Make build_graph output stable

poligrapher/scripts/build_graph.py

@@ -262,14 +262,16 @@ def normalize_terms():
             """Step 6: Run phrase normalization."""
 
             for src, token_type in token_type_map.items():
-                normalized_terms_map[src] = terms = set()
-
                 if (phrase := _expand_phrase(src)) is None:
+                    normalized_terms_map[src] = []
                     continue
 
                 # Fallback to lemmatization only if token types from NER and graph relation agrees
                 flag_use_stem = phrase.root.ent_type_ == token_type
 
+                # All candidate terms
+                terms = set()
+
                 match token_type:
                     case "DATA":
                         terms.update(self.data_phrase_normalizer.normalize(phrase, flag_use_stem))
@@ -318,7 +320,8 @@ def normalize_terms():
 
                 G_final.add_nodes_from(terms, type=token_type)
 
-                logging.info("Phrase %r (%s) -> %r", phrase.text, token_type, ", ".join(sorted(terms)))
+                normalized_terms_map[src] = sorted(terms, reverse=True)  # Stablize set order for reproducibility
+                logging.info("Phrase %r (%s) -> %r", phrase.text, token_type, ", ".join(normalized_terms_map[src]))
 
         def merge_subsum_graph():
             """Step 7: Populate SUBSUM edges in G_final from G_subsum."""