @@ def _geocode_nominatim(name: str) -> Optional[dict]:
             if "address" in raw:
                 country_code = raw["address"].get("country_code", "").upper()
 
+            normalized_name = location.address.split(",")[0] if location.address else name
+
+            # Plausibilitaetspruefung: Nominatim-Ergebnis muss zum Suchbegriff passen
+            similarity = SequenceMatcher(None, name.lower(), normalized_name.lower()).ratio()
+            if similarity < 0.3:
+                logger.debug(f"Nominatim-Ergebnis verworfen: '{name}' -> '{normalized_name}' (Aehnlichkeit {similarity:.2f})")
+                return None
+
             return {
                 "lat": float(location.latitude),
                 "lon": float(location.longitude),
                 "country_code": country_code,
-                "normalized_name": location.address.split(",")[0] if location.address else name,
+                "normalized_name": normalized_name,
                 "confidence": 0.4,  # Nominatim-Ergebnis = niedrigere Konfidenz
             }
     except (GeocoderTimedOut, GeocoderServiceError) as e:
@@ -276,6 +343,7 @@ async def geoparse_articles(articles: list[dict]) -> dict[int, list[dict]]:
         for loc in locations_raw:
             geo = await asyncio.to_thread(_geocode_location, loc["name"])
             if geo:
+                category = _classify_location(loc.get("source_text", ""), text)
                 locations.append({
                     "location_name": loc["name"],
                     "location_name_normalized": geo["normalized_name"],
@@ -284,6 +352,7 @@ async def geoparse_articles(articles: list[dict]) -> dict[int, list[dict]]:
                     "lon": geo["lon"],
                     "confidence": geo["confidence"],
                     "source_text": loc.get("source_text", ""),
+                    "category": category,
                 })
 
         if locations: