\ub525 \ub7ec\ub2dd\uc758 \ubc1c\uc804\uacfc \ud568\uaed8 \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\uc758 \ub2e4\uc591\ud55c \uacfc\uc81c\ub4e4\uc774 \ud574\uacb0\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\uc911\uc5d0\uc11c\ub3c4 \ud55c\uad6d\uc5b4\uc640 \uac19\uc740 \uace0\ub3c4\ub85c \uc735\ud569\ub41c \uc5b8\uc5b4\uc758 \ucc98\ub9ac\ub294 \ub3c4\uc804 \uacfc\uc81c\ub85c \ub0a8\uc544 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 \ub9c8\uc2a4\ud06c\ub4dc \uc5b8\uc5b4 \ubaa8\ub378(Masked Language Model, MLM)\uc758 \uac1c\ub150\uacfc \ud55c\uad6d\uc5b4 BERT \ubaa8\ub378\uc744 \ud65c\uc6a9\ud55c \uc2e4\uc2b5 \uacfc\uc815\uc744 \uc790\uc138\ud788 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub294 \uc778\uac04\uc758 \uc5b8\uc5b4\ub97c \ucef4\ud4e8\ud130\uac00 \uc774\ud574\ud558\uace0 \ucc98\ub9ac\ud560 \uc218 \uc788\ub3c4\ub85d \ud558\ub294 \uae30\uc220\ub85c, \ud14d\uc2a4\ud2b8 \ubd84\uc11d, \uae30\uacc4 \ubc88\uc5ed, \uac10\uc815 \ubd84\uc11d \ub4f1 \uad11\ubc94\uc704\ud55c \ubd84\uc57c\uc5d0 \uac78\uccd0 \ud65c\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ucd5c\uadfc \ub525 \ub7ec\ub2dd \uae30\ubc18\uc758 \ubaa8\ub378\ub4e4\uc774 \uc774\ub7ec\ud55c \uc791\uc5c5\uc744 \uc218\ud589\ud558\ub294 \ub370 \uc788\uc5b4 \ud070 \uc774\uc810\uc744 \uc81c\uacf5\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub294 \uc778\uacf5\uc9c0\ub2a5\uc758 \uc911\uc694\ud55c \ubd84\uc57c \uc911 \ud558\ub098\ub85c, \uc778\uac04\uacfc \ucef4\ud4e8\ud130 \uac04\uc758 \uc0c1\ud638\uc791\uc6a9\uc744 \uac1c\uc120\ud558\uace0, \uc815\ubcf4 \uac80\uc0c9, \ub370\uc774\ud130 \ubd84\uc11d \ub4f1\uc5d0 \uae30\uc5ec\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \uc0ac\uc6a9\uc790\uc758 \ub300\ud654, \uac80\uc0c9 \ucffc\ub9ac, \uace0\uac1d \ud53c\ub4dc\ubc31 \ub4f1\uc744 \uc774\ud574\ud558\ub294 \ub370 \ud544\uc218\uc801\uc785\ub2c8\ub2e4.<\/p>\n
BERT(Bidirectional Encoder Representations from Transformers)\ub294 \uad6c\uae00\uc774 \uac1c\ubc1c\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubaa8\ub378\ub85c, MLM\uacfc Next Sentence Prediction(NSP) \uc791\uc5c5\uc744 \ud1b5\ud574 \ubb38\ub9e5\uc744 \uc774\ud574\ud558\ub294 \ub370 \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. BERT\ub294 \uc591\ubc29\ud5a5 Transformer \uc778\ucf54\ub354\ub97c \uc0ac\uc6a9\ud558\uc5ec \ubb38\uc7a5\uc758 \ubaa8\ub4e0 \ub2e8\uc5b4\ub97c \ub3d9\uc2dc\uc5d0 \uace0\ub824\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
BERT\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uad6c\uc131 \uc694\uc18c\ub85c \uc124\uba85\ub429\ub2c8\ub2e4:<\/p>\n
\ub9c8\uc2a4\ud06c\ub4dc \uc5b8\uc5b4 \ubaa8\ub378\uc740 \ud2b9\uc815 \ub2e8\uc5b4\ub97c \ub9c8\uc2a4\ud0b9\ud558\uc5ec \uc774\ub97c \uc608\uce21\ud558\ub294 \uc791\uc5c5\uc785\ub2c8\ub2e4. BERT\ub294 \uc785\ub825 \ubb38\uc7a5\uc5d0\uc11c \ub2e8\uc5b4\uc758 15%\ub97c \ubb34\uc791\uc704\ub85c \uc120\ud0dd\ud558\uc5ec ‘[MASK]’ \ud1a0\ud070\uc73c\ub85c \ub300\uccb4\ud55c \ub4a4, \ubaa8\ub378\uc774 \uc774 \ub9c8\uc2a4\ud06c\ub41c \ub2e8\uc5b4\ub97c \uc608\uce21\ud558\ub3c4\ub85d \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc774 \ubc29\ubc95\uc740 \ubb38\ub9e5\uc744 \uc774\ud574\ud558\uace0 \uc885\ub958\uac00 \ub2e4\uc591\ud55c \ubb38\uc7a5\uc744 \uc0dd\uc131\ud558\ub294 \ub370 \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4.<\/p>\n
\ud55c\uad6d\uc5b4\uc5d0 \ud2b9\ud654\ub41c BERT \ubaa8\ub378\uc740 \ud55c\uad6d\uc5b4\uc758 \ubb38\ubc95\uc801 \ud2b9\uc131\uacfc \uc5b4\ud718\ub97c \ubc18\uc601\ud558\uc5ec \ud6c8\ub828\ub418\uc5c8\uc2b5\ub2c8\ub2e4. Hugging Face\uc758 Tranformers \ub77c\uc774\ube0c\ub7ec\ub9ac\uc5d0\uc11c\ub294 \ud55c\uad6d\uc5b4 BERT \ubaa8\ub378\uc744 \uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub294 API\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n
\ud55c\uad6d\uc5b4 BERT\ub294 \ub2e4\uc591\ud55c \ud55c\uad6d\uc5b4 \ub9d0\ubb49\uce58\ub97c \uae30\ubc18\uc73c\ub85c \ud6c8\ub828\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \ud55c\uad6d\uc5b4\uc758 \ub2e4\uc591\ud55c \ubb38\ub9e5\uacfc \uc758\ubbf8\ub97c \uc774\ud574\ud560 \uc218 \uc788\ub294 \ub2a5\ub825\uc744 \uac16\ucd94\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
\uc774\uc81c \ud55c\uad6d\uc5b4 BERT\ub97c \ud65c\uc6a9\ud558\uc5ec \ub9c8\uc2a4\ud06c\ub4dc \uc5b8\uc5b4 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\ub294 \uc2e4\uc2b5\uc744 \uc9c4\ud589\ud558\uaca0\uc2b5\ub2c8\ub2e4. Python\uacfc Hugging Face Transformers \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud658\uacbd\uc744 \uc124\uc815\ud569\ub2c8\ub2e4.<\/p>\n
pip install transformers\npip install torch\npip install tokenizers<\/code><\/pre>\n4.2 \uc2e4\uc2b5 \ucf54\ub4dc<\/h3>\n
\uc544\ub798 \ucf54\ub4dc\ub294 \ud55c\uad6d\uc5b4 BERT \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ubb38\uc7a5\uc5d0\uc11c \ud2b9\uc815 \ub2e8\uc5b4\ub97c \ub9c8\uc2a4\ud0b9\ud558\uace0 \uc774\ub97c \uc608\uce21\ud558\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4.<\/p>\n
from transformers import BertTokenizer, BertForMaskedLM\nimport torch\n\n# \ud1a0\ud06c\ub098\uc774\uc800\uc640 \ubaa8\ub378 \ubd88\ub7ec\uc624\uae30\ntokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-cased')\nmodel = BertForMaskedLM.from_pretrained('bert-base-multilingual-cased')\n\n# \uc608\uc2dc \ubb38\uc7a5\ntext = \"\ub098\ub294 [MASK]\ub97c \uc88b\uc544\ud55c\ub2e4.\"\n\n# \uc785\ub825 \ub370\uc774\ud130 \uc900\ube44\ninput_ids = tokenizer.encode(text, return_tensors='pt')\n\n# \ub9c8\uc2a4\ud06c\ub41c \ud1a0\ud070\uc758 \uc778\ub371\uc2a4 \ucc3e\uae30\nmask_index = torch.where(input_ids == tokenizer.mask_token_id)[1]\n\n# \uc608\uce21 \uc218\ud589\nwith torch.no_grad():\n outputs = model(input_ids)\n predictions = outputs[0]\n\n# \ub9c8\uc2a4\ud06c\ub41c \ud1a0\ud070 \uc608\uce21\npredicted_index = torch.argmax(predictions[0, mask_index], dim=1)\npredicted_token = tokenizer.decode(predicted_index)\n\nprint(f\"\uc608\uce21\ub41c \ub2e8\uc5b4: {predicted_token}\")<\/code><\/pre>\n\uc704 \ucf54\ub4dc\uc5d0\uc11c\ub294 \uc6b0\uc120 BERT \ubaa8\ub378\uacfc \ud55c\uad6d\uc5b4\ub97c \uc9c0\uc6d0\ud558\ub294 \ud1a0\ud06c\ub098\uc774\uc800\ub97c \ubd88\ub7ec\uc628 \ud6c4, \ub9c8\uc2a4\ud0b9\ub41c \ubb38\uc7a5\uc744 \uc785\ub825\uc73c\ub85c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \ubaa8\ub378\uc740 \ub9c8\uc2a4\ud06c\ub41c \uc704\uce58\uc5d0 \ud574\ub2f9\ud558\ub294 \ub2e8\uc5b4\ub97c \uc608\uce21\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
5. \ubaa8\ub378 \ud3c9\uac00<\/h2>\n
\ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\uae30 \uc704\ud574\uc11c\ub294 \uc5ec\ub7ec \ub2e4\uc591\ud55c \ubb38\uc7a5\uacfc \ube44\uc728\uc758 \ub9c8\uc2a4\ud06c\ub97c \uc801\uc6a9\ud558\uc5ec \uc77c\ubc18\ud654\ub41c \uacb0\uacfc\ub97c \ub3c4\ucd9c\ud574\uc57c \ud569\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc5d0\uc11c \uc815\ud655\ub3c4(accuracy), F1 \uc2a4\ucf54\uc5b4 \ub4f1\uc758 \uc9c0\ud45c\ub97c \ud65c\uc6a9\ud558\uc5ec \ubaa8\ub378\uc758 \uc2e0\ub8b0\uc131\uc744 \uac80\uc99d\ud569\ub2c8\ub2e4.<\/p>\n
5.1 \ud3c9\uac00 \uc9c0\ud45c<\/h3>\n
\ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\ub294 \uc8fc\uc694 \uc9c0\ud45c\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \uc815\ud655\ub3c4(Accuracy):<\/strong> \ubaa8\ub378\uc774 \uc62c\ubc14\ub974\uac8c \uc608\uce21\ud55c \uacbd\uc6b0\uc758 \ube44\uc728.<\/li>\n
- F1 \uc2a4\ucf54\uc5b4:<\/strong> \uc815\ubc00\ub3c4(precision)\uc640 \uc7ac\ud604\uc728(recall)\uc758 \uc870\ud654 \ud3c9\uade0.<\/li>\n<\/ul>\n
6. \uacb0\ub860<\/h2>\n
\uc774\ubc88 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 \ub525 \ub7ec\ub2dd \uae30\ubc18\uc758 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc5d0\uc11c \ud55c\uad6d\uc5b4 BERT\uc758 \ub9c8\uc2a4\ud06c\ub4dc \uc5b8\uc5b4 \ubaa8\ub378\uc744 \ud65c\uc6a9\ud55c \uc2e4\uc2b5\uc744 \uc9c4\ud589\ud588\uc2b5\ub2c8\ub2e4. \ud55c\uad6d\uc5b4 \ucc98\ub9ac\uc758 \ubcf5\uc7a1\uc131\uc744 \uace0\ub824\ud560 \ub54c, BERT\uc640 \uac19\uc740 \uace0\uae09 \ubaa8\ub378\uc744 \ud65c\uc6a9\ud568\uc73c\ub85c\uc368 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc758 \uc815\ud655\ub3c4\ub97c \ub192\uc77c \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c\ub3c4 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uae30\uc220\uc774 \ubc1c\uc804\ud558\uc5ec \ub9ce\uc740 \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ub420 \uc218 \uc788\uae30\ub97c \uae30\ub300\ud569\ub2c8\ub2e4.<\/p>\n
6.1 \ucc38\uace0 \ubb38\ud5cc<\/h3>\n