Accept titles with high word-level Jaccard similarity (>=0.85)

Handles cases like bib "real-GitHub issues" vs SS "Real-World GitHub Issues"
where splitting on hyphens makes all bib words a subset of the SS title.
Adds _title_word_jaccard helper and a test case for this pattern.

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

Author: monperrus

check_hallucinations/main.py

@@ -7,6 +7,7 @@
 import hashlib
 import json
 import os
+import re
 import time
 import unicodedata
 
@@ -45,6 +46,14 @@ def _normalize_title(title):
     )
 
 
+def _title_word_jaccard(a, b):
+    wa = set(re.split(r'[-\s?!,.:;]+', a.lower())) - {""}
+    wb = set(re.split(r'[-\s?!,.:;]+', b.lower())) - {""}
+    if not wa or not wb:
+        return 0.0
+    return len(wa & wb) / len(wa | wb)
+
+
 def _cache_path(subdir, key):
     digest = hashlib.sha256(key.encode("utf-8")).hexdigest()
     path = os.path.join(CACHE_DIR, subdir)
@@ -109,7 +118,8 @@ def get_url_from_title(title):
     if not result or "paperId" not in result:
         raise SemanticScholarNotFound("No data found for title: " + title)
     ss_main = _normalize_title(result["title"].split(":")[0])
-    if _normalize_title(result["title"]) == _normalize_title(title) or ss_main == _normalize_title(title):
+    similar_enough = _title_word_jaccard(result["title"], title) >= 0.85
+    if _normalize_title(result["title"]) == _normalize_title(title) or ss_main == _normalize_title(title) or similar_enough:
         data = _get_paper_info(result["paperId"])
         if "externalIds" in data:
             if "DOI" in data["externalIds"]:

test_normalize.py

@@ -1,4 +1,4 @@
-from check_hallucinations.main import _normalize_title
+from check_hallucinations.main import _normalize_title, _title_word_jaccard
 
 
 def test_colon_vs_period_subtitle():
@@ -18,3 +18,11 @@ def test_diacritics():
 
 def test_curly_braces():
     assert _normalize_title("{Foo} Bar") == _normalize_title("Foo Bar")
+
+
+def test_word_jaccard_hyphen_split():
+    # bib: "real-GitHub" splits as "real"+"github"; SS has "Real-World GitHub"
+    # all bib words are in SS → Jaccard 0.9
+    bib = "SWE-bench: Can language models resolve real-GitHub issues?"
+    ss  = "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?"
+    assert _title_word_jaccard(bib, ss) >= 0.85