\ub525\ub7ec\ub2dd\uc758 \ubc1c\uc804\uacfc \ud568\uaed8 \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP) \ubd84\uc57c\uc5d0\uc11c\ub3c4 \ub9ce\uc740 \ud601\uc2e0\uc774 \uc774\ub8e8\uc5b4\uc84c\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \ud5c8\uae45\ud398\uc774\uc2a4(Hugging Face)\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38(Transformers) \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \uc5ec\ub7ec \uac15\ub825\ud55c \uc0ac\uc804 \ud6c8\ub828(pre-trained) \ubaa8\ub378\uc744 \uc81c\uacf5\ud558\uba70, \uc5f0\uad6c\uc790\ub098 \uac1c\ubc1c\uc790\uac00 \uc190\uc27d\uac8c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubaa8\ub378\uc744 \ud65c\uc6a9\ud560 \uc218 \uc788\uac8c \ud574\uc8fc\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uac15\uc88c\uc5d0\uc11c\ub294 BERT(Bidirectional Encoder Representations from Transformers) \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud14d\uc2a4\ud2b8 \ubd84\ub958 \uc791\uc5c5\uc744 \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc744 \uc790\uc138\ud558\uac8c \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
BERT\ub294 Google\uc5d0\uc11c \ubc1c\ud45c\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubaa8\ub378\ub85c, ‘\uc591\ubc29\ud5a5'(Bidirectional)\uc774\ub77c\ub294 \ud2b9\uc9d5\uc774 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 \ud14d\uc2a4\ud2b8\uc758 \ucee8\ud14d\uc2a4\ud2b8\ub97c \uc774\ud574\ud558\ub294 \ub370 \ub9e4\uc6b0 \uac15\ub825\ud55c \uae30\ub2a5\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. BERT\ub294 \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud560 \ub54c, \ub2e8\uc5b4\uc758 \uc704\uce58\uc5d0 \uad00\uacc4\uc5c6\uc774 \ubb38\ub9e5\uc744 \uc774\ud574\ud558\uae30 \ub54c\ubb38\uc5d0 \uc804\ud1b5\uc801\uc778 \ub2e8\uc5b4 \uc784\ubca0\ub529 \uae30\ubc95\ubcf4\ub2e4 \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubcf4\uc785\ub2c8\ub2e4.<\/p>\n
\ud5c8\uae45\ud398\uc774\uc2a4\uc758 Transformers \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 BERT\ub97c \ud3ec\ud568\ud55c \ub2e4\uc591\ud55c \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ubaa8\ub378\uc744 \uc190\uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub294 \ud30c\uc774\uc36c \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. \ud604\uc7ac NLP \ubd84\uc57c\uc5d0\uc11c \ub110\ub9ac \uc0ac\uc6a9\ub418\uace0 \uc788\uc73c\uba70, \uc0ac\uc804 \ud6c8\ub828\ub41c \ubaa8\ub378\uc744 \ubd88\ub7ec\uc640\uc11c \ud30c\uc778\ud29c\ub2dd(fine-tuning)\ud558\uc5ec \ud2b9\uc815 \uc791\uc5c5\uc5d0 \ub9de\uac8c \ud6a8\uc728\uc801\uc73c\ub85c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\ud5c8\uae45\ud398\uc774\uc2a4 Transformers \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uae30 \uc704\ud574 \uc544\ub798\uc758 pip \uba85\ub839\uc5b4\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4:<\/p>\n
pip install transformers<\/code><\/pre>\n3. BERT\ub85c \ud14d\uc2a4\ud2b8 \ubd84\ub958\ud558\uae30<\/h2>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 IMDB \uc601\ud654 \ub9ac\ubdf0 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc601\ud654 \ub9ac\ubdf0\uac00 \uae0d\uc815\uc801\uc778\uc9c0 \ubd80\uc815\uc801\uc778\uc9c0\ub97c \ubd84\ub958\ud558\ub294 \ubaa8\ub378\uc744 \uad6c\ud604\ud569\ub2c8\ub2e4. \ub370\uc774\ud130\uc14b\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \ud615\ud0dc\ub85c \ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \ud14d\uc2a4\ud2b8: \uc601\ud654 \ub9ac\ubdf0<\/li>\n
- \ub808\uc774\ube14: \uae0d\uc815(1) \ub610\ub294 \ubd80\uc815(0)<\/li>\n<\/ul>\n
3.1 \ub370\uc774\ud130\uc14b \uc900\ube44\ud558\uae30<\/h3>\n
\uba3c\uc800 \ub370\uc774\ud130\uc14b\uc744 \ub2e4\uc6b4\ub85c\ub4dc\ud558\uace0 \uc804\ucc98\ub9ac\ud569\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\ud558\uae30\nurl = \"https:\/\/ai.stanford.edu\/~amaas\/data\/sentiment\/aclImdb_v1.tar.gz\"\n# \ub370\uc774\ud130\ub97c \ub85c\ub4dc\ud558\uace0 \uc804\ucc98\ub9ac\n# \uc5ec\uae30\uc5d0 \ub370\uc774\ud130 \ub85c\ub529 \ubc0f \uc804\ucc98\ub9ac \ucf54\ub4dc\uac00 \uc635\ub2c8\ub2e4. (\uac04\ub2e8\ud55c \uc608\uc2dc\uc785\ub2c8\ub2e4)\ndata = pd.read_csv(\"imdb_reviews.csv\")\ndata['label'] = data['sentiment'].apply(lambda x: 1 if x == 'positive' else 0)\n\ntrain_texts, val_texts, train_labels, val_labels = train_test_split(data['review'], data['label'], test_size=0.2)\n <\/code><\/pre>\n3.2 BERT Tokenization<\/h3>\n
\ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\ub97c BERT \ubaa8\ub378\uc5d0 \ub9de\uac8c \ubcc0\ud658\ud558\uae30 \uc704\ud574 \ud1a0\ud06c\ub098\uc774\uc800(tokenizer)\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \ud1a0\ud06c\ub098\uc774\uc800\ub294 \ud14d\uc2a4\ud2b8\ub97c \ubd84\ud560\ud558\uc5ec \ubaa8\ub378 \uc785\ub825 \ud615\uc2dd\uc73c\ub85c \ubcc0\ud658\ud558\ub294 \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4.<\/p>\n
\nfrom transformers import BertTokenizer\n\n# BERT \ud1a0\ud06c\ub098\uc774\uc800 \ucd08\uae30\ud654\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\n\n# BERT \uc785\ub825 \ud615\ud0dc\ub85c \ubcc0\ud658\ud558\ub294 \ud568\uc218\ndef encode(texts):\n return tokenizer(texts.tolist(), padding=True, truncation=True, return_tensors='pt')\n\ntrain_encodings = encode(train_texts)\nval_encodings = encode(val_texts)\n <\/code><\/pre>\n3.3 \ub370\uc774\ud130\uc14b \ub9cc\ub4e4\uae30<\/h3>\n
\ud1a0\ud06c\ub098\uc774\uc800\uac00 \ubcc0\ud658\ud55c \uc778\ucf54\ub529\uc744 \ud30c\uc774\ud1a0\uce58 \ud150\uc11c\ub85c \ubcc0\ud658\ud558\uc5ec \ub370\uc774\ud130\uc14b\uc744 \ub9cc\ub4ed\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'] = torch.tensor(self.labels[idx])\n return item\n\n def __len__(self):\n return len(self.labels)\n\ntrain_dataset = IMDbDataset(train_encodings, train_labels.values)\nval_dataset = IMDbDataset(val_encodings, val_labels.values)\n <\/code><\/pre>\n4. BERT \ubaa8\ub378 \uc815\uc758\ud558\uae30<\/h3>\n
\uc774\uc81c Hugging Face\uc758 Transforms \ub77c\uc774\ube0c\ub7ec\ub9ac\uc5d0\uc11c \uc81c\uacf5\ud558\ub294 BERT \ubaa8\ub378\uc744 \ubd88\ub7ec\uc640\uc11c \ubd84\ub958 \uc791\uc5c5\uc744 \uc704\ud55c \ud30c\uc778\ud29c\ub2dd\uc744 \uc9c4\ud589\ud569\ub2c8\ub2e4.<\/p>\n
\nfrom transformers import BertForSequenceClassification\n\n# BERT \ubaa8\ub378 \ubd88\ub7ec\uc624\uae30\nmodel = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)\n <\/code><\/pre>\n5. \ubaa8\ub378 \ud6c8\ub828\ud558\uae30<\/h3>\n
\ubaa8\ub378\uc744 \ud6c8\ub828\ud558\uae30 \uc704\ud574 Trainer API\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 API\ub294 \ud6c8\ub828 \ub8e8\ud504\ub97c \uc790\ub3d9\uc73c\ub85c \ucc98\ub9ac\ud558\ubbc0\ub85c \ub9e4\uc6b0 \ud3b8\ub9ac\ud569\ub2c8\ub2e4.<\/p>\n
\nfrom transformers import Trainer, TrainingArguments\n\n# \ud6c8\ub828 \ud658\uacbd \uc124\uc815\ntraining_args = TrainingArguments(\n output_dir='.\/results',\n num_train_epochs=3,\n per_device_train_batch_size=8,\n per_device_eval_batch_size=8,\n warmup_steps=500,\n weight_decay=0.01,\n logging_dir='.\/logs',\n)\n\ntrainer = Trainer(\n model=model,\n args=training_args,\n train_dataset=train_dataset,\n eval_dataset=val_dataset,\n)\n\n# \ud6c8\ub828 \uc2dc\uc791\ntrainer.train()\n <\/code><\/pre>\n6. \ubaa8\ub378 \ud3c9\uac00\ud558\uae30<\/h3>\n
\ud6c8\ub828\ub41c \ubaa8\ub378\uc744 \ud3c9\uac00\ud558\uc5ec \uc131\ub2a5\uc744 \ud655\uc778\ud569\ub2c8\ub2e4.<\/p>\n
\ntrainer.evaluate()\n <\/code><\/pre>\n\uacb0\ub860<\/h2>\n
\ubcf8 \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 \uc0ac\uc6a9\ud55c \ud14d\uc2a4\ud2b8 \ubd84\ub958 \uc791\uc5c5\uc744 \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc744 \ubc30\uc6e0\uc2b5\ub2c8\ub2e4. BERT\ub294 \ub2e4\uc591\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uc791\uc5c5\uc5d0\uc11c \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubc1c\ud718\ud558\uba70, \uc0ac\uc804 \ud6c8\ub828\ub41c \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uba74 \uc801\uc740 \ub370\uc774\ud130\ub9cc\uc73c\ub85c\ub3c4 \uc88b\uc740 \uacb0\uacfc\ub97c \ub0bc \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c \uc5ec\ub7ec\ubd84\uc774 \ub2e4\uc591\ud55c NLP \ud504\ub85c\uc81d\ud2b8\uc5d0\uc11c BERT\ub97c \ud65c\uc6a9\ud558\uae38 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0 \uc790\ub8cc<\/h2>\n