\ub525\ub7ec\ub2dd\uacfc \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\uac00 \ubc1c\uc804\ud558\uba74\uc11c \ub2e4\uc591\ud55c \ubaa8\ub378\ub4e4\uc774 \ub4f1\uc7a5\ud558\uc600\uace0, \uadf8\uc911 BERT(Bidirectional Encoder Representations from Transformers)\ub294 \uc624\ub298\ub0a0 NLP\uc5d0\uc11c \uac00\uc7a5 \uc601\ud5a5\ub825 \uc788\ub294 \ubaa8\ub378 \uc911 \ud558\ub098\ub85c \uc790\ub9ac \uc7a1\uc558\uc2b5\ub2c8\ub2e4. \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec BERT\ub97c \uc559\uc0c1\ube14 \ubaa8\ub378\ub85c \uad6c\ud604\ud558\uae30 \uc704\ud574 \ub370\uc774\ud130\uc138\ud2b8\ub97c \uc900\ube44\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \ub2e4\ub8e8\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc559\uc0c1\ube14 \ud559\uc2b5\uc774\ub780 \uc5ec\ub7ec \uac1c\uc758 \ubaa8\ub378\uc744 \uacb0\ud569\ud558\uc5ec \uc131\ub2a5\uc744 \uac1c\uc120\ud558\ub294 \uae30\ubc95\uc785\ub2c8\ub2e4. \uc5ec\ub7ec \ubaa8\ub378\uc758 \uc608\uce21 \uacb0\uacfc\ub97c \uc870\ud569\ud568\uc73c\ub85c\uc368 \uac01\uac01\uc758 \ubaa8\ub378\uc774 \uac00\uc9c0\ub294 \ub2e8\uc810\uc744 \ubcf4\uc644\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc559\uc0c1\ube14 \ud559\uc2b5\uc740 \uc77c\ubc18\uc801\uc73c\ub85c \ub450 \uac00\uc9c0 \ubc29\uc2dd\uc73c\ub85c \uc774\ub8e8\uc5b4\uc9d1\ub2c8\ub2e4:<\/p>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 BERT \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ub2e4\uc218\uc758 \ubaa8\ub378\uc744 \uc870\ud569\ud574 \uc559\uc0c1\ube14 \ud559\uc2b5\uc744 \uad6c\ud604\ud558\ub294 \ub370 \uc911\uc810\uc744 \ub450\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ub2e4\uc591\ud55c \uc0ac\uc804 \ud6c8\ub828\ub41c \uc5b8\uc5b4 \ubaa8\ub378\uc744 \uc190\uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ub3c4\uc640\uc8fc\ub294 \ud30c\uc774\uc36c \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. BERT \ubaa8\ub378\ubfd0 \uc544\ub2c8\ub77c GPT, T5 \ub4f1\uc758 \ub2e4\uc591\ud55c \ubaa8\ub378\ub4e4\uc774 \ud3ec\ud568\ub418\uc5b4 \uc788\uc5b4 NLP \ud0dc\uc2a4\ud06c\ub97c \uc218\ud589\ud558\ub294 \ub370 \uc720\uc6a9\ud569\ub2c8\ub2e4. \uc774 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc758 \uc8fc\uc694 \uae30\ub2a5\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\uc774\uc81c BERT \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uae30 \uc704\ud574 \ud544\uc694\ud55c \ub370\uc774\ud130\uc14b\uc744 \uc900\ube44\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc6b0\uc120, \uc559\uc0c1\ube14 \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\uae30 \uc704\ud574 \uc0ac\uc6a9\ud560 \ub370\uc774\ud130\uc14b\uc744 \uc900\ube44\ud574\uc57c \ud569\ub2c8\ub2e4. BERT \ubaa8\ub378\uc744 \ud559\uc2b5\ud558\uae30 \uc704\ud574 \uc77c\ubc18\uc801\uc73c\ub85c\ub294 \ud14d\uc2a4\ud2b8\uc640 \ub77c\ubca8\uc774 \ud544\uc694\ud55c \ub370\uc774\ud130\uc14b\uc774 \ud544\uc694\ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uac10\uc815 \ubd84\uc11d \ubaa8\ub378\uc744 \ud6c8\ub828\ud55c\ub2e4\uace0 \uac00\uc815\ud560 \ub54c \ub2e4\uc74c\uacfc \uac19\uc740 \ud615\uc2dd\uc758 \ub370\uc774\ud130\uac00 \ud544\uc694\ud569\ub2c8\ub2e4:<\/p>\n
\n | \ud14d\uc2a4\ud2b8 | \ub77c\ubca8 |\n |------------------|------|\n | \"\uc88b\uc544\uc694!\" | 1 |\n | \"\uc2e4\ub9dd\ud588\uc5b4\uc694\" | 0 |\n | \"\ucd5c\uace0\uc758 \uacbd\ud5d8!\" | 1 |\n | \"\ub2e4\uc2dc\ub294 \uc548 \ud560\ub798\uc694\" | 0 |\n <\/code><\/pre>\n\ub370\uc774\ud130\ub97c \uc900\ube44\ud55c \ud6c4 CSV \ud30c\uc77c\ub85c \uc800\uc7a5\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \uc608\uc2dc\uc5d0\uc11c\ub294 \ud30c\uc774\uc36c\uc758 pandas \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec \ub370\uc774\ud130\ub97c CSV \ud615\uc2dd\uc73c\ub85c \uc800\uc7a5\ud569\ub2c8\ub2e4.<\/p>\n
\n import pandas as pd\n\n # \uc608\uc2dc \ub370\uc774\ud130 \uc0dd\uc131\n data = {\n 'text': [\n '\uc88b\uc544\uc694!', \n '\uc2e4\ub9dd\ud588\uc5b4\uc694', \n '\ucd5c\uace0\uc758 \uacbd\ud5d8!', \n '\ub2e4\uc2dc\ub294 \uc548 \ud560\ub798\uc694'\n ],\n 'label': [1, 0, 1, 0]\n }\n \n # \ub370\uc774\ud130\ud504\ub808\uc784\uc73c\ub85c \ubcc0\ud658\n df = pd.DataFrame(data)\n\n # CSV \ud30c\uc77c\ub85c \uc800\uc7a5\n df.to_csv('sentiment_data.csv', index=False, encoding='utf-8-sig')\n <\/code><\/pre>\n4. \ub370\uc774\ud130\uc14b \ub85c\ub529 \ubc0f \uc804\ucc98\ub9ac<\/h2>\n
CSV \ud30c\uc77c\ub85c \uc800\uc7a5\ub41c \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub529\ud558\uace0, BERT \ubaa8\ub378\uc5d0 \ub9de\uac8c \uc804\ucc98\ub9ac\ud574\uc57c \ud569\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 ‘transformers’ \ub77c\uc774\ube0c\ub7ec\ub9ac\uc5d0\uc11c \uc81c\uacf5\ud558\ub294 tokenizer\ub97c \uc0ac\uc6a9\ud558\uc5ec \ub370\uc774\ud130\ub97c \uc804\ucc98\ub9ac\ud569\ub2c8\ub2e4. \uba3c\uc800 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uace0 \ub85c\ub529\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\n !pip install transformers\n !pip install torch\n <\/code><\/pre>\n\uc774\uc81c \ud30c\uc774\uc36c \ucf54\ub4dc\ub85c \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub529\ud558\uace0 \uc804\ucc98\ub9ac\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\n from transformers import BertTokenizer\n\n # \ud1a0\ud06c\ub098\uc774\uc800 \ub85c\ub4dc\n tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\n\n # \ub370\uc774\ud130\uc14b \ub85c\ub529\n df = pd.read_csv('sentiment_data.csv')\n\n # \ud14d\uc2a4\ud2b8 \uc804\ucc98\ub9ac\n encodings = tokenizer(df['text'].tolist(), truncation=True, padding=True, max_length=128)\n\n # \ud14d\uc2a4\ud2b8\uc640 \ub77c\ubca8 \ud655\uc778\n print(encodings['input_ids'])\n print(df['label'].tolist())\n <\/code><\/pre>\n\uc704 \ucf54\ub4dc\uc5d0\uc11c ‘input_ids’\ub294 \uac01 \ub2e8\uc5b4\uac00 BERT \ubaa8\ub378\uc5d0 \uc785\ub825\ub418\uae30 \uc704\ud574 \ub9e4\ud551\ub41c \uc778\ub371\uc2a4 \uac12\uc774\uba70, \ub77c\ubca8\uc740 \uc608\uce21\ud558\uace0\uc790 \ud558\ub294 \ubaa9\ud45c\uc785\ub2c8\ub2e4. \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\uae30 \uc704\ud574 \ub370\uc774\ud130\uc14b\uc758 \ud615\ud0dc\ub85c \ubcc0\ud658\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
5. \ub370\uc774\ud130 \ub85c\ub354 \uc0dd\uc131<\/h2>\n
\ub370\uc774\ud130\ub97c \ubaa8\ub378\uc5d0 \uc804\ub2ec\ud558\uae30 \uc704\ud574 PyTorch\uc758 DataLoader\ub97c \uc0ac\uc6a9\ud558\uc5ec \ubc30\uce58 \ub2e8\uc704\ub85c \ub370\uc774\ud130\ub97c \ub9ac\ud134\ud558\ub294 \ud074\ub798\uc2a4\ub97c \ub9cc\ub4e4\uc5b4\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
\n import torch\n from torch.utils.data import Dataset, DataLoader\n\n class SentimentDataset(Dataset):\n def __init__(self, encodings, labels):\n self.encodings = encodings\n self.labels = labels\n\n def __getitem__(self, idx):\n item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}\n item['labels'] = torch.tensor(self.labels[idx])\n return item\n\n def __len__(self):\n return len(self.labels)\n\n # \ub370\uc774\ud130\uc14b \uac1d\uccb4 \ub9cc\ub4e4\uae30\n dataset = SentimentDataset(encodings, df['label'].tolist())\n\n # DataLoader \uc0dd\uc131\n train_loader = DataLoader(dataset, batch_size=2, shuffle=True)\n <\/code><\/pre>\n6. \ubaa8\ub378 \ud6c8\ub828<\/h2>\n
\ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\uae30 \uc704\ud574 BERT \ubaa8\ub378\uc744 \ub85c\ub4dc\ud558\uace0, \uc635\ud2f0\ub9c8\uc774\uc800\uc640 \uc190\uc2e4 \ud568\uc218\ub97c \uc124\uc815\ud569\ub2c8\ub2e4. \ud6c8\ub828 \ubc0f \ud3c9\uac00 \uacfc\uc815\uc5d0\uc11c BERT \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
\n from transformers import BertForSequenceClassification, AdamW\n\n # \ubaa8\ub378 \ub85c\ub4dc\n model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)\n\n # \uc635\ud2f0\ub9c8\uc774\uc800 \uc124\uc815\n optimizer = AdamW(model.parameters(), lr=5e-5)\n\n # GPU \uc0ac\uc6a9 \uc2dc, \ubaa8\ub378\uc744 GPU\ub85c \uc774\ub3d9\n if torch.cuda.is_available():\n model = model.cuda()\n\n # \ubaa8\ub378 \ud6c8\ub828\n model.train()\n for epoch in range(3): # \uc5d0\ud3ed \uc218\n for batch in train_loader:\n optimizer.zero_grad()\n \n # GPU \uc0ac\uc6a9 \uc2dc, \ubc30\uce58\ub97c GPU\ub85c \uc774\ub3d9\n if torch.cuda.is_available():\n batch = {k: v.cuda() for k, v in batch.items()}\n\n outputs = model(**batch)\n loss = outputs.loss\n loss.backward()\n optimizer.step()\n print(f'Epoch {epoch}, Loss: {loss.item()}')\n <\/code><\/pre>\n\ubaa8\ub378 \ud6c8\ub828 \uc2dc \ucd9c\ub825\ub41c \uc190\uc2e4 \uac12\uc740 \ubaa8\ub378\uc774 \uc5bc\ub9c8\ub098 \uc798 \ud559\uc2b5\ud558\uace0 \uc788\ub294\uc9c0\ub97c \ub098\ud0c0\ub0b4\ub294 \uc9c0\ud45c\uc785\ub2c8\ub2e4. \uc190\uc2e4 \uac12\uc774 \ub0ae\uc544\uc9c8\uc218\ub85d \ubaa8\ub378\uc758 \uc608\uce21 \uc131\ub2a5\uc774 \uc88b\uc544\uc9c4\ub2e4\uace0 \ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
7. \uc559\uc0c1\ube14 \ubaa8\ub378 \uad6c\ucd95<\/h2>\n
\uc5ec\ub7ec BERT \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a8 \ud6c4 \uc559\uc0c1\ube14\ud558\ub294 \ubc29\ubc95\uc5d0\ub294 \uc5ec\ub7ec \uac00\uc9c0\uac00 \uc788\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 \uac04\ub2e8\ud55c \ubc29\ubc95\uc73c\ub85c \ubaa8\ub378\uc758 \uc608\uce21 \uacb0\uacfc\ub97c \ud3c9\uade0\ub0b4\ub294 \ubc29\uc2dd\uc744 \uc0ac\uc6a9\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\n predictions = []\n\n # \uc559\uc0c1\ube14\ud560 \ubaa8\ub378 \uc218 \uc124\uc815\n model_count = 3\n for i in range(model_count):\n model = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)\n # \ubaa8\ub378 \ud6c8\ub828\uc740 \uc0dd\ub7b5(\uc704\uc5d0\uc11c \uc124\uba85\ud55c \ud6c8\ub828 \ucf54\ub4dc \uc0ac\uc6a9)\n # ...\n \n # \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc608\uce21\n model.eval()\n with torch.no_grad():\n outputs = model(**batch)\n logits = outputs.logits\n predictions.append(logits.softmax(dim=-1))\n\n # \uc608\uce21 \ud3c9\uade0\n final_predictions = torch.mean(torch.stack(predictions), dim=0)\n predicted_labels = final_predictions.argmax(dim=-1).tolist()\n <\/code><\/pre>\n8. \uacb0\uacfc \uac80\uc99d<\/h2>\n
\ubaa8\ub378\uc758 \uc608\uce21\ub825\uc744 \ud3c9\uac00\ud558\uae30 \uc704\ud574, \uc2e4\uc81c \ub77c\ubca8\uacfc\uc758 \ube44\uad50\ub97c \ud1b5\ud574 \uc815\ud655\ub3c4\ub97c \uacc4\uc0b0\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \uc815\ud655\ub3c4\ub97c \uacc4\uc0b0\ud558\uace0 \ucd9c\ub825\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/p>\n
\n from sklearn.metrics import accuracy_score\n\n # \uc2e4\uc81c \ub77c\ubca8\n true_labels = df['label'].tolist()\n\n # \uc815\ud655\ub3c4 \uacc4\uc0b0\n accuracy = accuracy_score(true_labels, predicted_labels)\n print(f'Accuracy: {accuracy * 100:.2f}%')\n <\/code><\/pre>\n9. \ucd5c\uc885 \uc815\ub9ac<\/h2>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec BERT \ubaa8\ub378\uc744 \uc559\uc0c1\ube14\ub85c \uad6c\uc131\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \ub370\uc774\ud130\uc138\ud2b8\ub97c \uc900\ube44\ud558\uace0, \uc804\ucc98\ub9ac \ubc0f DataLoader \uc0dd\uc131\uc744 \ud1b5\ud574 BERT \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ucf30\uc2b5\ub2c8\ub2e4. \ub9c8\uc9c0\ub9c9\uc73c\ub85c \uc5ec\ub7ec \ubaa8\ub378\uc758 \uc608\uce21 \uacb0\uacfc\ub97c \uc559\uc0c1\ube14\ud558\uc5ec \ucd5c\uc885 \uacb0\uacfc\ub97c \ub3c4\ucd9c\ud558\uc600\uc2b5\ub2c8\ub2e4.<\/p>\n
\ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \uc131\ub2a5 \ud5a5\uc0c1\uc744 \uc704\ud574 \uc559\uc0c1\ube14 \uae30\ubc95\uc744 \uc801\uc6a9\ud558\ub294 \uac83\uc740 \ub9e4\uc6b0 \ud6a8\uacfc\uc801\uc778 \ubc29\ubc95\uc785\ub2c8\ub2e4. \uc774 \uac15\uc88c\uc5d0\uc11c \ubc30\uc6b4 \ub0b4\uc6a9\uc744 \ubc14\ud0d5\uc73c\ub85c \ub2e4\uc591\ud55c \ubaa8\ub378\uacfc \ub370\uc774\ud130\uc14b\uc5d0 \uc2e4\ud5d8\ud574 \ubcf4\uae38 \uad8c\uc7a5\ud569\ub2c8\ub2e4.<\/p>\n
10. \ucc38\uace0 \uc790\ub8cc<\/h2>\n