Spaces:
Sleeping
Sleeping
File size: 15,590 Bytes
b4c92f5 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 |
"""
FastText embedding module for Tibetan text.
This module provides functions to train and use FastText models for Tibetan text.
"""
import os
import math
import logging
import numpy as np
import fasttext
from typing import List, Optional
from huggingface_hub import hf_hub_download
# Set up logging
logger = logging.getLogger(__name__)
# Default parameters optimized for Tibetan
DEFAULT_DIM = 100
DEFAULT_EPOCH = 5
DEFAULT_MIN_COUNT = 5
DEFAULT_WINDOW = 5
DEFAULT_MINN = 3
DEFAULT_MAXN = 6
DEFAULT_NEG = 5
# Define paths for model storage
DEFAULT_MODEL_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "models")
DEFAULT_MODEL_PATH = os.path.join(DEFAULT_MODEL_DIR, "fasttext_model.bin")
# Facebook's official Tibetan FastText model
FACEBOOK_TIBETAN_MODEL_ID = "facebook/fasttext-bo-vectors"
FACEBOOK_TIBETAN_MODEL_FILE = "model.bin"
# Create models directory if it doesn't exist
os.makedirs(DEFAULT_MODEL_DIR, exist_ok=True)
def ensure_dir_exists(directory: str) -> None:
"""
Ensure that a directory exists, creating it if necessary.
Args:
directory: Directory path to ensure exists
"""
if not os.path.exists(directory):
os.makedirs(directory, exist_ok=True)
def train_fasttext_model(
corpus_path: str,
model_path: str = DEFAULT_MODEL_PATH,
dim: int = DEFAULT_DIM,
epoch: int = DEFAULT_EPOCH,
min_count: int = DEFAULT_MIN_COUNT,
window: int = DEFAULT_WINDOW,
minn: int = DEFAULT_MINN,
maxn: int = DEFAULT_MAXN,
neg: int = DEFAULT_NEG,
model_type: str = "skipgram"
) -> fasttext.FastText._FastText:
"""
Train a FastText model on Tibetan corpus using optimized parameters.
Args:
corpus_path: Path to the corpus file
model_path: Path where to save the trained model
dim: Embedding dimension (default: 300)
epoch: Number of training epochs (default: 15)
min_count: Minimum count of words (default: 3)
window: Size of context window (default: 5)
minn: Minimum length of char n-gram (default: 3)
maxn: Maximum length of char n-gram (default: 6)
neg: Number of negatives in negative sampling (default: 10)
model_type: FastText model type ('skipgram' or 'cbow')
Returns:
Trained FastText model
"""
ensure_dir_exists(os.path.dirname(model_path))
logger.info("Training FastText model with %s, dim=%d, epoch=%d, window=%d, minn=%d, maxn=%d...",
model_type, dim, epoch, window, minn, maxn)
# Preprocess corpus for Tibetan - segment by syllable points
# This is based on research showing syllable segmentation works better for Tibetan
try:
with open(corpus_path, 'r', encoding='utf-8') as f:
content = f.read()
# Ensure syllable segmentation by adding spaces after Tibetan syllable markers (if not already present)
# This improves model quality for Tibetan text according to research
processed_content = content.replace('་', '་ ')
# Write back the processed content
with open(corpus_path, 'w', encoding='utf-8') as f:
f.write(processed_content)
logger.info("Preprocessed corpus with syllable segmentation for Tibetan text")
except Exception as e:
logger.warning("Could not preprocess corpus for syllable segmentation: %s", str(e))
# Train the model with optimized parameters
if model_type == "skipgram":
model = fasttext.train_unsupervised(
corpus_path,
model="skipgram",
dim=dim,
epoch=epoch,
minCount=min_count,
wordNgrams=1,
minn=minn,
maxn=maxn,
neg=neg,
window=window
)
else: # cbow
model = fasttext.train_unsupervised(
corpus_path,
model="cbow",
dim=dim,
epoch=epoch,
minCount=min_count,
wordNgrams=1,
minn=minn,
maxn=maxn,
neg=neg,
window=window
)
# Save the model
model.save_model(model_path)
logger.info("FastText model trained and saved to %s", model_path)
return model
def load_fasttext_model(model_path: str = DEFAULT_MODEL_PATH) -> Optional[fasttext.FastText._FastText]:
"""
Load a FastText model from file, with fallback to official Facebook model.
Args:
model_path: Path to the model file
Returns:
Loaded FastText model or None if loading fails
"""
try:
# First try to load the official Facebook FastText Tibetan model
try:
# Try to download the official Facebook FastText Tibetan model
logger.info("Attempting to download and load official Facebook FastText Tibetan model")
facebook_model_path = hf_hub_download(
repo_id=FACEBOOK_TIBETAN_MODEL_ID,
filename=FACEBOOK_TIBETAN_MODEL_FILE,
cache_dir=DEFAULT_MODEL_DIR
)
logger.info("Loading official Facebook FastText Tibetan model from %s", facebook_model_path)
return fasttext.load_model(facebook_model_path)
except Exception as e:
logger.warning("Could not load official Facebook FastText Tibetan model: %s", str(e))
logger.info("Falling back to local model")
# Fall back to local model
if os.path.exists(model_path):
logger.info("Loading local FastText model from %s", model_path)
return fasttext.load_model(model_path)
else:
logger.warning("Model path %s does not exist", model_path)
return None
except Exception as e:
logger.error("Error loading FastText model: %s", str(e))
return None
def get_text_embedding(
text: str,
model: fasttext.FastText._FastText,
tokenize_fn=None,
use_stopwords: bool = True,
stopwords_set=None,
use_tfidf_weighting: bool = True, # Enabled by default for better results
corpus_token_freq=None
) -> np.ndarray:
"""
Get embedding for a text using a FastText model with optional TF-IDF weighting.
Args:
text: Input text
model: FastText model
tokenize_fn: Optional tokenization function or pre-tokenized list
use_stopwords: Whether to filter out stopwords before computing embeddings
stopwords_set: Set of stopwords to filter out (if use_stopwords is True)
use_tfidf_weighting: Whether to use TF-IDF weighting for averaging word vectors
corpus_token_freq: Dictionary of token frequencies across corpus (required for TF-IDF)
Returns:
Text embedding vector
"""
if not text.strip():
return np.zeros(model.get_dimension())
# Handle tokenization
if tokenize_fn is None:
# Simple whitespace tokenization as fallback
tokens = text.split()
elif isinstance(tokenize_fn, list):
# If tokenize_fn is already a list of tokens, use it directly
tokens = tokenize_fn
elif callable(tokenize_fn):
# If tokenize_fn is a function, call it
tokens = tokenize_fn(text)
else:
# If tokenize_fn is something else (like a string), use whitespace tokenization
logger.warning(f"Unexpected tokenize_fn type: {type(tokenize_fn)}. Using default whitespace tokenization.")
tokens = text.split()
# Filter out stopwords if enabled and stopwords_set is provided
if use_stopwords and stopwords_set is not None:
tokens = [token for token in tokens if token not in stopwords_set]
# If all tokens were filtered out as stopwords, return zero vector
if not tokens:
return np.zeros(model.get_dimension())
# Filter out empty tokens
tokens = [token for token in tokens if token.strip()]
if not tokens:
return np.zeros(model.get_dimension())
# Calculate TF-IDF weighted average if requested
if use_tfidf_weighting and corpus_token_freq is not None:
# Calculate term frequencies in this document
token_counts = {}
for token in tokens:
token_counts[token] = token_counts.get(token, 0) + 1
# Calculate IDF for each token with improved stability
N = sum(corpus_token_freq.values()) # Total number of tokens in corpus
N = max(N, 1) # Ensure N is at least 1 to avoid division by zero
# Compute TF-IDF weights with safeguards against extreme values
weights = []
for token in tokens:
# Term frequency in this document
tf = token_counts.get(token, 0) / max(len(tokens), 1) if len(tokens) > 0 else 0
# Inverse document frequency with smoothing to avoid extreme values
token_freq = corpus_token_freq.get(token, 0)
idf = math.log((N + 1) / (token_freq + 1)) + 1 # Add 1 for smoothing
# TF-IDF weight with bounds to prevent extreme values
weight = tf * idf
weight = min(max(weight, 0.1), 10.0) # Limit to reasonable range
weights.append(weight)
# Normalize weights to sum to 1 with stability checks
total_weight = sum(weights)
if total_weight > 0:
weights = [w / total_weight for w in weights]
else:
# If all weights are 0, use uniform weights
weights = [1.0 / len(tokens) if len(tokens) > 0 else 0 for _ in tokens]
# Check for NaN or infinite values and replace with uniform weights if found
if any(math.isnan(w) or math.isinf(w) for w in weights):
logger.warning("Found NaN or infinite weights in TF-IDF calculation. Using uniform weights instead.")
weights = [1.0 / len(tokens) if len(tokens) > 0 else 0 for _ in tokens]
# Get vectors for each token and apply weights
vectors = [model.get_word_vector(token) for token in tokens]
weighted_vectors = [w * v for w, v in zip(weights, vectors)]
# Sum the weighted vectors
return np.sum(weighted_vectors, axis=0)
else:
# Simple averaging if TF-IDF is not enabled or corpus frequencies not provided
vectors = [model.get_word_vector(token) for token in tokens]
return np.mean(vectors, axis=0)
def get_batch_embeddings(
texts: List[str],
model: fasttext.FastText._FastText,
tokenize_fn=None,
use_stopwords: bool = True,
stopwords_set=None,
use_tfidf_weighting: bool = True, # Enabled by default for better results
corpus_token_freq=None
) -> np.ndarray:
"""
Get embeddings for a batch of texts with optional TF-IDF weighting.
Args:
texts: List of input texts
model: FastText model
tokenize_fn: Optional tokenization function or pre-tokenized list of tokens
use_stopwords: Whether to filter out stopwords before computing embeddings
stopwords_set: Set of stopwords to filter out (if use_stopwords is True)
use_tfidf_weighting: Whether to use TF-IDF weighting for averaging word vectors
corpus_token_freq: Dictionary of token frequencies across corpus (required for TF-IDF)
Returns:
Array of text embedding vectors
"""
# If corpus_token_freq is not provided but TF-IDF is requested, build it from the texts
if use_tfidf_weighting and corpus_token_freq is None:
logger.info("Building corpus token frequency dictionary for TF-IDF weighting")
corpus_token_freq = {}
# Process each text to build corpus token frequencies
for text in texts:
if not text.strip():
continue
# Handle tokenization
if tokenize_fn is None:
tokens = text.split()
elif isinstance(tokenize_fn, list):
# In this case, tokenize_fn should be a list of lists (one list of tokens per text)
# This is not a common use case, so we'll just use the first one as fallback
tokens = tokenize_fn[0] if tokenize_fn else []
else:
tokens = tokenize_fn(text)
# Filter out stopwords if enabled
if use_stopwords and stopwords_set is not None:
tokens = [token for token in tokens if token not in stopwords_set]
# Update corpus token frequencies
for token in tokens:
if token.strip(): # Skip empty tokens
corpus_token_freq[token] = corpus_token_freq.get(token, 0) + 1
logger.info("Built corpus token frequency dictionary with %d unique tokens", len(corpus_token_freq))
# Get embeddings for each text
embeddings = []
for i, text in enumerate(texts):
# Handle pre-tokenized input
tokens = None
if isinstance(tokenize_fn, list):
if i < len(tokenize_fn):
tokens = tokenize_fn[i]
embedding = get_text_embedding(
text,
model,
tokenize_fn=tokens, # Pass the tokens directly, not the function
use_stopwords=use_stopwords,
stopwords_set=stopwords_set,
use_tfidf_weighting=use_tfidf_weighting,
corpus_token_freq=corpus_token_freq
)
embeddings.append(embedding)
return np.array(embeddings)
def generate_embeddings(
texts: List[str],
model: fasttext.FastText._FastText,
device: str,
model_type: str = "sentence_transformer",
tokenize_fn=None,
use_stopwords: bool = True,
use_lite_stopwords: bool = False
) -> np.ndarray:
"""
Generate embeddings for a list of texts using a FastText model.
Args:
texts: List of input texts
model: FastText model
device: Device to use for computation (not used for FastText)
model_type: Model type ('sentence_transformer' or 'fasttext')
tokenize_fn: Optional tokenization function or pre-tokenized list of tokens
use_stopwords: Whether to filter out stopwords
use_lite_stopwords: Whether to use a lighter set of stopwords
Returns:
Array of text embedding vectors
"""
if model_type != "fasttext":
logger.warning("Model type %s not supported for FastText. Using FastText anyway.", model_type)
# Generate embeddings using FastText
try:
# Load stopwords if needed
stopwords_set = None
if use_stopwords:
from .tibetan_stopwords import get_stopwords
stopwords_set = get_stopwords(use_lite=use_lite_stopwords)
logger.info("Loaded %d Tibetan stopwords", len(stopwords_set))
# Generate embeddings
embeddings = get_batch_embeddings(
texts,
model,
tokenize_fn=tokenize_fn,
use_stopwords=use_stopwords,
stopwords_set=stopwords_set,
use_tfidf_weighting=True # Enable TF-IDF weighting for better results
)
logger.info("FastText embeddings generated with shape: %s", str(embeddings.shape))
return embeddings
except Exception as e:
logger.error("Error generating FastText embeddings: %s", str(e))
# Return empty embeddings as fallback
return np.zeros((len(texts), model.get_dimension()))
|