epsilon-compositionality

commit 1569435e4c9674c02987d325b9baed018bf819f0
parent d7b80318340fa4315f6c693ffd592ebae8e6fdee
Author: Sondre Wold <sondrewold1@gmail.com>
Date:   Mon, 21 Jul 2025 16:26:01 +0200

remove word2vec

Diffstat:
M
experiment.py
 | 
35
-----------------------------------

1 file changed, 0 insertions(+), 35 deletions(-)
diff --git a/experiment.py b/experiment.py
@@ -12,8 +12,6 @@ import torch
 import pickle
 import tqdm
 
-from gensim.models import KeyedVectors
-
 from transformers import AutoTokenizer, AutoModel
 from sentence_transformers import SentenceTransformer
 
@@ -95,39 +93,6 @@ class ModelFactory:
         return cls._factory[name](*args, **kwargs)
 
 
-# @ModelFactory("Word2Vec")
-class Word2vec(Model):
-    """
-    It does not make conceptually sense to test Word2Vec, as sim(\mu(ac), \mu(bc)) will 
-    be equal to sim(\mu(a), \mu(b)) by definition if \mu(ac) 
-    is the sum of the embeddings of the subwords in a and c. 
-
-    But, maybe we can do something else with it, so here it is.
-    """
-
-    def __init__(self) -> None:
-        self.model_path: str = "./models/GoogleNews-vectors-negative300.bin.gz"
-        self.encoder = KeyedVectors.load_word2vec_format(self.model_path, binary=True)
-
-
-    def embed(self, string: str) -> torch.Tensor:
-        words: list[str] = string.split()
-
-        # In He et.al (2025), they ignore OOV word word2vec.
-        present_words = [word for word in words if word in self.encoder]
-
-        embeddings = [self.encoder[word] for word in present_words]
-        sentence_embedding = numpy.sum(embeddings, axis=0)
-        return torch.tensor(sentence_embedding)
-
-
-    def calculate_similarity(self, a: torch.Tensor, b: torch.Tensor) -> float:
-        return torch.nn.functional.cosine_similarity(
-                a, b, dim=0
-        ).item()
-
-
-
 @ModelFactory("BERT_cls")
 class BERT_cls(Model):