\u60c5\u611f\u5206\u6790\u4f5c\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u7684\u6838\u5fc3\u4efb\u52a1\uff0c\u65e8\u5728\u901a\u8fc7\u7b97\u6cd5\u8bc6\u522b\u6587\u672c\u4e2d\u7684\u60c5\u611f\u503e\u5411\uff08\u79ef\u6781\/\u6d88\u6781\/\u4e2d\u6027\uff09\u3002\u5728\u4e2d\u6587\u573a\u666f\u4e0b\uff0c\u6c49\u5b57\u7279\u6709\u7684\u8868\u610f\u7279\u5f81\u548c\u590d\u6742\u8bed\u6cd5\u7ed3\u6784\u5bf9\u6a21\u578b\u8bbe\u8ba1\u63d0\u51fa\u66f4\u9ad8\u8981\u6c42\u3002PyTorch\u51ed\u501f\u52a8\u6001\u8ba1\u7b97\u56fe\u548c\u4e30\u5bcc\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u5e93\uff0c\u6210\u4e3a\u4e2d\u6587\u60c5\u611f\u5206\u6790\u7684\u4e3b\u6d41\u5f00\u53d1\u6846\u67b6\u3002<\/p>\n
\u5178\u578b\u5e94\u7528\u573a\u666f\u5305\u62ec\uff1a<\/p>\n
\u793e\u4ea4\u5a92\u4f53\u8206\u60c5\u76d1\u63a7\uff08\u5982\u5fae\u535a\u3001\u6296\u97f3\u8bc4\u8bba\u5206\u6790\uff09<\/p>\n<\/li>\n
\u7535\u5546\u5e73\u53f0\u4ea7\u54c1\u8bc4\u4ef7\u5206\u7c7b\uff08\u5982\u6dd8\u5b9d\u3001\u4eac\u4e1c\u8bc4\u8bba\u6781\u6027\u5224\u65ad\uff09<\/p>\n<\/li>\n
\u5ba2\u6237\u670d\u52a1\u53cd\u9988\u5904\u7406\uff08\u5982\u94f6\u884c\u5de5\u5355\u60c5\u611f\u5206\u7ea7\uff09<\/p>\n<\/li>\n<\/ul>\n
\u63a8\u8350\u4f7f\u7528\u4ee5\u4e0b\u516c\u5f00\u6570\u636e\u96c6\uff1a<\/p>\n
ChnSentiCorp<\/span>\uff1a\u5305\u542b\u9152\u5e97\u3001\u7535\u8111\u3001\u4e66\u7c4d\u9886\u57df\u8bc4\u8bba\uff0c\u6807\u6ce8\u79ef\u6781\/\u6d88\u6781\u6807\u7b7e<\/p>\n<\/li>\n NLPCC\u60c5\u611f\u5206\u6790\u6570\u636e\u96c6<\/span>\uff1a\u8986\u76d6\u5fae\u535a\u3001\u65b0\u95fb\u7b49\u573a\u666f\uff0c\u6807\u6ce8\u66f4\u7ec6\u7c92\u5ea6\u60c5\u611f<\/p>\n<\/li>\n THUCNews\u60c5\u611f\u5206\u7c7b\u6570\u636e\u96c6<\/span>\uff1a\u65b0\u95fb\u6587\u672c\u60c5\u611f\u6807\u6ce8\uff0c\u9002\u5408\u957f\u6587\u672c\u5206\u6790<\/p>\n<\/li>\n<\/ul>\n \u5173\u952e\u53c2\u6570\u63a7\u5236<\/span>\uff1a<\/p>\n \u6700\u5927\u5e8f\u5217\u957f\u5ea6\uff08max_len\uff09\uff1a\u5efa\u8bae\u901a\u8fc7\u6570\u636e\u7edf\u8ba1\u786e\u5b9a95%\u5206\u4f4d\u6570<\/p>\n<\/li>\n \u8bcd\u6c47\u8868\u5927\u5c0f\uff1a\u901a\u5e38\u9650\u5236\u57282\u4e07-5\u4e07\u8bcd<\/p>\n<\/li>\n \u505c\u7528\u8bcd\u8fc7\u6ee4\uff1a\u9700\u79fb\u9664"\u7684"\u3001"\u4e86"\u7b49\u9ad8\u9891\u65e0\u610f\u4e49\u8bcd<\/p>\n<\/li>\n<\/ul>\n \u4f18\u5316\u6280\u5de7<\/span>\uff1a<\/p>\n \u4f7f\u7528\u53cc\u5411LSTM\u6355\u6349\u4e0a\u4e0b\u6587\u4fe1\u606f<\/p>\n<\/li>\n \u6dfb\u52a0Dropout\u5c42\u9632\u6b62\u8fc7\u62df\u5408\uff08\u5efa\u8bae0.3-0.5\uff09<\/p>\n<\/li>\n \u91c7\u7528PackedSequence\u5904\u7406\u53d8\u957f\u5e8f\u5217<\/p>\n<\/li>\n<\/ul>\n \u6027\u80fd\u5bf9\u6bd4<\/span>\uff1a<\/p>\n <\/p>\n <\/p>\n \u4e8c\u5206\u7c7b\u4efb\u52a1\uff1a \u591a\u5206\u7c7b\u4efb\u52a1\uff1a <\/p>\n","protected":false},"excerpt":{"rendered":"2.2 \u5173\u952e\u9884\u5904\u7406\u6b65\u9aa4<\/h3>\n
import jieba\nfrom sklearn.model_selection import train_test_split\n\n# \u793a\u4f8b\u6570\u636e\u9884\u5904\u7406\u6d41\u7a0b\ndef preprocess_data(raw_data):\n # 1. \u6807\u7b7e\u6620\u5c04\n label_map = {'positive': 1, 'negative': 0}\n data = [(text, label_map[label]) for text, label in raw_data]\n \n # 2. \u4e2d\u6587\u5206\u8bcd\uff08\u4f7f\u7528jieba\u5e93\uff09\n data = [(jieba.lcut(text), label) for text, label in data]\n \n # 3. \u6784\u5efa\u8bcd\u6c47\u8868\uff08\u9700\u7edf\u8ba1\u8bcd\u9891\u5e76\u8fc7\u6ee4\u4f4e\u9891\u8bcd\uff09\n vocab = set()\n for text, _ in data:\n vocab.update(text)\n word_to_idx = {word: idx+1 for idx, word in enumerate(vocab)} # 0\u4fdd\u7559\u4e3apadding\n \n # 4. \u5e8f\u5217\u586b\u5145\u4e0e\u5206\u5272\n max_len = 128 # \u6839\u636e\u6570\u636e\u7edf\u8ba1\u8bbe\u5b9a\n X = []\n y = []\n for text, label in data:\n text_idx = [word_to_idx[word] for word in text if word in word_to_idx]\n padded_text = text_idx[:max_len] + [0]*(max_len - len(text_idx[:max_len]))\n X.append(padded_text)\n y.append(label)\n \n return train_test_split(X, y, test_size=0.2, random_state=42)<\/pre>\n\n
\u4e09\u3001\u6a21\u578b\u67b6\u6784\u8bbe\u8ba1<\/h2>\n
3.1 LSTM\u57fa\u7840\u6a21\u578b<\/h3>\n
import torch\nimport torch.nn as nn\n\nclass LSTMModel(nn.Module):\n def __init__(self, vocab_size, embedding_dim=128, hidden_dim=256, output_dim=2):\n super().__init__()\n self.embedding = nn.Embedding(vocab_size, embedding_dim, padding_idx=0)\n self.lstm = nn.LSTM(embedding_dim, hidden_dim, \n bidirectional=True, batch_first=True)\n self.fc = nn.Linear(hidden_dim*2, output_dim)\n self.dropout = nn.Dropout(0.5)\n \n def forward(self, x):\n # x shape: [batch_size, seq_len]\n embedded = self.dropout(self.embedding(x)) # [batch, seq, emb_dim]\n output, (hidden, _) = self.lstm(embedded)\n # \u53cc\u5411LSTM\u62fc\u63a5\u6700\u540e\u65f6\u523b\u7684\u9690\u85cf\u72b6\u6001\n hidden = torch.cat([hidden[-2], hidden[-1]], dim=1)\n return self.fc(self.dropout(hidden))<\/pre>\n
\n
3.2 BERT\u9884\u8bad\u7ec3\u6a21\u578b<\/h3>\n
from transformers import BertTokenizer, BertForSequenceClassification\n\nclass BertSentimentModel:\n def __init__(self, model_name='bert-base-chinese'):\n self.tokenizer = BertTokenizer.from_pretrained(model_name)\n self.model = BertForSequenceClassification.from_pretrained(\n model_name, num_labels=2)\n \n def predict(self, text):\n inputs = self.tokenizer(text, return_tensors="pt", \n max_length=128, truncation=True, padding='max_length')\n with torch.no_grad():\n outputs = self.model(**inputs)\n return torch.sigmoid(outputs.logits).item() > 0.5<\/pre>\n
\n\n
\n \n\u6a21\u578b\u7c7b\u578b<\/th>\n \u8bad\u7ec3\u6570\u636e\u91cf<\/th>\n \u51c6\u786e\u7387<\/th>\n \u63a8\u7406\u901f\u5ea6\uff08ms\/\u6837\u672c\uff09<\/th>\n<\/tr>\n<\/thead>\n \n LSTM<\/td>\n 10\u4e07\u6761<\/td>\n 82.7%<\/td>\n 12.5<\/td>\n<\/tr>\n \n BERT-base<\/td>\n 10\u4e07\u6761<\/td>\n 92.3%<\/td>\n 45.8<\/td>\n<\/tr>\n \n BERT-wwm-ext<\/td>\n 10\u4e07\u6761<\/td>\n 93.1%<\/td>\n 48.2<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u56db\u3001\u8bad\u7ec3\u4e0e\u4f18\u5316\u7b56\u7565<\/h2>\n
4.1 \u635f\u5931\u51fd\u6570\u9009\u62e9<\/h3>\n
\n
nn.BCEWithLogitsLoss()<\/code>\uff08\u63a8\u8350\uff09<\/p>\n<\/li>\nnn.CrossEntropyLoss()<\/code><\/p>\n<\/li>\n<\/ul>\n4.2 \u4f18\u5316\u5668\u914d\u7f6e<\/h3>\n
def train_model(model, train_loader, val_loader, epochs=10):\n criterion = nn.BCEWithLogitsLoss()\n optimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n \n for epoch in range(epochs):\n model.train()\n train_loss = 0\n for batch in train_loader:\n inputs, labels = batch\n optimizer.zero_grad()\n outputs = model(inputs)\n loss = criterion(outputs.squeeze(), labels.float())\n loss.backward()\n optimizer.step()\n train_loss += loss.item()\n \n # \u9a8c\u8bc1\u9636\u6bb5\n val_loss, val_acc = evaluate(model, val_loader)\n print(f"Epoch {epoch+1}: Train Loss={train_loss\/len(train_loader):.4f}, "\n f"Val Acc={val_acc*100:.2f}%")<\/pre>\n4.3 \u8bad\u7ec3\u5faa\u73af\u5b9e\u73b0<\/h3>\n
def train_model(model, train_loader, val_loader, epochs=10):\n criterion = nn.BCEWithLogitsLoss()\n optimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n \n for epoch in range(epochs):\n model.train()\n train_loss = 0\n for batch in train_loader:\n inputs, labels = batch\n optimizer.zero_grad()\n outputs = model(inputs)\n loss = criterion(outputs.squeeze(), labels.float())\n loss.backward()\n optimizer.step()\n train_loss += loss.item()\n \n # \u9a8c\u8bc1\u9636\u6bb5\n val_loss, val_acc = evaluate(model, val_loader)\n print(f"Epoch {epoch+1}: Train Loss={train_loss\/len(train_loader):.4f}, "\n f"Val Acc={val_acc*100:.2f}%")<\/pre>\n\u4e94\u3001\u90e8\u7f72\u4e0e\u5e94\u7528\u5b9e\u8df5<\/h2>\n
5.1 \u6a21\u578b\u5bfc\u51fa\u4e0e\u63a8\u7406\u4f18\u5316<\/h3>\n
# \u5bfc\u51fa\u4e3aTorchScript\u683c\u5f0f\ntraced_model = torch.jit.trace(model.eval(), example_input)\ntraced_model.save("sentiment_model.pt")\n\n# \u4f7f\u7528ONNX\u52a0\u901f\u63a8\u7406\ndummy_input = torch.randint(0, 20000, (1, 128))\ntorch.onnx.export(model, dummy_input, "sentiment.onnx",\n input_names=['input'], output_names=['output'])<\/pre>\n5.2 \u5b9e\u9645\u573a\u666f\u5e94\u7528\u793a\u4f8b<\/h3>\n
class SentimentAnalyzer:\n def __init__(self, model_path):\n self.model = torch.jit.load(model_path)\n self.vocab = self._load_vocab() # \u9700\u4fdd\u5b58\u8bad\u7ec3\u65f6\u7684\u8bcd\u6c47\u8868\n \n def analyze(self, text):\n tokens = jieba.lcut(text)\n token_ids = [self.vocab.get(token, 0) for token in tokens][:128]\n padded_input = token_ids + [0]*(128 - len(token_ids))\n input_tensor = torch.tensor([padded_input])\n \n with torch.no_grad():\n output = torch.sigmoid(self.model(input_tensor))\n return "\u79ef\u6781" if output.item() > 0.5 else "\u6d88\u6781"<\/pre>\n
<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-191","post","type-post","status-publish","format-standard","hentry","category-pythonbiji"],"_links":{"self":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts\/191","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=191"}],"version-history":[{"count":1,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts\/191\/revisions"}],"predecessor-version":[{"id":193,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts\/191\/revisions\/193"}],"wp:attachment":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=191"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=191"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=191"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}