File size: 8,561 Bytes
db764ae 9f87ec0 db764ae 9f87ec0 db764ae 9f87ec0 db764ae | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 | """
Word2Vec Baseline (gensim)
Trains a Word2Vec model on your corpus and provides the same interface
as the transformer engine, so you can compare results side by side.
Key limitation: Word2Vec gives ONE static vector per word regardless of
context. "pizza" always has the same embedding whether it means food or school.
The only contextual signal comes from averaging word vectors in a sentence.
Usage:
w2v = Word2VecEngine()
w2v.add_document("doc1", text)
w2v.build_index() # trains Word2Vec on your corpus
results = w2v.query("a place where children learn", top_k=5)
score = w2v.compare_texts("pizza gives me homework", "school gives me homework")
"""
import json
import re
import logging
from dataclasses import dataclass
from pathlib import Path
from typing import Optional
import numpy as np
from gensim.models import Word2Vec
logger = logging.getLogger(__name__)
@dataclass
class W2VResult:
"""A single similarity result."""
text: str
doc_id: str
score: float
rank: int
class Word2VecEngine:
"""
Word2Vec baseline for comparison with the transformer engine.
Trains Word2Vec on your corpus, represents sentences as averaged
word vectors, and uses cosine similarity for matching.
"""
def __init__(
self,
vector_size: int = 100,
window: int = 5,
min_count: int = 1,
epochs: int = 50,
sg: int = 1,
):
"""
Args:
vector_size: Dimensionality of word vectors.
window: Context window size.
min_count: Ignore words with frequency below this.
epochs: Training epochs.
sg: 1 for skip-gram, 0 for CBOW.
"""
self.vector_size = vector_size
self.window = window
self.min_count = min_count
self.epochs = epochs
self.sg = sg
self.model: Optional[Word2Vec] = None
self.sentences: list[str] = []
self.sentence_docs: list[str] = []
self.sentence_vecs: Optional[np.ndarray] = None
def add_document(self, doc_id: str, text: str) -> int:
"""Add a document. Returns number of sentences extracted."""
sents = self._split_sentences(text)
self.sentences.extend(sents)
self.sentence_docs.extend([doc_id] * len(sents))
return len(sents)
def build_index(self) -> dict:
"""Train Word2Vec on the corpus and compute sentence vectors."""
tokenized = [self._tokenize(s) for s in self.sentences]
self.model = Word2Vec(
sentences=tokenized,
vector_size=self.vector_size,
window=self.window,
min_count=self.min_count,
epochs=self.epochs,
sg=self.sg,
workers=4,
)
# Compute sentence vectors (average of word vectors)
vecs = []
for tokens in tokenized:
vecs.append(self._sentence_vector(tokens))
self.sentence_vecs = np.array(vecs, dtype=np.float32)
vocab_size = len(self.model.wv)
logger.info(f"Word2Vec trained: {vocab_size} words, {len(self.sentences)} sentences")
return {
"vocab_size": vocab_size,
"sentences": len(self.sentences),
"vector_size": self.vector_size,
}
def compare_texts(self, text_a: str, text_b: str) -> float:
"""Cosine similarity between two texts (averaged word vectors)."""
vec_a = self._sentence_vector(self._tokenize(text_a))
vec_b = self._sentence_vector(self._tokenize(text_b))
return float(self._cosine(vec_a, vec_b))
def query(self, text: str, top_k: int = 10) -> list[W2VResult]:
"""Find most similar sentences to a query."""
query_vec = self._sentence_vector(self._tokenize(text))
scores = self.sentence_vecs @ query_vec
norms = np.linalg.norm(self.sentence_vecs, axis=1) * np.linalg.norm(query_vec)
norms[norms == 0] = 1e-10
scores = scores / norms
top_idx = np.argsort(scores)[::-1][:top_k]
return [
W2VResult(
text=self.sentences[i],
doc_id=self.sentence_docs[i],
score=float(scores[i]),
rank=rank + 1,
)
for rank, i in enumerate(top_idx)
]
def most_similar_words(self, word: str, top_k: int = 10) -> list[tuple[str, float]]:
"""Find words most similar to a given word (static, no context)."""
word = word.lower()
if word not in self.model.wv:
return []
return self.model.wv.most_similar(word, topn=top_k)
def word_similarity(self, word_a: str, word_b: str) -> float:
"""Cosine similarity between two individual words."""
a, b = word_a.lower(), word_b.lower()
if a not in self.model.wv or b not in self.model.wv:
return 0.0
return float(self.model.wv.similarity(a, b))
# ------------------------------------------------------------------ #
# Persistence
# ------------------------------------------------------------------ #
def save(self, directory: str) -> dict:
"""Save trained Word2Vec state to disk for later restore."""
save_dir = Path(directory)
save_dir.mkdir(parents=True, exist_ok=True)
if self.model is None:
raise RuntimeError("Cannot save: model has not been trained yet.")
self.model.save(str(save_dir / "w2v.model"))
np.save(save_dir / "sentence_vecs.npy", self.sentence_vecs)
meta = {
"vector_size": self.vector_size,
"window": self.window,
"min_count": self.min_count,
"epochs": self.epochs,
"sg": self.sg,
"num_sentences": len(self.sentences),
"vocab_size": len(self.model.wv),
}
with open(save_dir / "w2v_meta.json", "w") as f:
json.dump(meta, f, indent=2)
# Save sentences and their doc mappings
with open(save_dir / "w2v_sentences.json", "w") as f:
json.dump({"sentences": self.sentences, "sentence_docs": self.sentence_docs}, f)
logger.info("Word2Vec saved to %s: %d sentences, %d vocab",
directory, len(self.sentences), len(self.model.wv))
return meta
@classmethod
def load(cls, directory: str) -> "Word2VecEngine":
"""Load a previously saved Word2Vec state from disk."""
save_dir = Path(directory)
if not (save_dir / "w2v_meta.json").is_file():
raise FileNotFoundError(f"No saved Word2Vec state at {directory}")
with open(save_dir / "w2v_meta.json") as f:
meta = json.load(f)
engine = cls(
vector_size=meta["vector_size"],
window=meta["window"],
min_count=meta["min_count"],
epochs=meta["epochs"],
sg=meta["sg"],
)
engine.model = Word2Vec.load(str(save_dir / "w2v.model"))
engine.sentence_vecs = np.load(save_dir / "sentence_vecs.npy")
with open(save_dir / "w2v_sentences.json") as f:
data = json.load(f)
engine.sentences = data["sentences"]
engine.sentence_docs = data["sentence_docs"]
logger.info("Word2Vec loaded from %s: %d sentences, %d vocab",
directory, len(engine.sentences), len(engine.model.wv))
return engine
@staticmethod
def has_saved_state(directory: str) -> bool:
"""Check if a saved Word2Vec state exists at the given directory."""
return (Path(directory) / "w2v_meta.json").is_file()
# ------------------------------------------------------------------ #
def _sentence_vector(self, tokens: list[str]) -> np.ndarray:
"""Average word vectors for a sentence."""
vecs = [self.model.wv[t] for t in tokens if t in self.model.wv]
if not vecs:
return np.zeros(self.vector_size, dtype=np.float32)
return np.mean(vecs, axis=0).astype(np.float32)
@staticmethod
def _cosine(a: np.ndarray, b: np.ndarray) -> float:
dot = np.dot(a, b)
norm = np.linalg.norm(a) * np.linalg.norm(b)
return dot / norm if norm > 0 else 0.0
@staticmethod
def _tokenize(text: str) -> list[str]:
return re.findall(r"\b[a-z]+\b", text.lower())
@staticmethod
def _split_sentences(text: str) -> list[str]:
parts = re.split(r"(?<=[.!?])\s+", text.strip())
return [s.strip() for s in parts if len(s.split()) >= 4]
|