Skip to content

word2vec

Word2Vec embedding implementation.

Word2VecEmbedding

Bases: BaseEmbedding

Word2Vec embedding backend.

Source code in src/xwhy/embeddings/word2vec.py
 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
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
class Word2VecEmbedding(BaseEmbedding):
    """Word2Vec embedding backend."""

    _MODEL_FILE_MAP: ClassVar[dict[str, dict[str, Any]]] = {
        "word2vec-google-news-300": {
            "file": "GoogleNews-vectors-negative300.bin",
            "binary": True,
            "gensim": True,
            "no_header": False,
            "google_id": "1vAjPzr5R1RQiuh9NOgHVFGhBmCuYWnJU",
        },
        "glove.840B.300d": {
            "file": "glove.840B.300d.txt",
            "binary": False,
            "gensim": True,
            "no_header": True,
            "google_id": "19cJAKkgrYAiT1gU-OnWTN6GaGcd7pLZI",
        },
        "paragram_300_sl999": {
            "file": "paragram_300_sl999.txt",
            "binary": False,
            "gensim": True,
            "no_header": True,
            "google_id": "1c-16FP0jvaJeyaM8JcqqKPdoVw7uqVkK",
        },
        "paragram-300-WS353": {
            "file": "paragram_300_ws353.txt",
            "binary": False,
            "gensim": True,
            "no_header": True,
            "google_id": "1bBLz6F6MJ_qx9xnSZUgI8W0QZhJYgBdu",
        },
    }

    def __init__(
        self,
        *,
        settings: Settings,
        model_name: str = "word2vec-google-news-300",
        force_download: bool = False,
        **kwargs: Any,  # noqa: ANN401
    ) -> None:
        """Initialize Word2Vec embedding backend."""
        self._settings = settings
        self._model_name = model_name
        self._force_download = force_download
        self._model: KeyedVectors | None = None

    def load(self) -> KeyedVectors:
        """Load embedding model with caching strategy."""
        if self._model is not None:
            return self._model

        cache_dir = self._get_cache_dir()
        cache_dir.mkdir(parents=True, exist_ok=True)

        if self._model_name not in self._MODEL_FILE_MAP:
            raise ValueError(f"Unsupported model: {self._model_name}")

        model_info = self._MODEL_FILE_MAP[self._model_name]
        original_model_path = cache_dir / model_info["file"]

        # 1. cache
        bin_cache_path = original_model_path.with_suffix(".bin")

        if bin_cache_path.exists() and not self._force_download:
            logger.debug(
                f"Loading embedding model from fast binary cache: {bin_cache_path}"
            )
            self._model = KeyedVectors.load_word2vec_format(
                str(bin_cache_path),
                binary=True,
            )
            return self._model

        # 2. gdown direct download option
        gdown_model = self._try_gdown_download(
            model_info=model_info,
            bin_cache_path=bin_cache_path,
        )
        if gdown_model is not None:
            return gdown_model

        # 3. gensim
        if model_info["gensim"]:
            logger.debug(f"Loading embedding model via gensim: {self._model_name}")
            try:
                model = api.load(self._model_name)
                logger.info(f"Saving to fast binary cache: {bin_cache_path}")
                model.save_word2vec_format(str(bin_cache_path), binary=True)
                self._model = model
                return model

            except Exception:
                pass

        # 4. fallback download for specific models
        if self._model_name == "word2vec-google-news-300":
            return self._download_google_news(cache_dir, bin_cache_path)
        elif self._model_name == "glove.840B.300d":
            return self._download_glove(cache_dir, bin_cache_path, original_model_path)
        elif self._model_name in ("paragram_300_sl999", "paragram-300-WS353"):
            return self._download_paragram(
                cache_dir, bin_cache_path, original_model_path
            )

        raise RuntimeError(f"Failed to load embedding model: {self._model_name}")

    def encode(self, text: str) -> list[float]:
        """Encode text using averaged word vectors."""
        model = self.load()
        words = text.split()

        vectors: list[list[float]] = [
            model[word].tolist() for word in words if word in model
        ]

        if not vectors:
            return [0.0] * 300

        dim = len(vectors[0])
        result = [0.0] * dim

        for vec in vectors:
            for i, val in enumerate(vec):
                result[i] += val

        return [x / len(vectors) for x in result]

    def _get_cache_dir(self) -> Path:
        cache_dir = self._settings.embedding_cache_dir
        return cache_dir if cache_dir else Path(Path.home() / ".cache/xwhy/embeddings")

    def _try_gdown_download(
        self,
        model_info: dict[str, Any],
        bin_cache_path: Path,
    ) -> KeyedVectors | None:
        """Attempt to download pre-converted binary model from Google Drive."""
        google_id: str | None = model_info.get("google_id")

        if not google_id or self._force_download:
            return None

        logger.debug(f"Attempting binary download via gdown for ID: {google_id}")

        try:
            # Download directly to bin_cache_path since the GDrive file is already .bin
            gdown.download(id=google_id, output=str(bin_cache_path), quiet=False)

            # All files on GDrive are pre-converted binary format. Force binary=True.
            model = KeyedVectors.load_word2vec_format(
                str(bin_cache_path),
                binary=True,
                no_header=False,
            )

            self._model = model
            return self._model

        except Exception as err:
            logger.debug(f"gdown route failed for {self._model_name}: {err}")
            if bin_cache_path.exists():
                bin_cache_path.unlink()
            return None

    def _download_google_news(
        self,
        cache_dir: Path,
        bin_cache_path: Path,
    ) -> KeyedVectors:
        """Download GoogleNews model."""
        gz_path = cache_dir / "GoogleNews-vectors-negative300.bin.gz"
        url = (
            "https://public.ukp.informatik.tu-darmstadt.de/"
            "reimers/wordembeddings/GoogleNews-vectors-negative300.bin.gz"
        )

        try:
            if not gz_path.exists():
                self._download_file(url, gz_path)

            self._extract_gzip(gz_path, bin_cache_path)

            self._model = KeyedVectors.load_word2vec_format(
                str(bin_cache_path),
                binary=True,
            )

            return self._model

        except Exception as error:
            raise RuntimeError("Failed to load GoogleNews model") from error

    def _download_glove(
        self, cache_dir: Path, bin_path: Path, txt_path: Path
    ) -> KeyedVectors:
        """Download, clean, and convert GloVe model to binary."""
        zip_path = cache_dir / "glove.840B.300d.zip"
        url = "http://nlp.stanford.edu/data/glove.840B.300d.zip"

        try:
            if not zip_path.exists():
                self._download_file(url, zip_path)

            if not txt_path.exists():
                logger.info("Extracting GloVe zip file...")
                with zipfile.ZipFile(zip_path, "r") as zip_ref:
                    zip_ref.extractall(cache_dir)

            cleaned_path = txt_path.with_name(txt_path.name + ".cleaned")
            if not cleaned_path.exists():
                seen = set()
                logger.info("De-duplicating GloVe file...")
                with (
                    open(txt_path, encoding="utf-8", errors="ignore") as fin,
                    open(cleaned_path, "w", encoding="utf-8") as fout,
                ):
                    for line in tqdm(fin, desc="Cleaning GloVe"):
                        word = line.split(maxsplit=1)[0]
                        if word not in seen:
                            fout.write(line)
                            seen.add(word)

            logger.info("Loading cleaned GloVe text...")
            model = KeyedVectors.load_word2vec_format(
                str(cleaned_path), binary=False, no_header=True
            )

            logger.info("Converting GloVe to fast binary format...")
            model.save_word2vec_format(str(bin_path), binary=True)

            for path_to_remove in [cleaned_path, txt_path, zip_path]:
                if path_to_remove.exists():
                    path_to_remove.unlink()

            self._model = model
            return self._model
        except Exception as error:
            raise RuntimeError("Failed to load GloVe model") from error

    def _download_paragram(
        self, cache_dir: Path, bin_path: Path, txt_path: Path
    ) -> KeyedVectors:
        """Download and convert Paragram models to binary."""
        gdrive_ids = {
            "paragram_300_sl999": "0B9w48e1rj-MOck1fRGxaZW1LU2M",
            "paragram-300-WS353": "0B9w48e1rj-MOLVdZRzFfTlNsem8",
        }
        file_id = gdrive_ids[self._model_name]

        try:
            if not txt_path.exists():
                logger.info(f"Downloading {self._model_name} via gdown...")

                temp_path = txt_path.with_suffix(".tmp")
                gdown.download(id=file_id, output=str(temp_path), quiet=False)

                if zipfile.is_zipfile(str(temp_path)):
                    logger.info("Extracting zip archive...")
                    with zipfile.ZipFile(str(temp_path), "r") as zf:
                        # Find the text file name inside the zip archive
                        txt_filename = next(
                            (name for name in zf.namelist() if name.endswith(".txt")),
                            None,
                        )
                        if not txt_filename:
                            raise FileNotFoundError("No .txt file found inside zip.")

                        # Manually extract with Python to bypass CRC errors
                        with open(txt_path, "wb") as f_out:
                            try:
                                with zf.open(txt_filename) as f_in:
                                    shutil.copyfileobj(f_in, f_out)
                            except zipfile.BadZipFile:
                                logger.warning("Ignored CRC error.")
                    temp_path.unlink()
                else:
                    temp_path.rename(txt_path)

            logger.info("Sanitizing text file and adding header...")
            clean_txt_path = txt_path.with_suffix(".clean.txt")
            expected_dim = 300

            # Step 1: Count valid lines (lines with exactly 301
            # parts: 1 word + 300 numbers)
            valid_lines = 0
            with open(txt_path, encoding="utf-8", errors="ignore") as f:
                for line in f:
                    if len(line.rstrip("\n").split(" ")) == expected_dim + 1:
                        valid_lines += 1

            # Step 2: Rewrite the file with only valid lines
            # + add standard word2vec header
            with (
                open(txt_path, encoding="utf-8", errors="ignore") as f_in,
                open(clean_txt_path, "w", encoding="utf-8") as f_out,
            ):
                f_out.write(f"{valid_lines} {expected_dim}\n")
                for line in f_in:
                    if len(line.rstrip("\n").split(" ")) == expected_dim + 1:
                        f_out.write(line)

            logger.info(f"Loading {self._model_name} text into Gensim...")
            # Since we added the header to the cleaned file
            # ourselves, set no_header=False
            model = KeyedVectors.load_word2vec_format(
                str(clean_txt_path),
                binary=False,
                no_header=False,
                unicode_errors="ignore",
            )

            logger.info(f"Converting {self._model_name} to binary...")
            model.save_word2vec_format(str(bin_path), binary=True)

            # Clean up both text files (original and cleaned) to free up disk space
            if txt_path.exists():
                txt_path.unlink()
            if clean_txt_path.exists():
                clean_txt_path.unlink()

            self._model = model
            return self._model
        except Exception as error:
            raise RuntimeError(f"Failed to load {self._model_name}") from error

    @staticmethod
    def _download_file(url: str, path: Path) -> None:
        """Download a file using streaming requests."""
        path.parent.mkdir(parents=True, exist_ok=True)
        try:
            logger.debug(f"Attempting direct download from {url}...")
            response = requests.get(url, stream=True, timeout=600)
            response.raise_for_status()
            total_size = int(response.headers.get("content-length", 0))

            with (
                path.open("wb") as file,
                tqdm(
                    desc=path.name,
                    total=total_size,
                    unit="iB",
                    unit_scale=True,
                    unit_divisor=1024,
                ) as bar,
            ):
                for chunk in response.iter_content(chunk_size=8192):
                    if chunk:
                        size = file.write(chunk)
                        bar.update(size)

            if path.stat().st_size < 100:
                raise OSError("Downloaded file too small")

        except Exception:
            if path.exists():
                path.unlink()

            raise

    @staticmethod
    def _extract_gzip(src: Path, dst: Path) -> None:
        """Extract gzip file."""
        logger.debug("Extracting %s...", src)
        with gzip.open(src, "rb") as f_in, dst.open("wb") as f_out:
            shutil.copyfileobj(f_in, f_out)

__init__(*, settings, model_name='word2vec-google-news-300', force_download=False, **kwargs)

Initialize Word2Vec embedding backend.

Source code in src/xwhy/embeddings/word2vec.py
56
57
58
59
60
61
62
63
64
65
66
67
68
def __init__(
    self,
    *,
    settings: Settings,
    model_name: str = "word2vec-google-news-300",
    force_download: bool = False,
    **kwargs: Any,  # noqa: ANN401
) -> None:
    """Initialize Word2Vec embedding backend."""
    self._settings = settings
    self._model_name = model_name
    self._force_download = force_download
    self._model: KeyedVectors | None = None

encode(text)

Encode text using averaged word vectors.

Source code in src/xwhy/embeddings/word2vec.py
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
def encode(self, text: str) -> list[float]:
    """Encode text using averaged word vectors."""
    model = self.load()
    words = text.split()

    vectors: list[list[float]] = [
        model[word].tolist() for word in words if word in model
    ]

    if not vectors:
        return [0.0] * 300

    dim = len(vectors[0])
    result = [0.0] * dim

    for vec in vectors:
        for i, val in enumerate(vec):
            result[i] += val

    return [x / len(vectors) for x in result]

load()

Load embedding model with caching strategy.

Source code in src/xwhy/embeddings/word2vec.py
 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
def load(self) -> KeyedVectors:
    """Load embedding model with caching strategy."""
    if self._model is not None:
        return self._model

    cache_dir = self._get_cache_dir()
    cache_dir.mkdir(parents=True, exist_ok=True)

    if self._model_name not in self._MODEL_FILE_MAP:
        raise ValueError(f"Unsupported model: {self._model_name}")

    model_info = self._MODEL_FILE_MAP[self._model_name]
    original_model_path = cache_dir / model_info["file"]

    # 1. cache
    bin_cache_path = original_model_path.with_suffix(".bin")

    if bin_cache_path.exists() and not self._force_download:
        logger.debug(
            f"Loading embedding model from fast binary cache: {bin_cache_path}"
        )
        self._model = KeyedVectors.load_word2vec_format(
            str(bin_cache_path),
            binary=True,
        )
        return self._model

    # 2. gdown direct download option
    gdown_model = self._try_gdown_download(
        model_info=model_info,
        bin_cache_path=bin_cache_path,
    )
    if gdown_model is not None:
        return gdown_model

    # 3. gensim
    if model_info["gensim"]:
        logger.debug(f"Loading embedding model via gensim: {self._model_name}")
        try:
            model = api.load(self._model_name)
            logger.info(f"Saving to fast binary cache: {bin_cache_path}")
            model.save_word2vec_format(str(bin_cache_path), binary=True)
            self._model = model
            return model

        except Exception:
            pass

    # 4. fallback download for specific models
    if self._model_name == "word2vec-google-news-300":
        return self._download_google_news(cache_dir, bin_cache_path)
    elif self._model_name == "glove.840B.300d":
        return self._download_glove(cache_dir, bin_cache_path, original_model_path)
    elif self._model_name in ("paragram_300_sl999", "paragram-300-WS353"):
        return self._download_paragram(
            cache_dir, bin_cache_path, original_model_path
        )

    raise RuntimeError(f"Failed to load embedding model: {self._model_name}")