1 files changed, 36 insertions, 6 deletions
diff --git a/worldalign/natural_pipeline.py b/worldalign/natural_pipeline.py
index 1a4f63e..1844a6d 100644
--- a/worldalign/natural_pipeline.py
+++ b/worldalign/natural_pipeline.py
@@ -53,6 +53,9 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--views", type=int, default=1)
     parser.add_argument("--moment-degree", type=int, default=2, choices=[2, 3])
     parser.add_argument("--third-width", type=int, default=12)
+    parser.add_argument(
+        "--phrase-encoding", default="mean", choices=["mean", "structured"]
+    )
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--output", default="artifacts/vg_5k/natural_pipeline.pt")
     return parser.parse_args()
@@ -160,13 +163,39 @@ def main() -> None:
 
     vectors = word_vectors(load_phrases(vg_dir / "text_nodes.jsonl", "region_closed"), args)
 
+    width = args.word_vectors
+
+    def encode_phrase(phrase: str) -> np.ndarray | None:
+        """One vector per region description.
+
+        Averaging every word of a phrase collapses "red bus" and "bus red"
+        and, worse, mixes the thing named with what is said about it. The
+        synthetic world never had to face this because its captions were
+        encoded into separate factor blocks. Splitting the head from its
+        modifiers recovers part of that structure from English word order
+        alone -- the last token of an English noun phrase is its head -- which
+        is a fact about one language, not a claim about the other modality,
+        so it stays inside Tier 0.
+        """
+        tokens = [t for t in re.findall(r"[a-z]+", phrase.lower()) if t in vectors]
+        if not tokens:
+            return None
+        if args.phrase_encoding == "mean":
+            return np.mean([vectors[t] for t in tokens], axis=0)
+        head = vectors[tokens[-1]]
+        modifiers = (
+            np.mean([vectors[t] for t in tokens[:-1]], axis=0)
+            if len(tokens) > 1
+            else np.zeros(width)
+        )
+        return np.concatenate([head, modifiers])
+
     def language_state(node: str) -> np.ndarray | None:
-        rows = []
-        for phrase in records[node]["region_closed"]:
-            hits = [vectors[token] for token in re.findall(r"[a-z]+", phrase.lower())
-                    if token in vectors]
-            if hits:
-                rows.append(np.mean(hits, axis=0))
+        rows = [
+            row
+            for row in (encode_phrase(p) for p in records[node]["region_closed"])
+            if row is not None
+        ]
         return np.stack(rows) if rows else None
 
     def vision_state(node: str, view: int) -> np.ndarray | None:
@@ -245,6 +274,7 @@ def main() -> None:
         "kept_directions": keep,
         "weight_power": args.weight_power,
         "moment_degree": args.moment_degree,
+        "phrase_encoding": args.phrase_encoding,
         "field_correlation_at_truth": correlation,
         "recovery_threshold": 0.9,
     }