Add analyze_vectors.py

analyze_vectors.py

@@ -0,0 +1,148 @@
+#!/usr/bin/env python3
+"""Analyze extracted emotion vectors: cosine similarity, clustering, visualization."""
+
+import json
+import os
+import numpy as np
+
+EXP_DIR = os.path.dirname(os.path.abspath(__file__))
+RESULTS_DIR = os.path.join(EXP_DIR, "results")
+
+
+def load_vectors():
+    vectors_file = os.path.join(RESULTS_DIR, "emotion_vectors.npz")
+    data = np.load(vectors_file)
+    return {name: data[name] for name in data.files}
+
+
+def cosine_sim(a, b):
+    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b) + 1e-8)
+
+
+def cosine_similarity_matrix(vectors):
+    emotions = sorted(vectors.keys())
+    n = len(emotions)
+    matrix = np.zeros((n, n))
+    for i, e1 in enumerate(emotions):
+        for j, e2 in enumerate(emotions):
+            matrix[i, j] = cosine_sim(vectors[e1], vectors[e2])
+    return emotions, matrix
+
+
+def print_similarity_matrix(emotions, matrix):
+    print("\n=== Cosine Similarity Matrix ===\n")
+    # Header
+    header = f"{'':12s}" + "".join(f"{e[:6]:>7s}" for e in emotions)
+    print(header)
+    for i, e in enumerate(emotions):
+        row = f"{e:12s}" + "".join(f"{matrix[i,j]:7.2f}" for j in range(len(emotions)))
+        print(row)
+
+
+def find_clusters(emotions, matrix, threshold=0.5):
+    """Find emotion pairs with high similarity."""
+    print(f"\n=== High Similarity Pairs (>{threshold}) ===\n")
+    pairs = []
+    for i in range(len(emotions)):
+        for j in range(i + 1, len(emotions)):
+            if matrix[i, j] > threshold:
+                pairs.append((emotions[i], emotions[j], matrix[i, j]))
+    pairs.sort(key=lambda x: -x[2])
+    for e1, e2, sim in pairs:
+        print(f"  {e1:12s} <-> {e2:12s}  sim={sim:.3f}")
+    if not pairs:
+        print("  (none found)")
+    return pairs
+
+
+def find_opposites(emotions, matrix, threshold=-0.3):
+    """Find emotion pairs with negative similarity (opposites)."""
+    print(f"\n=== Opposite Pairs (<{threshold}) ===\n")
+    pairs = []
+    for i in range(len(emotions)):
+        for j in range(i + 1, len(emotions)):
+            if matrix[i, j] < threshold:
+                pairs.append((emotions[i], emotions[j], matrix[i, j]))
+    pairs.sort(key=lambda x: x[2])
+    for e1, e2, sim in pairs:
+        print(f"  {e1:12s} <-> {e2:12s}  sim={sim:.3f}")
+    if not pairs:
+        print("  (none found)")
+    return pairs
+
+
+def valence_arousal_check(emotions, pca_results):
+    """Check if PC1≈valence, PC2≈arousal based on known emotion groupings."""
+    print("\n=== Valence-Arousal Structure Check ===\n")
+
+    positive = {"happy", "proud", "inspired", "loving", "hopeful", "calm", "playful"}
+    negative = {"sad", "angry", "afraid", "desperate", "guilty", "disgusted", "lonely", "spiteful"}
+    high_arousal = {"angry", "afraid", "surprised", "desperate", "inspired", "nervous", "anxious"}
+    low_arousal = {"calm", "sad", "brooding", "lonely", "guilty"}
+
+    for pc_name, pc_vals in [("PC1", pca_results["pc1"]), ("PC2", pca_results["pc2"])]:
+        pos_vals = [pc_vals[i] for i, e in enumerate(pca_results["emotions"]) if e in positive]
+        neg_vals = [pc_vals[i] for i, e in enumerate(pca_results["emotions"]) if e in negative]
+        hi_vals = [pc_vals[i] for i, e in enumerate(pca_results["emotions"]) if e in high_arousal]
+        lo_vals = [pc_vals[i] for i, e in enumerate(pca_results["emotions"]) if e in low_arousal]
+
+        pos_mean = np.mean(pos_vals) if pos_vals else 0
+        neg_mean = np.mean(neg_vals) if neg_vals else 0
+        hi_mean = np.mean(hi_vals) if hi_vals else 0
+        lo_mean = np.mean(lo_vals) if lo_vals else 0
+
+        valence_sep = abs(pos_mean - neg_mean)
+        arousal_sep = abs(hi_mean - lo_mean)
+
+        print(f"  {pc_name}:")
+        print(f"    Positive mean: {pos_mean:+.3f}  Negative mean: {neg_mean:+.3f}  → Valence separation: {valence_sep:.3f}")
+        print(f"    High-A mean:   {hi_mean:+.3f}  Low-A mean:    {lo_mean:+.3f}  → Arousal separation: {arousal_sep:.3f}")
+
+        if valence_sep > arousal_sep:
+            print(f"    → {pc_name} ≈ VALENCE axis")
+        else:
+            print(f"    → {pc_name} ≈ AROUSAL axis")
+
+
+def main():
+    print("=== Emotion Vector Analysis ===\n")
+
+    # Load vectors
+    vectors = load_vectors()
+    print(f"Loaded {len(vectors)} emotion vectors")
+    print(f"Vector dimension: {next(iter(vectors.values())).shape[0]}")
+
+    # Load experiment results for PCA
+    results_file = os.path.join(RESULTS_DIR, "experiment_results.json")
+    with open(results_file, "r") as f:
+        results = json.load(f)
+
+    # Similarity analysis
+    emotions, matrix = cosine_similarity_matrix(vectors)
+    print_similarity_matrix(emotions, matrix)
+    find_clusters(emotions, matrix, threshold=0.4)
+    find_opposites(emotions, matrix, threshold=-0.2)
+
+    # Valence-Arousal check
+    if "pca" in results:
+        valence_arousal_check(emotions, results["pca"])
+
+    # Summary
+    print("\n=== SUMMARY ===\n")
+    avg_sim = matrix[np.triu_indices_from(matrix, k=1)].mean()
+    print(f"  Average pairwise similarity: {avg_sim:.3f}")
+    print(f"  Variance explained by PC1+PC2: {(results['pca']['explained_variance_pc1'] + results['pca']['explained_variance_pc2'])*100:.1f}%")
+
+    # Anthropic found ~30% variance in first 2 PCs for 171 emotions
+    # With 20 emotions, we'd expect higher concentration
+    var_12 = results['pca']['explained_variance_pc1'] + results['pca']['explained_variance_pc2']
+    if var_12 > 0.3:
+        print("  ✓ Strong 2D structure detected (>30% in PC1+PC2)")
+    else:
+        print("  ✗ Weak 2D structure (<30% in PC1+PC2)")
+
+    print("\n=== ANALYSIS COMPLETE ===")
+
+
+if __name__ == "__main__":
+    main()