\uc791\uc131\uc790: [\uc870\uad11\ud615]<\/p>\n
\uc791\uc131\uc77c: [\ub0a0\uc9dc]<\/p>\n<\/header>\n \uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub294 \uc778\uac04 \uc5b8\uc5b4\ub97c \ucef4\ud4e8\ud130\uac00 \uc774\ud574\ud558\uace0 \ucc98\ub9ac\ud558\ub294 \uae30\uc220\ub85c, \uae30\uacc4 \ubc88\uc5ed, \uac10\uc815 \ubd84\uc11d, \uc9c8\ubb38 \uc751\ub2f5 \uc2dc\uc2a4\ud15c \ub4f1 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0 \ud65c\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud2b9\ud788 \ucd5c\uadfc\uc5d0\ub294 \ub525 \ub7ec\ub2dd \uae30\uc220\uc758 \ubc1c\uc804\uc73c\ub85c \uc778\ud574 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c\ub3c4 \ub9ce\uc740 \ud601\uc2e0\uc774 \uc77c\uc5b4\ub098\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 KoBERT \ubaa8\ub378\uc744 \uc774\uc6a9\ud558\uc5ec KorNLI \ub370\uc774\ud130\uc14b\uc744 \ud65c\uc6a9\ud55c \ub2e4\uc911 \ud074\ub798\uc2a4 \ubd84\ub958 \ubb38\uc81c\ub97c \uae4a\uc774 \uc788\uac8c \ud0d0\uad6c\ud558\uace0\uc790 \ud569\ub2c8\ub2e4.<\/p>\n<\/section>\n \uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub97c \uc704\ud574 \ub525 \ub7ec\ub2dd \uae30\uc220\uc774 \uc0ac\uc6a9\ub418\ub294 \uc774\uc720\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4. \uccab\uc9f8, \ub525 \ub7ec\ub2dd \ubaa8\ub378\uc740 \ub300\ub7c9\uc758 \ub370\uc774\ud130\ub97c \uae30\ubc18\uc73c\ub85c \ud559\uc2b5\ud560 \uc218 \uc788\uc5b4, \uc5b8\uc5b4\uc758 \ubcf5\uc7a1\ud55c \ud328\ud134\uc744 \ud559\uc2b5\ud558\ub294 \ub370 \uc801\ud569\ud569\ub2c8\ub2e4. \ub458\uc9f8, \uc2e0\uacbd\ub9dd \uc544\ud0a4\ud14d\ucc98\ub294 \ub2e4\uc591\ud55c \ub370\uc774\ud130(\ud14d\uc2a4\ud2b8, \uc774\ubbf8\uc9c0 \ub4f1)\ub97c \ud1b5\ud569\ud558\uc5ec \ucc98\ub9ac\ud560 \uc218 \uc788\ub294 \ub2a5\ub825\uc774 \uc788\uc2b5\ub2c8\ub2e4. \ub9c8\uc9c0\ub9c9\uc73c\ub85c, \ucd5c\uadfc\uc5d0\ub294 Transformer \uae30\ubc18\uc758 \ubaa8\ub378\ub4e4\uc774 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c \ub9e4\uc6b0 \ub192\uc740 \uc131\ub2a5\uc744 \ubcf4\uc774\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n KoBERT\ub294 \ud55c\uad6d\uc5b4\uc5d0 \ud2b9\ud654\ub41c BERT(Bidirectional Encoder Representations from Transformers) \ubaa8\ub378\ub85c, \ud55c\uad6d\uc5b4 \ub370\uc774\ud130\uc14b\uc744 \uc774\uc6a9\ud558\uc5ec \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\uc785\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \ud55c\uad6d\uc5b4 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uc791\uc5c5\uc5d0\uc11c \ub192\uc740 \uc131\ub2a5\uc744 \ubc1c\ud718\ud558\uba70, \ub2e4\uc591\ud55c \ud558\uc704 \uc791\uc5c5(\uac10\uc815 \ubd84\uc11d, \uac1c\uccb4\uba85 \uc778\uc2dd \ub4f1)\uc5d0 \uc27d\uac8c \uc801\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. KoBERT\uc758 \uc8fc\uc694 \ud2b9\uc9d5\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n KorNLI(Korean Natural Language Inference)\ub294 \ud55c\uad6d\uc5b4 \uc790\uc5f0\uc5b4 \ucd94\ub860(NLI) \ub370\uc774\ud130\uc14b\uc73c\ub85c, \ubb38\uc7a5 \uc30d\uc5d0 \ub300\ud55c \uad00\uacc4\ub97c \ub2e4\uc911 \ud074\ub798\uc2a4(Entailment, Neutral, Contradiction)\ub85c \ubd84\ub958\ud558\ub294 \ud0dc\uc2a4\ud06c\ub97c \uc218\ud589\ud569\ub2c8\ub2e4. \uc774 \ub370\uc774\ud130\uc14b\uc740 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubaa8\ub378\uc758 \ucd94\ub860 \ub2a5\ub825\uc744 \ud3c9\uac00\ud558\ub294 \ub370 \uc801\ud569\ud569\ub2c8\ub2e4. KorNLI \ub370\uc774\ud130\uc14b\uc758 \ud2b9\uc9d5\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\uae30 \uc704\ud574 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc8fc\ub85c PyTorch\uc640 Hugging Face\uc758 Transformers \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/p>\n KorNLI \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud558\uace0, \uc801\uc808\ud55c \ud615\uc2dd\uc73c\ub85c \uc804\ucc98\ub9ac\ud558\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4. \ubb38\uc7a5 \uc30d\uc744 \uc785\ub825\uc73c\ub85c \uc0ac\uc6a9\ud558\uace0, \ud574\ub2f9 \ub808\uc774\ube14\uc744 \ucd9c\ub825\uc73c\ub85c \ud560\ub2f9\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n KoBERT \ubaa8\ub378\uc744 \ubd88\ub7ec\uc624\uace0, \ubd84\ub958 \uc791\uc5c5\uc744 \uc704\ud55c \ub808\uc774\uc5b4\ub97c \ucd94\uac00\ud569\ub2c8\ub2e4. \uc544\ub798\ub294 \uae30\ubcf8\uc801\uc778 \ubaa8\ub378 \uc815\uc758 \ucf54\ub4dc\uc785\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378 \ud559\uc2b5\uc744 \uc704\ud574 PyTorch\uc758 DataLoader\ub97c \uc0ac\uc6a9\ud558\uc5ec \ubc30\uce58 \ub2e8\uc704\ub85c \ub370\uc774\ud130\ub97c \ub85c\ub4dc\ud569\ub2c8\ub2e4. \uc774\ud6c4 epochs \ub9cc\ud07c \ubaa8\ub378\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/p>\n \ud559\uc2b5\ub41c \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\uae30 \uc704\ud574 \uac80\uc99d \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. accuracy\uc640 \uac19\uc740 \uc131\ub2a5 \uc9c0\ud45c\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n KoBERT\ub97c \ud65c\uc6a9\ud558\uc5ec KorNLI \ub370\uc774\ud130\uc14b\uc744 \uc774\uc6a9\ud55c \ub2e4\uc911 \ud074\ub798\uc2a4 \ubd84\ub958 \ubb38\uc81c\ub97c \ud574\uacb0\ud568\uc73c\ub85c\uc368, \ud55c\uad6d\uc5b4 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc758 \ubc1c\uc804 \uac00\ub2a5\uc131\uacfc \ub525 \ub7ec\ub2dd \uae30\uc220\uc758 \ud6a8\uc6a9\uc131\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c\ub3c4 \ub525 \ub7ec\ub2dd \uae30\ubc18 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uae30\uc220\uc758 \ubc1c\uc804\uc740 \ub354\uc6b1 \uac00\uc18d\ud654\ub420 \uac83\uc774\uba70, \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c\uc758 \uc751\uc6a9\uc774 \uae30\ub300\ub429\ub2c8\ub2e4.<\/p>\n<\/section>\n1. \uc11c\ub860<\/h2>\n
2. \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc640 \ub525 \ub7ec\ub2dd<\/h2>\n
3. KoBERT \uc18c\uac1c<\/h2>\n
\n
4. KorNLI \ub370\uc774\ud130\uc14b<\/h2>\n
\n
5. KoBERT\ub97c \ud65c\uc6a9\ud55c KorNLI \ubaa8\ub378 \uad6c\ucd95<\/h2>\n
5.1. \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uce58<\/h3>\n
!pip install torch torchvision transformers<\/code><\/pre>\n5.2. \ub370\uc774\ud130 \uc804\ucc98\ub9ac<\/h3>\n
\nimport pandas as pd\n\n# \ub370\uc774\ud130\uc14b \ub85c\ub4dc\ndata = pd.read_csv('kornli_dataset.csv')\n\n# \ub370\uc774\ud130 \ud655\uc778\nprint(data.head())\n <\/code><\/pre>\n5.3. KoBERT \ubaa8\ub378 \uc815\uc758<\/h3>\n
\nfrom transformers import BertTokenizer, BertForSequenceClassification\n\n# \ud1a0\ud06c\ub098\uc774\uc800\uc640 \ubaa8\ub378 \ucd08\uae30\ud654\ntokenizer = BertTokenizer.from_pretrained('monologg\/kobert')\nmodel = BertForSequenceClassification.from_pretrained('monologg\/kobert', num_labels=3)\n <\/code><\/pre>\n5.4. \ubaa8\ub378 \ud559\uc2b5<\/h3>\n
\nfrom torch.utils.data import DataLoader, Dataset\n\nclass KorNliDataset(Dataset):\n def __init__(self, texts, labels, tokenizer, max_len):\n self.labels = labels\n self.texts = texts\n self.tokenizer = tokenizer\n self.max_len = max_len\n\n def __len__(self):\n return len(self.texts)\n\n def __getitem__(self, idx):\n text = self.texts[idx]\n label = self.labels[idx]\n\n encoding = self.tokenizer.encode_plus(\n text,\n add_special_tokens=True,\n max_length=self.max_len,\n return_token_type_ids=False,\n padding='max_length',\n return_attention_mask=True,\n return_tensors='pt',\n truncation=True\n )\n\n return {\n 'input_ids': encoding['input_ids'].flatten(),\n 'attention_mask': encoding['attention_mask'].flatten(),\n 'labels': torch.tensor(label, dtype=torch.long)\n }\n\n# DataLoader \uc815\uc758\ndataset = KorNliDataset(data['text'].values, data['label'].values, tokenizer, max_len=128)\ndataloader = DataLoader(dataset, batch_size=16, shuffle=True)\n\n# \uc635\ud2f0\ub9c8\uc774\uc800 \uc815\uc758\noptimizer = torch.optim.AdamW(model.parameters(), lr=2e-5)\n\n# \ubaa8\ub378 \ud559\uc2b5\nmodel.train()\nfor epoch in range(epochs):\n for batch in dataloader:\n optimizer.zero_grad()\n outputs = model(\n input_ids=batch['input_ids'],\n attention_mask=batch['attention_mask'],\n labels=batch['labels']\n )\n loss = outputs.loss\n loss.backward()\n optimizer.step()\n <\/code><\/pre>\n5.5. \ubaa8\ub378 \ud3c9\uac00<\/h3>\n
\nmodel.eval()\npredictions = []\ntrue_labels = []\n\nwith torch.no_grad():\n for batch in validation_dataloader:\n outputs = model(\n input_ids=batch['input_ids'],\n attention_mask=batch['attention_mask']\n )\n\n preds = torch.argmax(outputs.logits, dim=1)\n predictions.extend(preds.numpy())\n true_labels.extend(batch['labels'].numpy())\n\n# \uc815\ud655\ub3c4 \uacc4\uc0b0\nfrom sklearn.metrics import accuracy_score\naccuracy = accuracy_score(true_labels, predictions)\nprint(f'Accuracy: {accuracy}')\n <\/code><\/pre>\n<\/section>\n6. \uacb0\ub860<\/h2>\n