\ub525\ub7ec\ub2dd\uacfc \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\ub294 \ud604\ub300\uc758 \uc778\uacf5\uc9c0\ub2a5 \uae30\uc220\uc5d0\uc11c \uc911\uc694\ud55c \uc5ed\ud560\uc744 \ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uadf8 \uc911\uc5d0\uc11c\ub3c4 BERT(Bidirectional Encoder Representations from Transformers)\ub294 \uad6c\uae00\uc774 \uac1c\ubc1c\ud55c \uac15\ub825\ud55c \uc5b8\uc5b4 \ubaa8\ub378\ub85c, \ub9ce\uc740 NLP \uc791\uc5c5\uc5d0\uc11c \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc8fc\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec BERT \ubaa8\ub378\uc744 \ud30c\uc778\ud29c\ub2dd\ud558\uace0, \ud14d\uc2a4\ud2b8 \ubd84\ub958 \uc791\uc5c5\uc744 \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\ud5c8\uae45\ud398\uc774\uc2a4(Hugging Face)\ub294 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubaa8\ub378\uc744 \uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ub2e4\uc591\ud55c \ubaa8\ub378\uacfc \ud234\uc744 \uc81c\uacf5\ud558\ub294 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. \ud2b9\ud788, BERT\uc640 \uac19\uc740 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \uae30\ubc18\uc758 \ubaa8\ub378\uc744 \uac04\ud3b8\ud558\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ud574\uc90d\ub2c8\ub2e4. BERT \ubaa8\ub378\uc740 \ubb38\ub9e5\uc758 \uc591\ucabd \uc815\ubcf4\ub97c \ub3d9\uc2dc\uc5d0 \uace0\ub824\ud558\uc5ec \ub354 \uae4a\uc774 \uc788\ub294 \uc774\ud574\ub97c \uac00\ub2a5\ud558\uac8c \ud569\ub2c8\ub2e4. \uc774\ub294 \uc804\ud1b5\uc801\uc778 RNN\uc774\ub098 LSTM \uae30\ubc18 \ubaa8\ub378\ub4e4\ubcf4\ub2e4 \ub354 \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubc1c\ud718\ud560 \uc218 \uc788\ub294 \uc774\uc720\uc785\ub2c8\ub2e4.<\/p>\n
BERT\ub294 \uc778\ucf54\ub354 – \ub514\ucf54\ub354 \uad6c\uc870\uc758 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \ubaa8\ub378\ub85c, \uc8fc\ub85c \uc778\ucf54\ub354 \ubd80\ubd84\uc774 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. BERT\uc758 \uc8fc\uc694 \ud2b9\uc9d5\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\uba3c\uc800, \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574\uc11c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uba85\ub839\uc5b4\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
pip install transformers<\/code><\/pre>\n4. \ub370\uc774\ud130\uc14b \uc900\ube44\ud558\uae30<\/h2>\n
\ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\uae30 \uc704\ud574\uc11c\ub294 \uc801\uc808\ud55c \ub370\uc774\ud130\uc14b\uc774 \ud544\uc694\ud569\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 \uac04\ub2e8\ud55c \ud14d\uc2a4\ud2b8 \ubd84\ub958 \uc791\uc5c5\uc744 \uc704\ud574 IMDB \uc601\ud654 \ub9ac\ubdf0 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \ub370\uc774\ud130\uc14b\uc740 \uae0d\uc815\uc801\uc778 \ub9ac\ubdf0\uc640 \ubd80\uc815\uc801\uc778 \ub9ac\ubdf0\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\nimport pandas as pd\nfrom sklearn.model_selection import train_test_split\n\n# IMDB \ub370\uc774\ud130\uc14b \ub85c\ub4dc\nurl = 'https:\/\/ai.stanford.edu\/~amaas\/data\/sentiment\/aclImdb_v1.tar.gz'\n# \ub370\uc774\ud130\ub97c \ub2e4\uc6b4\ub85c\ub4dc\ud558\uace0 \uc555\ucd95 \ud574\uc81c\ud569\ub2c8\ub2e4.\n!wget {url}\n!tar -xvf aclImdb_v1.tar.gz\n\n# \uae0d\uc815 \ubc0f \ubd80\uc815 \ub9ac\ubdf0 \ub85c\ub4dc\npos_reviews = pd.read_csv('aclImdb\/train\/pos\/*.txt', delimiter=\"\\n\", header=None)\nneg_reviews = pd.read_csv('aclImdb\/train\/neg\/*.txt', delimiter=\"\\n\", header=None)\n\n# \ub370\uc774\ud130 \uc900\ube44\npositive = [(1, review) for review in pos_reviews[0]]\nnegative = [(0, review) for review in neg_reviews[0]]\n\ndata = positive + negative\ndf = pd.DataFrame(data, columns=['label', 'review'])\ndf['label'] = df['label'].map({0: 'negative', 1: 'positive'})\n\n# \ud6c8\ub828 \ub370\uc774\ud130\uc640 \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\ub85c \ubd84\ub9ac\ntrain_df, test_df = train_test_split(df, test_size=0.2, random_state=42)\n <\/code><\/pre>\n5. \ub370\uc774\ud130 \uc804\ucc98\ub9ac<\/h2>\n
BERT \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uae30 \uc704\ud574\uc11c\ub294 \ub370\uc774\ud130\ub97c \uc801\uc808\ud55c \ud615\uc2dd\uc73c\ub85c \uc804\ucc98\ub9ac\ud574\uc57c \ud569\ub2c8\ub2e4. \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc5d0\uc11c \uc81c\uacf5\ud558\ub294 BERT \ud1a0\ud06c\ub098\uc774\uc800\ub97c \uc0ac\uc6a9\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\nfrom transformers import BertTokenizer\n\n# BertTokenizer \ub85c\ub4dc\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\n\n# \ub370\uc774\ud130 \ud1a0\ud070\ud654\ndef tokenize_data(data):\n return tokenizer(data['review'].tolist(), padding=True, truncation=True, return_tensors='pt')\n\ntrain_encodings = tokenize_data(train_df)\ntest_encodings = tokenize_data(test_df)\n <\/code><\/pre>\n6. \ub370\uc774\ud130\uc14b \uc0dd\uc131<\/h2>\n
\ud1a0\ud070\ud654\ub41c \ub370\uc774\ud130\ub97c Dataset \ud074\ub798\uc2a4\ub97c \uc0ac\uc6a9\ud574 PyTorch Dataset\uc73c\ub85c \ubcc0\ud658\ud569\ub2c8\ub2e4.<\/p>\n
\nimport torch\n\nclass IMDbDataset(torch.utils.data.Dataset):\n def __init__(self, encodings, labels):\n self.encodings = encodings\n self.labels = labels\n\n def __getitem__(self, idx):\n item = {key: val[idx] for key, val in self.encodings.items()}\n item['labels'] = self.labels[idx]\n return item\n\n def __len__(self):\n return len(self.labels)\n\ntrain_dataset = IMDbDataset(train_encodings, train_df['label'].tolist())\ntest_dataset = IMDbDataset(test_encodings, test_df['label'].tolist())\n <\/code><\/pre>\n7. \ubaa8\ub378 \uc124\uc815 \ubc0f \ud30c\uc778\ud29c\ub2dd<\/h2>\n
\uc774\uc81c BERT \ubaa8\ub378\uc744 \ub85c\ub4dc\ud558\uace0, \ubd84\ub958 \uc791\uc5c5\uc744 \uc704\ud574 \ud30c\uc778\ud29c\ub2dd\uc744 \uc9c4\ud589\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\nfrom transformers import BertForSequenceClassification, Trainer, TrainingArguments\n\n# BERT \ubaa8\ub378 \ub85c\ub4dc\nmodel = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)\n\n# \ud6c8\ub828 \uc778\uc790 \uc124\uc815\ntraining_args = TrainingArguments(\n output_dir='.\/results',\n evaluation_strategy='epoch',\n learning_rate=2e-5,\n per_device_train_batch_size=16,\n per_device_eval_batch_size=64,\n num_train_epochs=3,\n weight_decay=0.01,\n)\n\n# \ud2b8\ub808\uc774\ub108 \uac1d\uccb4 \uc0dd\uc131\ntrainer = Trainer(\n model=model,\n args=training_args,\n train_dataset=train_dataset,\n eval_dataset=test_dataset,\n)\n \n# \ubaa8\ub378 \ud6c8\ub828\ntrainer.train()\n <\/code><\/pre>\n8. \ud3c9\uac00 \ubc0f \uc608\uce21<\/h2>\n
\ubaa8\ub378 \ud6c8\ub828 \ud6c4, \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc14b\uc744 \uc774\uc6a9\ud558\uc5ec \ubaa8\ub378 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\uace0 \uc608\uce21\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\n# \ubaa8\ub378 \ud3c9\uac00\ntrainer.evaluate()\n\n# \uc608\uce21\npredictions = trainer.predict(test_dataset)\npredicted_labels = predictions.predictions.argmax(-1)\n <\/code><\/pre>\n9. \uacb0\uacfc \ud574\uc11d<\/h2>\n
\ubaa8\ub378\uc774 \uc608\uce21\ud55c \ub77c\ubca8\uc744 \uc6d0\ubcf8 \ub77c\ubca8\uacfc \ube44\uad50\ud558\uc5ec \uc815\ud655\ub3c4\ub97c \uacc4\uc0b0\ud558\uace0, \uc218\uc815\ud560 \ubd80\ubd84\uc774 \uc788\ub294\uc9c0 \ubd84\uc11d\ud569\ub2c8\ub2e4. Confusion Matrix\ub97c \ud1b5\ud574\uc11c \ub354\uc6b1 \uc790\uc138\ud55c \uacb0\uacfc\ub97c \uc2dc\uac01\ud654\ud560 \uc218\ub3c4 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\nfrom sklearn.metrics import confusion_matrix\nimport seaborn as sns\nimport matplotlib.pyplot as plt\n\nconfusion_mtx = confusion_matrix(test_df['label'].tolist(), predicted_labels)\nplt.figure(figsize=(10,7))\nsns.heatmap(confusion_mtx, annot=True, fmt='d', cmap='Blues')\nplt.xlabel('Predicted')\nplt.ylabel('True')\nplt.title('Confusion Matrix')\nplt.show()\n <\/code><\/pre>\n10. \uacb0\ub860<\/h2>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud65c\uc6a9\ud558\uc5ec BERT \ubaa8\ub378\uc744 \ud30c\uc778\ud29c\ub2dd\ud558\uace0 \ud14d\uc2a4\ud2b8 \ubd84\ub958 \uc791\uc5c5\uc744 \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. BERT\uc640 \uac19\uc740 \uc0ac\uc804 \ud6c8\ub828\ub41c \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\ub294 \uac83\uc740 \uc2dc\uac04\uacfc \uc790\uc6d0\uc744 \uc808\uc57d\ud560 \ubfd0\ub9cc \uc544\ub2c8\ub77c \uc131\ub2a5\ub3c4 \ud06c\uac8c \ud5a5\uc0c1\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c\ub3c4 \ub2e4\uc591\ud55c NLP \uc791\uc5c5\uc5d0\uc11c BERT\uc640 \uac19\uc740 \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uc5ec \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \ub2ec\uc131\ud560 \uc218 \uc788\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0\uc0ac\ud56d<\/h2>\n