epsilon-compositionality

commit d7b80318340fa4315f6c693ffd592ebae8e6fdee
parent e6d15a7c6d4173253ae34d4ac3c99297eb7a796a
Author: Sondre Wold <sondrewold1@gmail.com>
Date:   Mon, 21 Jul 2025 16:20:49 +0200

update experiment class

Diffstat:
M
experiment.py
 | 
29
++++++++++++++++++++++++++---

1 file changed, 26 insertions(+), 3 deletions(-)
diff --git a/experiment.py b/experiment.py
@@ -26,7 +26,7 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("-data_path", type=pathlib.Path)
     parser.add_argument("-result_path", type=pathlib.Path)
     parser.add_argument("-normalization_value", type=float)
-    parser.add_argument("-experiment_type", type=str, choices=["sick", "ncip"])
+    parser.add_argument("-experiment_type", type=str, choices=["sick", "ncip", "ncip-non"])
     parser.add_argument("-model", type=str, choices=["BERT_cls", "BERT_mean", "S_BERT", "S_GPT", "E5", "QWEN"])
     return parser.parse_args()
 
@@ -50,6 +50,16 @@ class SentencePairs(NamedTuple):
     sentence_B: list[str]
 
 
+def normalize_v2():
+    """
+    For a context c and a substring x (to be replaced), replace x with n candidates, embed them, calculate the mean the std.dev.
+    Subtract the mean and divide by the std.dev (gaussian assumption) (mu(ac) - (mean(c with x)) / std.dev
+    
+    Sample x from a list of random words.
+    """
+    pass
+
+
 class Model(ABC):
 
     @abstractmethod
@@ -208,9 +218,14 @@ class E5(Model):
 
     def __init__(self):
         self.encoder = SentenceTransformer("intfloat/multilingual-e5-large-instruct", device=DEVICE)
+        self.task = "Given a sentence, encode its semantic properties"
+    
+    def get_detailed_instruct(self, query: str) -> str:
+        return f'Instruct: {self.task}\nQuery: {query}'
 
     def embed(self, string: str) -> torch.Tensor:
-        return self.encoder.encode([string])
+        prepared_string = self.get_detailed_instruct(string)
+        return self.encoder.encode([prepared_string])
 
     def calculate_similarity(self, a: torch.Tensor, b: torch.Tensor) -> float:
         return self.encoder.similarity(a, b)
@@ -373,7 +388,7 @@ class Experiment:
 def main() -> None:
     args = parse_args()
     sep = "\t"
-    if args.experiment_type == "ncip":
+    if args.experiment_type in ["ncip", "ncipv2"]:
         sep = ","
 
     df: pandas.DataFrame = pandas.read_csv(args.data_path, sep=sep)
@@ -389,6 +404,14 @@ def main() -> None:
         sentence_B: list[str] = (
             synonym_sentence + synonym_sentence2 + synonym_sentence3 + synonym_sentence4
         )
+    elif args.experiment_type == "ncip-non":
+        A: list[str] = df["synonym both"].values.tolist()
+        B: list[str]= df["both synonym alt2_1"].values.tolist()
+        C: list[str] = df["both synonym alt3_1"].values.tolist()
+        D: list[str] = df["both synonym alt4_1"].values.tolist()
+
+        sentence_A = A + B + C + D
+        sentence_B = B + C + D + A
     elif args.experiment_type == "sick":
         lexA = df.loc[df["sentence_A_expRule"] == "S2_lex"]
         lexB = df.loc[df["sentence_B_expRule"] == "S2_lex"]