\n \ub525\ub7ec\ub2dd\uacfc \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\uc758 \ubc1c\uc804\uc5d0 \ubc1c\ub9de\ucd94\uc5b4, \ud5c8\uae45\ud398\uc774\uc2a4(Hugging Face)\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38(transformers) \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ub9ce\uc740 \ub370\uc774\ud130 \uacfc\ud559\uc790\ub4e4\uacfc \uac1c\ubc1c\uc790\ub4e4\uc5d0\uac8c \ud544\uc218\uc801\uc778 \ub3c4\uad6c\uac00 \ub418\uc5c8\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, BERT(Bidirectional Encoder Representations from Transformers) \ubaa8\ub378\uc740 \ubb38\ub9e5\uc744 \uc774\ud574\ud558\ub294 \ub370 \uac15\ub825\ud55c \uc131\ub2a5\uc744 \ubc1c\ud718\ud558\uc5ec \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uc791\uc5c5\uc5d0\uc11c \ub110\ub9ac \uc0ac\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uc5ec BERT \ubaa8\ub378\uc744 \ud65c\uc6a9\ud55c \ub3c4\ud050\uba3c\ud2b8 \ubca1\ud130 \ucc98\ub9ac \ubc29\ubc95\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ucd5c\uc2e0 \uc790\uc5f0\uc5b4 \ucc98\ub9ac(SQ) \ubaa8\ub378\uc744 \uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ud574\uc8fc\ub294 \ud30c\uc774\uc36c \ud328\ud0a4\uc9c0\uc785\ub2c8\ub2e4. \uc774 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ub300\uaddc\ubaa8 \uc0ac\uc804 \ud6c8\ub828\ub41c \ubaa8\ub378\uc744 \uc81c\uacf5\ud558\uc5ec \uac1c\ubc1c\uc790\ub4e4\uc774 \ubcf5\uc7a1\ud55c \ubaa8\ub378\uc744 \uc190\uc27d\uac8c \uad6c\ud604\ud558\uace0 \ud65c\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \uc9c0\uc6d0\ud569\ub2c8\ub2e4. \ud2b9\ud788 BERT, GPT-2, RoBERTa \ub4f1\uc758 \uc778\uae30 \ubaa8\ub378\uc744 \ud3ec\ud568\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n BERT(Bidirectional Encoder Representations from Transformers)\ub294 \ubb38\ub9e5\uc744 \uc591\ubc29\ud5a5\uc5d0\uc11c \uc774\ud574\ud558\uae30 \uc704\ud574 \uc124\uacc4\ub41c \ubaa8\ub378\uc785\ub2c8\ub2e4. \uae30\uc874\uc758 RNN \uacc4\uc5f4 \ubaa8\ub378\ub4e4\uc740 \ub2e8\ud5a5\uc131\uc73c\ub85c \uc815\ubcf4\uc758 \ud750\ub984\uc744 \uc774\ud574\ud558\uae30 \ub54c\ubb38\uc5d0 \ubb38\ub9e5\uc744 \uc644\ubcbd\ud788 \ucc98\ub9ac\ud558\uae30 \uc5b4\ub824\uc6e0\uc73c\ub098, BERT\ub294 \uc591\ubc29\ud5a5\uc758 \uc815\ubcf4\ub97c \ubaa8\ub450 \ud65c\uc6a9\ud560 \uc218 \uc788\uac8c \ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \ub192\uc740 \uc218\uc900\uc758 \ubb38\uc7a5 \uc774\ud574 \ubc0f \ubb38\uc11c \ubd84\ub958 \uc791\uc5c5\uc5d0\uc11c \uc131\ub2a5\uc744 \uadf9\ub300\ud654 \ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \uc2dc\uc791\ud558\uae30 \uc804\uc5d0 \ud544\uc694\ud55c \uac1c\ubc1c \ud658\uacbd\uc744 \uc900\ube44\ud574\uc57c \ud569\ub2c8\ub2e4. \uc544\ub798\uc758 \ub2e8\uacc4\ub97c \ub530\ub77c \ud544\uc694\ud55c \ud328\ud0a4\uc9c0\ub97c \uc124\uce58\ud558\uace0 \ud658\uacbd\uc744 \uc124\uc815\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \uba3c\uc800 Python\uc774 \uc124\uce58\ub418\uc5b4 \uc788\uc5b4\uc57c \ud569\ub2c8\ub2e4. [Python \uacf5\uc2dd \uc6f9\uc0ac\uc774\ud2b8](https:\/\/www.python.org\/downloads\/)\uc5d0\uc11c Python\uc744 \ub2e4\uc6b4\ub85c\ub4dc\ud558\uace0 \uc124\uce58\ud569\ub2c8\ub2e4. Python 3.6 \uc774\uc0c1\uc774 \ud544\uc694\ud569\ub2c8\ub2e4.\n <\/p>\n
\n \uac00\uc0c1 \ud658\uacbd\uc744 \uc0dd\uc131\ud558\ub294 \uac83\uc740 \ud504\ub85c\uc81d\ud2b8\uc758 \uc885\uc18d\uc131\uc744 \uad00\ub9ac\ud558\ub294 \ub370 \uc720\ub9ac\ud569\ub2c8\ub2e4. \uc544\ub798 \uba85\ub839\uc5b4\ub97c \ud1b5\ud574 \uc0c8\ub85c\uc6b4 \uac00\uc0c1 \ud658\uacbd\uc744 \uc0dd\uc131\ud558\uace0 \ud65c\uc131\ud654\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
python -m venv bert-env\nsource bert-env\/bin\/activate # Linux \/ macOS\n.\\bert-env\\Scripts\\activate # Windows<\/code><\/pre>\n3.3. \ud328\ud0a4\uc9c0 \uc124\uce58<\/h3>\n
\n \uc774\uc81c BERT \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uae30 \uc704\ud55c \ud328\ud0a4\uc9c0\ub97c \uc124\uce58\ud569\ub2c8\ub2e4. \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\uc640 \ucd94\uac00\uc801\uc73c\ub85c \ud544\uc694\ud55c \ud328\ud0a4\uc9c0\ub97c \uc124\uce58\ud569\ub2c8\ub2e4.\n <\/p>\n
pip install transformers torch<\/code><\/pre>\n4. BERT \ubaa8\ub378 \uc0ac\uc6a9\ud558\uae30<\/h2>\n
\n \uc124\uce58\uac00 \uc644\ub8cc\ub418\uba74 BERT \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ubb38\uc11c \ubca1\ud130\ub97c \ucc98\ub9ac\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\uc758 \uc608\uc81c \ucf54\ub4dc\ub97c \ud1b5\ud574 \uac04\ub2e8\ud55c \ubb38\uc11c \ubca1\ud130\ucc98\ub9ac \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
4.1. \ucf54\ub4dc \uc608\uc81c<\/h3>\nimport torch\nfrom transformers import BertTokenizer, BertModel\n\n# BERT \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800 \ucd08\uae30\ud654\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\nmodel = BertModel.from_pretrained('bert-base-uncased')\n\n# \uc608\uc2dc \ubb38\uc11c\ndocument = \"\ud5c8\uae45\ud398\uc774\uc2a4 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub97c \uc704\ud55c \ub9e4\uc6b0 \uc911\uc694\ud55c \ub3c4\uad6c\uc785\ub2c8\ub2e4.\"\n\n# \ubb38\uc11c \ud1a0\ud06c\ub098\uc774\uc988 \ubc0f \ud150\uc11c \ubcc0\ud658\ninputs = tokenizer(document, return_tensors='pt')\n\n# BERT \ubaa8\ub378\uc5d0 \uc785\ub825\ud558\uace0 \ucd9c\ub825 \uac12\uc744 \uc5bb\uc74c\nwith torch.no_grad():\n outputs = model(**inputs)\n\n# \ub9c8\uc9c0\ub9c9 \uc740\ub2c9 \uc0c1\ud0dc \uc5bb\uae30\nlast_hidden_states = outputs.last_hidden_state\nprint(last_hidden_states.shape) # (1, \ubb38\uc7a5 \uae38\uc774, 768)<\/code><\/pre>\n4.2. \ucf54\ub4dc \uc124\uba85<\/h3>\n
\n \uc704 \ucf54\ub4dc\ub294 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 BERT \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ub2e8\uc77c \ubb38\uc11c\ub97c \ucc98\ub9ac\ud558\ub294 \uacfc\uc815\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. <\/p>\n
\nfrom transformers import BertTokenizer, BertModel<\/code> : \ud5c8\uae45\ud398\uc774\uc2a4\uc5d0\uc11c BERT \ubaa8\ub378\uacfc \uad00\ub828\ub41c \ud074\ub798\uc2a4\ub4e4\uc744 \ubd88\ub7ec\uc635\ub2c8\ub2e4.<\/li>\nBertTokenizer.from_pretrained('bert-base-uncased')<\/code> : \uc0ac\uc804\ud6c8\ub828\ub41c BERT\uc758 \ud1a0\ud06c\ub098\uc774\uc800\ub97c \ubd88\ub7ec\uc635\ub2c8\ub2e4. ‘bert-base-uncased’\ub294 \ub300\uc18c\ubb38\uc790\ub97c \uad6c\ubd84\ud558\uc9c0 \uc54a\ub294 \ubaa8\ub378\uc785\ub2c8\ub2e4.<\/li>\nmodel = BertModel.from_pretrained('bert-base-uncased')<\/code> : BERT \ubaa8\ub378\uc744 \ucd08\uae30\ud654\ud569\ub2c8\ub2e4.<\/li>\ntokenizer(document, return_tensors='pt')<\/code> : \uc785\ub825 \ubb38\uc11c\ub97c \ud1a0\ud06c\ub098\uc774\uc988\ud558\uace0 PyTorch \ud150\uc11c \ud615\uc2dd\uc73c\ub85c \ubcc0\ud658\ud569\ub2c8\ub2e4.<\/li>\nmodel(**inputs)<\/code> : \ubaa8\ub378\uc5d0 \uc785\ub825\uc744 \uc8fc\uace0 \ucd9c\ub825\uac12\uc744 \ubc1b\uc2b5\ub2c8\ub2e4. \ucd9c\ub825\uc758 \ub9c8\uc9c0\ub9c9 \uc740\ub2c9 \uc0c1\ud0dc\ub97c \ud1b5\ud574 \ubb38\uc11c\uc758 \uc758\ubbf8\uc801 \uc815\ubcf4\ub97c \uc5bb\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<\/ul>\n5. \uc2e4\uc81c \ubb38\uc11c \ubca1\ud130 \ucc98\ub9ac \uc608\uc81c<\/h2>\n
\n \ub2e4\uc218\uc758 \ubb38\uc11c\ub97c \ucc98\ub9ac\ud558\uc5ec \uac01 \ubb38\uc11c\uc758 \ubca1\ud130\ub97c \uc5bb\ub294 \ubc29\ubc95\ub3c4 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc544\ub798\uc758 \ucf54\ub4dc\ub294 \uc5ec\ub7ec \ubb38\uc11c\ub97c \ubca1\ud130\ud654\ud558\ub294 \uacfc\uc815\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.\n <\/p>\n
documents = [\n \"\ud5c8\uae45\ud398\uc774\uc2a4\ub294 NLP\ub97c \uc704\ud55c \uac15\ub825\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4.\",\n \"\ub525\ub7ec\ub2dd\uc740 \ub9ce\uc740 \uc0b0\uc5c5\uc5d0\uc11c \uc911\uc694\ud55c \uc5ed\ud560\uc744 \ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.\",\n \"AI\uc758 \ubc1c\uc804\uc740 \uae30\uc220\uc758 \uc9c4\ud654\ub97c \uc774\ub04c\uace0 \uc788\uc2b5\ub2c8\ub2e4.\"\n]\n\n# \uac01 \ubb38\uc11c\ub97c \ubc18\ubcf5\ud558\uc5ec \ubca1\ud130\ud654\ndoc_vectors = []\nfor doc in documents:\n inputs = tokenizer(doc, return_tensors='pt')\n with torch.no_grad():\n outputs = model(**inputs)\n doc_vectors.append(outputs.last_hidden_state.mean(dim=1).squeeze().numpy())\n\nprint(f\"\ubb38\uc11c \ubca1\ud130\ub4e4: {doc_vectors}\") # \ub9ac\uc2a4\ud2b8 \ud615\ud0dc\ub85c \uac01 \ubb38\uc11c\uc758 \ubca1\ud130 \ucd9c\ub825<\/code><\/pre>\n5.1. \ucf54\ub4dc \uc124\uba85<\/h3>\n
\n \uc704 \ucf54\ub4dc\uc5d0\uc11c\ub294 \uc5ec\ub7ec \uac1c\uc758 \ubb38\uc11c\ub97c \uc785\ub825\ubc1b\uc544 \uac01 \ubb38\uc11c\uc758 \ubca1\ud130\ub97c \uacc4\uc0b0\ud569\ub2c8\ub2e4. <\/p>\n
\n- \ubb38\uc11c\ub97c \ud558\ub098\uc529 \ubc18\ubcf5\ud558\uc5ec \uac01 \ubb38\uc11c\uc5d0 \ub300\ud574 \ud1a0\ud06c\ub098\uc774\uc988 \ubc0f \ubca1\ud130\ub97c \uacc4\uc0b0\ud569\ub2c8\ub2e4.<\/li>\n
outputs.last_hidden_state.mean(dim=1)<\/code> : \uac01 \ubb38\uc11c\uc758 \ubaa8\ub4e0 \ud1a0\ud070\uc5d0 \ub300\ud55c \uc740\ub2c9 \uc0c1\ud0dc\ub97c \ud3c9\uade0\ud558\uc5ec \ubb38\uc11c\uc758 \ub300\ud45c \ubca1\ud130\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4.<\/li>\n- \ubb38\uc11c \ubca1\ud130\ub4e4\uc740 \ub9ac\uc2a4\ud2b8 \ud615\ud0dc\ub85c \uc800\uc7a5\ub418\uba70, \ucd5c\uc885\uc801\uc73c\ub85c
doc_vectors<\/code>\uc5d0 \uc800\uc7a5\ub429\ub2c8\ub2e4.<\/li>\n<\/ul>\n6. \uace0\uae09 \ud65c\uc6a9: \ubb38\uc11c \uc720\uc0ac\ub3c4 \uce21\uc815<\/h2>\n
\n \ubca1\ud130\ud654\ub97c \ud1b5\ud574 \uac01 \ubb38\uc11c\uc758 \uc758\ubbf8\ub97c \uc218\uce58\ud654\ud560 \uc218 \uc788\uc73c\uba70, \uc774\ub97c \ubc14\ud0d5\uc73c\ub85c \ubb38\uc11c \uac04\uc758 \uc720\uc0ac\ub3c4\ub97c \uce21\uc815\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc720\uc0ac\ub3c4 \uce21\uc815 \ubc29\ubc95 \uc911 \ud558\ub098\ub85c\ub294 \ucf54\uc0ac\uc778 \uc720\uc0ac\ub3c4\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
from sklearn.metrics.pairwise import cosine_similarity\n\n# \ubb38\uc11c \ubca1\ud130 \uac04\uc758 \ucf54\uc0ac\uc778 \uc720\uc0ac\ub3c4 \uacc4\uc0b0\nsimilarity_matrix = cosine_similarity(doc_vectors)\nprint(f\"\ubb38\uc11c \uc720\uc0ac\ub3c4 \ud589\ub82c: \\n{similarity_matrix}\")<\/code><\/pre>\n6.1. \ucf54\ub4dc \uc124\uba85<\/h3>\n
\n \ucf54\uc0ac\uc778 \uc720\uc0ac\ub3c4\ub97c \ud1b5\ud574 \uac01 \ubb38\uc11c \uac04\uc758 \uc720\uc0ac\ub3c4\ub97c \uacc4\uc0b0\ud558\uc5ec \uc720\uc0ac\ub3c4 \ud589\ub82c\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4. <\/p>\n
\nfrom sklearn.metrics.pairwise import cosine_similarity<\/code> : \ucf54\uc0ac\uc778 \uc720\uc0ac\ub3c4 \uacc4\uc0b0\uc744 \uc704\ud55c \ud568\uc218 \ubd88\ub7ec\uc624\uae30<\/li>\ncosine_similarity(doc_vectors)<\/code> : \ubca1\ud130 \ub9ac\uc2a4\ud2b8\ub97c \ubc14\ud0d5\uc73c\ub85c \uc720\uc0ac\ub3c4 \ud589\ub82c\uc744 \uacc4\uc0b0\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n7. \ub9c8\uce58\uba70<\/h2>\n
\n \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uace0 BERT \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uc5ec \ubb38\uc11c \ubca1\ud130 \ucc98\ub9ac \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. BERT\uc640 \uac19\uc740 \uace0\uae09 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubaa8\ub378\uc740 \ub2e4\uc591\ud55c NLP \uc791\uc5c5\uc5d0\uc11c \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubc1c\ud718\ud558\ubbc0\ub85c \uc774\ub97c \ud65c\uc6a9\ud558\uc5ec \ub2e4\uc591\ud55c \ub370\uc774\ud130 \ubd84\uc11d \uc791\uc5c5\uc5d0 \uc801\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c\ub3c4 \ub354 \ub9ce\uc740 \uc608\uc81c\uc640 \uc2ec\ud654 \ub0b4\uc6a9\uc744 \ub2e4\ub8f0 \uc608\uc815\uc774\ub2c8 \ub9ce\uc740 \uad00\uc2ec \ubd80\ud0c1\ub4dc\ub9bd\ub2c8\ub2e4.\n <\/p>\n
\ub525\ub7ec\ub2dd\uacfc \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc758 \uc138\uacc4\uc5d0 \ubc1c \ube60\ub974\uac8c \uc801\uc751\ud558\uc5ec \uc2e4\ub825\uc744 \uc313\ub294 \uc5ec\ub7ec\ubd84\uc774 \ub418\uae38 \ubc14\ub78d\ub2c8\ub2e4!<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"