epsilon-compositionality

commit 1e6e244f0400af38515a26df6b321ae32f44cb0b
parent d5509b1a6efa4fd84eef40c2b61a19fd46eab8a1
Author: Sondre Wold <sondrewold1@gmail.com>
Date:   Sat, 12 Jul 2025 20:19:37 +0200

add plot script

Diffstat:
M
experiment.py
 | 
4
++--
A
load_and_plot.py
 | 
38
++++++++++++++++++++++++++++++++++++++

2 files changed, 40 insertions(+), 2 deletions(-)
diff --git a/experiment.py b/experiment.py
@@ -131,10 +131,10 @@ def find_diff(a: str, b: str) -> list[str] | None:
             break
 
     diff_a = words_a[
-        prefix_len:len(words_a) - suffix_len if suffix_len > 0 else len(words_a)
+        prefix_len : len(words_a) - suffix_len if suffix_len > 0 else len(words_a)
     ]
     diff_b = words_b[
-        prefix_len:len(words_b) - suffix_len if suffix_len > 0 else len(words_b)
+        prefix_len : len(words_b) - suffix_len if suffix_len > 0 else len(words_b)
     ]
 
     if not diff_a or not diff_b:
diff --git a/load_and_plot.py b/load_and_plot.py
@@ -0,0 +1,38 @@
+import sys
+import pickle
+import numpy
+from torch import histogram
+from experiment import Point
+import matplotlib.pyplot as plt
+
+if __name__ == "__main__":
+    points_path: str = sys.argv[1]
+    with open(points_path, "rb") as file:
+        points: list[Point] = pickle.load(file)
+        n: int = len(points)
+        word_sims: numpy.ndarray  = numpy.array([point.phrase_similarity for point in points])
+        context_sims: numpy.ndarray = numpy.array(
+            [point.contextualized_similarity for point in points]
+        )
+
+        epsilons: list[float] = []
+        for i in range(len(word_sims)):
+            if word_sims[i] > 0:
+                epsilons.append(1 - context_sims[i] / word_sims[i])
+        final_epsilon: float = float(numpy.percentile(epsilons, 80))
+
+        plt.scatter((1 - final_epsilon) * word_sims, context_sims)
+        xpoints = ypoints = plt.xlim()
+        plt.plot(
+            xpoints,
+            ypoints,
+            linestyle="--",
+            color="k",
+            lw=3,
+            scalex=False,
+            scaley=False,
+        )
+
+        plt.xlabel("Word similarity")
+        plt.ylabel("Contextualized similarity")
+        plt.show()