\n1. \uc11c\ub860<\/h2>\n
\ucd5c\uadfc \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\uc758 \uc911\uc694\uc131\uc774 \uc810\uc810 \ucee4\uc9d0\uc5d0 \ub530\ub77c, \ub2e4\uc591\ud55c \ub525\ub7ec\ub2dd \ubaa8\ub378\uc774 \uac1c\ubc1c\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\uc911\uc5d0\uc11c\ub3c4 \ud5c8\uae45\ud398\uc774\uc2a4(Hugging Face)<\/strong>\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38(Transformers)<\/strong> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \uc5ec\ub7ec \uac00\uc9c0 NLP \ud0dc\uc2a4\ud06c\uc5d0\uc11c \ud65c\uc6a9\ub418\ub294 \uc720\uba85\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 DistilBERT<\/strong> \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uc5ec \ub9c8\uc2a4\ud06c\ub4dc \uc5b8\uc5b4 \ubaa8\ub378\ub9c1(Masking Language Modeling, MLM) \ud30c\uc774\ud504\ub77c\uc778\uc744 \uad6c\ucd95\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 discuss\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n\n2. \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \uc18c\uac1c<\/h2>\n
\ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 BERT, GPT-2, T5\uc640 \uac19\uc774 \ub2e4\uc591\ud55c \uc804\uc774 \ud559\uc2b5 \ubaa8\ub378\uc744 \uc81c\uacf5\ud558\uba70, NLP \uad00\ub828 \ud0dc\uc2a4\ud06c\uc5d0 \ub300\ud55c \uace0\uc131\ub2a5\uc758 \ucc98\ub9ac \ub2a5\ub825\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc774 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \uc27d\uac8c \ubaa8\ub378\uc744 \ub85c\ub529\ud558\uace0 \uc0ac\uc6a9\ud560 \uc218 \uc788\uac8c \ub3d5\ub294 API\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n
\n- \uac04\ud3b8\ud55c API:<\/strong> \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800\ub97c \uc190\uc27d\uac8c \ub85c\ub4dc\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n
- \ub2e4\uc591\ud55c \ubaa8\ub378:<\/strong> BERT, GPT, T5 \ub4f1 \ub2e4\uc591\ud55c \ucd5c\uc2e0 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n
- \ucee4\ubba4\ub2c8\ud2f0\uc758 \uc9c0\uc6d0:<\/strong> \ud65c\uc131\ud654\ub41c \ucee4\ubba4\ub2c8\ud2f0\uc640 \uc9c0\uc18d\uc801\uc778 \uc5c5\ub370\uc774\ud2b8\uac00 \uc774\ub8e8\uc5b4\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<\/ul>\n<\/section>\n\n
3. DistilBERT\ub780?<\/h2>\n
DistilBERT<\/strong>\ub294 BERT \ubaa8\ub378\uc758 \uacbd\ub7c9\ud654 \ubaa8\ub378\ub85c, \uc6d0\ub798\uc758 BERT\ubcf4\ub2e4 60% \ub354 \ube60\ub974\uace0 40% \ub354 \uc801\uc740 \ud30c\ub77c\ubbf8\ud130\ub97c \uac00\uc9c4 \ubaa8\ub378\uc785\ub2c8\ub2e4. \uadf8\ub7fc\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0 \ube44\uc2b7\ud55c \uc131\ub2a5\uc744 \uc720\uc9c0\ud558\uba74\uc11c \uc2e4\ubb34\uc5d0\uc11c \ub354\uc6b1 \uc720\uc6a9\ud558\uac8c \uc0ac\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n\uc774 \ubaa8\ub378\uc740 \uc5ec\ub7ec NLP \ud0dc\uc2a4\ud06c\uc5d0\uc11c \uc131\uacf5\uc801\uc73c\ub85c \uc0ac\uc6a9\ub418\uace0 \uc788\uc73c\uba70, \ud2b9\ud788 \ubb38\ub9e5 \uc774\ud574\uc640 \uad00\ub828\ub41c \uc791\uc5c5\uc5d0\uc11c \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubcf4\uc785\ub2c8\ub2e4.<\/p>\n<\/section>\n\n4. MLM(Masked Language Modeling) \ud30c\uc774\ud504\ub77c\uc778 \uc774\ud574\ud558\uae30<\/h2>\n
MLM\uc740 \ubaa8\ub974\ub294 \ub2e8\uc5b4\ub97c \ubb38\ub9e5\uc5d0\uc11c \uc608\uce21\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, “\ub098\ub294 [MASK]\ub97c \uc88b\uc544\ud569\ub2c8\ub2e4.”\uc640 \uac19\uc774 `[MASK]`\ub85c \ud45c\uc2dc\ub41c \ubd80\ubd84\uc5d0 \ub4e4\uc5b4\uac08 \ub2e8\uc5b4\ub97c \uc608\uce21\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uae30\ubc95\uc740 BERT\uc640 \uadf8 \ud30c\uc0dd \ubaa8\ub378\ub4e4\uc774 \ud6c8\ub828\ub418\ub294 \ubc29\uc2dd \uc911 \ud558\ub098\uc785\ub2c8\ub2e4.<\/p>\n
MLM\uc758 \uc8fc\uc694 \uc774\uc810\uc740 \ubaa8\ub378\uc774 \uc5b8\uc5b4\uc758 \ub2e4\uc591\ud55c \ud328\ud134\uc744 \ud559\uc2b5\ud560 \uc218 \uc788\ub3c4\ub85d \ub3d5\ub294 \uac83\uc774\uba70, \uc774\ub294 \uc790\uc5f0\uc5b4 \uc774\ud574\uc758 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \ub370 \ub3c4\uc6c0\uc744 \uc90d\ub2c8\ub2e4.<\/p>\n<\/section>\n\n5. DistilBERT \ubaa8\ub378 \ub85c\ub529\ud558\uae30<\/h2>\n
\uc774\uc81c DistilBERT \ubaa8\ub378\uc744 \ub85c\ub529\ud558\uace0 \uac04\ub2e8\ud55c MLM \ud30c\uc774\ud504\ub77c\uc778\uc744 \uad6c\ucd95\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uba3c\uc800 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\n pip install transformers torch<\/code>\n <\/pre>\n5.1 DistilBERT \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800 \ub85c\ub529\ud558\uae30<\/h3>\n
\ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc774\uc6a9\ud574 DistilBERT \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800\ub97c \ub85c\ub529\ud569\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 \ub2e4\uc74c\uacfc \uac19\uc740 \ucf54\ub4dc\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\n \nfrom transformers import DistilBertTokenizer, DistilBertForMaskedLM\nimport torch\n\n# DistilBERT \ud1a0\ud06c\ub098\uc774\uc800\uc640 \ubaa8\ub378 \ub85c\ub529\ntokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')\nmodel = DistilBertForMaskedLM.from_pretrained('distilbert-base-uncased')\n <\/code>\n <\/pre>\n\uc774 \ucf54\ub4dc\ub294 DistilBERT \ubaa8\ub378\uacfc \uadf8\uc5d0 \uc0c1\uc751\ud558\ub294 \ud1a0\ud06c\ub098\uc774\uc800\ub97c \ub85c\ub4dc\ud569\ub2c8\ub2e4. \ud1a0\ud06c\ub098\uc774\uc800\ub294 \ud14d\uc2a4\ud2b8\ub97c \uc778\ub371\uc2a4 \ud615\ud0dc\ub85c \ubcc0\ud658\ud558\ub294 \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4.<\/p>\n<\/section>\n\n6. MLM \ud30c\uc774\ud504\ub77c\uc778 \uad6c\ud604\ud558\uae30<\/h2>\n
\uc774\uc81c MLM\uc744 \uc608\uc2dc\ub85c \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc6b0\uc120 \uc785\ub825 \ubb38\uc7a5\uc744 \uc900\ube44\ud558\uace0, `[MASK]` \ud1a0\ud070\uc744 \ucd94\uac00\ud55c \ub4a4, \ubaa8\ub378 \uc608\uce21\uc744 \ud574\ubcf4\ub3c4\ub85d \ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\n \n# \uc785\ub825 \ubb38\uc7a5\ninput_text = \"\ub098\ub294 [MASK]\ub97c \uc88b\uc544\ud569\ub2c8\ub2e4.\"\n\n# \ud1a0\ud070\ud654\ninput_ids = tokenizer.encode(input_text, return_tensors='pt')\n\n# \uc608\uce21\nwith torch.no_grad():\n outputs = model(input_ids)\n predictions = outputs.logits\n\n# \uc608\uce21\ub41c \ud1a0\ud070\uc758 \uc778\ub371\uc2a4\nmasked_index = torch.where(input_ids == tokenizer.mask_token_id)[1]\npredicted_index = predictions[0, masked_index].argmax(dim=-1)\n\n# \uc608\uce21\ub41c \ub2e8\uc5b4\npredicted_token = tokenizer.decode(predicted_index)\nprint(f\"\uc608\uce21\ub41c \ub2e8\uc5b4: {predicted_token}\")\n <\/code>\n <\/pre>\n\uc704 \ucf54\ub4dc\ub294 \uc785\ub825 \ubb38\uc7a5\uc744 \ud1a0\ud070\ud654\ud558\uace0 \ubaa8\ub378\uc744 \ud1b5\ud574 \uc608\uce21\ud55c \uacb0\uacfc\ub97c \ucd9c\ub825\ud569\ub2c8\ub2e4. \ucd5c\uc885\uc801\uc73c\ub85c \uc608\uce21\ub41c \ub2e8\uc5b4\ub97c \ucd9c\ub825\ud558\uc5ec \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n\n7. \uacb0\uacfc \ubd84\uc11d\ud558\uae30<\/h2>\n
\uc704\uc758 \uc608\uc81c\uc5d0\uc11c \uc0ac\uc6a9\ub41c \ubb38\uc7a5 “\ub098\ub294 [MASK]\ub97c \uc88b\uc544\ud569\ub2c8\ub2e4.”\uc758 \uacbd\uc6b0, \ubaa8\ub378\uc740 \uac00\uc7a5 \uc801\ud569\ud55c \ub2e8\uc5b4\ub97c `{predicted_token}` \ud615\ud0dc\ub85c \ucd9c\ub825\ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4 “\ub098\ub294 \uc0ac\uacfc\ub97c \uc88b\uc544\ud569\ub2c8\ub2e4.”\uc640 \uac19\uc740 \ucd9c\ub825\uc774 \uc608\uc0c1\ub429\ub2c8\ub2e4.<\/p>\n
\uc774\ub7ec\ud55c \uacb0\uacfc\ub97c \ubc14\ud0d5\uc73c\ub85c \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\uac70\ub098, \uc2e4\uc81c \ub370\uc774\ud130\uc5d0\uc11c \uc5b4\ub5bb\uac8c \ud65c\uc6a9\ud560 \uc218 \uc788\uc744\uc9c0 \uace0\ubbfc\ud574\ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n\n8. \uacb0\ub860<\/h2>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc758 DistilBERT \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uc5ec MLM \ud30c\uc774\ud504\ub77c\uc778\uc744 \uad6c\ud604\ud558\ub294 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubc29\ubc95\uc740 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c \ud544\uc694\ud55c \ub2e4\uc591\ud55c \ub370\uc774\ud130 \uc804\ucc98\ub9ac \ubc0f \ubaa8\ub378 \uc801\uc6a9 \uae30\uc220\uc744 \uc2b5\ub4dd\ud558\ub294 \ub370 \ud070 \ub3c4\uc6c0\uc774 \ub420 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
\uc55e\uc73c\ub85c\ub3c4 \ub2e4\uc591\ud55c \ubaa8\ub378\uacfc \ud0dc\uc2a4\ud06c\uc5d0 \ub300\ud55c \ud559\uc2b5\uc744 \uacc4\uc18d \uc774\uc5b4\uac00\uc2dc\uae38 \ubc14\ub78d\ub2c8\ub2e4. \uac10\uc0ac\ud569\ub2c8\ub2e4!<\/p>\n<\/section>\n