\ucd5c\uadfc \uba87 \ub144 \uac04 \ubc14\uc774\uc624 \ubd84\uc57c\uc5d0\uc11c\ub294 \ub525\ub7ec\ub2dd \uae30\uc220\uc758 \ubc1c\uc804\uc73c\ub85c \uc778\ud574 \ub9ce\uc740 \ud601\uc2e0\uc774 \uc77c\uc5b4\ub0ac\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \ud5c8\uae45\ud398\uc774\uc2a4(Hugging Face) \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\uc5d0\uc11c \ud070 \uc8fc\ubaa9\uc744 \ubc1b\uc73c\uba70, \ub2e4\uc591\ud55c \ubaa8\ub378\uacfc \ub3c4\uad6c\ub4e4\uc744 \uc81c\uacf5\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub97c \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub354\ub098(Moderna)\uc640 \ud654\uc774\uc790(Pfizer)\uc758 \ubc31\uc2e0 \uad00\ub828 \ub370\uc774\ud130\ub97c \uc2dc\uac01\ud654\ud558\ub294 t-SNE \ubc29\ubc95\uc5d0 \ub300\ud574 \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\ud2b8\ub79c\uc2a4\ud3ec\uba38 \ubaa8\ub378\uc740 2017\ub144 ‘Attention is All You Need’\ub77c\ub294 \ub17c\ubb38\uc5d0\uc11c \ucc98\uc74c \uc18c\uac1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uae30\uc874\uc758 RNN \ubc0f LSTM \ubaa8\ub378\uacfc\ub294 \ub2ec\ub9ac, \ud2b8\ub79c\uc2a4\ud3ec\uba38\ub294 \uc785\ub825 \ub370\uc774\ud130\uc758 \ubaa8\ub4e0 \uc704\uce58\ub97c \ub3d9\uc2dc\uc5d0 \ucc98\ub9ac\ud560 \uc218 \uc788\ub294 \uc790\uae30 \uc8fc\uc758(attention) \uba54\ucee4\ub2c8\uc998\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ud2b9\uc9d5 \ub355\ubd84\uc5d0 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ubaa8\ub378\uc740 \ub300\uaddc\ubaa8 \ub370\uc774\ud130\uc14b\uc5d0\uc11c \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubc1c\ud718\ud569\ub2c8\ub2e4.<\/p>\n
\ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud558\uae30 \uc704\ud574\uc11c\ub294 \uba3c\uc800 \ud574\ub2f9 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. \ub2e4\uc74c\uacfc \uac19\uc740 \uba85\ub839\uc5b4\ub85c \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
pip install transformers datasets<\/code><\/pre>\n3. \ub370\uc774\ud130 \uc900\ube44<\/h2>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ubaa8\ub354\ub098\uc640 \ud654\uc774\uc790\uc758 \ubc31\uc2e0\uc5d0 \uad00\ud55c \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud558\uc5ec t-SNE \uc2dc\uac01\ud654\ub97c \uc9c4\ud589\ud569\ub2c8\ub2e4. \ub370\uc774\ud130\ub97c \uc218\uc9d1\ud558\uace0 \uc804\ucc98\ub9ac\ud558\ub294 \ub2e8\uacc4\ub294 \ub2e4\uc74c\uacfc \uac19\uc774 \uc27d\uac8c \uc124\uba85\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \uac01 \ubc31\uc2e0\uc5d0 \ub300\ud55c \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130 \uc218\uc9d1<\/li>\n
- \uc218\uc9d1\ub41c \ub370\uc774\ud130\uc5d0\uc11c \ud14d\uc2a4\ud2b8 \uc804\ucc98\ub9ac (\uc18c\ubb38\uc790 \ubcc0\ud658, \uad6c\ub450\uc810 \uc81c\uac70 \ub4f1)<\/li>\n
- \uc804\ucc98\ub9ac\ub41c \ud14d\uc2a4\ud2b8\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc784\ubca0\ub529 \uc0dd\uc131<\/li>\n<\/ul>\n
3.1 \ub370\uc774\ud130 \uc218\uc9d1 \uc608\uc81c<\/h3>\n
\ubaa8\ub354\ub098\uc640 \ud654\uc774\uc790\uc5d0 \ub300\ud55c \uae30\uc0ac\ub97c \ud06c\ub864\ub9c1\ud558\uac70\ub098, \ubbf8\ub9ac \uc900\ube44\ud55c CSV \ud30c\uc77c\uc744 \uc0ac\uc6a9\ud558\uc5ec \ub370\uc774\ud130\uc14b\uc744 \ub9cc\ub4e4 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\ub294 \ubaa8\ub354\ub098\uc640 \ud654\uc774\uc790\uc758 \ub370\uc774\ud130\uc14b\ub97c \ub85c\ub4dc\ud558\ub294 \uc608\uc2dc\uc785\ub2c8\ub2e4.<\/p>\n
import pandas as pd\n\nmoderna_df = pd.read_csv('moderna.csv')\npfizer_df = pd.read_csv('pfizer.csv')\n\n# \ub370\uc774\ud130 \ud655\uc778\nprint(moderna_df.head())\nprint(pfizer_df.head())<\/code><\/pre>\n4. \ud14d\uc2a4\ud2b8 \uc784\ubca0\ub529 \uc0dd\uc131<\/h2>\n
\ubaa8\ub378\uc774 \ud14d\uc2a4\ud2b8\ub97c \uc774\ud574\ud560 \uc218 \uc788\ub3c4\ub85d \uc784\ubca0\ub529 \ubca1\ud130\ub97c \uc0dd\uc131\ud574\uc57c \ud569\ub2c8\ub2e4. \ud5c8\uae45\ud398\uc774\uc2a4\uc758 ‘BERT’\ub098 ‘DistilBERT’ \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uc5ec \ud14d\uc2a4\ud2b8\ub97c \uc784\ubca0\ub529\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798 \ucf54\ub4dc\ub97c \ucc38\uace0\ud558\uc5ec \uc784\ubca0\ub529\uc744 \uc0dd\uc131\ud574 \ubcf4\uc138\uc694.<\/p>\n
from transformers import DistilBertTokenizer, DistilBertModel\nimport torch\n\n# \ubaa8\ub354\ub098\uc640 \ud654\uc774\uc790 \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130 \ud569\uce58\uae30\ntexts = list(moderna_df['text']) + list(pfizer_df['text'])\n\n# \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800 \ucd08\uae30\ud654\ntokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')\nmodel = DistilBertModel.from_pretrained('distilbert-base-uncased')\n\n# \uc785\ub825 \ub370\uc774\ud130 \ud1a0\ud070\ud654\ninputs = tokenizer(texts, padding=True, truncation=True, return_tensors='pt')\nwith torch.no_grad():\n outputs = model(**inputs)\n\nembeddings = outputs.last_hidden_state.mean(dim=1) # \ud3c9\uade0\uc73c\ub85c \uc784\ubca0\ub529 \uc0dd\uc131<\/code><\/pre>\n5. t-SNE \uc2dc\uac01\ud654<\/h2>\n
t-SNE(\ub450 \uac1c\uc758 \uace0\ucc28\uc6d0 \ub370\uc774\ud130 \ud3ec\uc778\ud2b8 \uac04\uc758 \ucc28\uc774\ub97c \ubcf4\uc874\ud558\ub294 \uae30\uc220)\ub294 \uace0\ucc28\uc6d0 \ub370\uc774\ud130\ub97c 2\ucc28\uc6d0 \ub610\ub294 3\ucc28\uc6d0\uc73c\ub85c \uc2dc\uac01\ud654\ud558\ub294 \ub370\uc5d0 \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4. \uc544\ub798 \ucf54\ub4dc\ub294 t-SNE\ub97c \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub354\ub098\uc640 \ud654\uc774\uc790\uc758 \ub370\uc774\ud130 \ubd84\ud3ec\ub97c \uc2dc\uac01\ud654\ud558\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n
from sklearn.manifold import TSNE\nimport matplotlib.pyplot as plt\n\n# t-SNE\ub97c \ud1b5\ud574 2\ucc28\uc6d0\uc73c\ub85c \ucc28\uc6d0 \ucd95\uc18c\ntsne = TSNE(n_components=2, random_state=0)\ntsne_results = tsne.fit_transform(embeddings)\n\n# \uacb0\uacfc \uc2dc\uac01\ud654\nplt.figure(figsize=(10, 7))\nplt.scatter(tsne_results[:len(moderna_df), 0], tsne_results[:len(moderna_df), 1], label='Moderna', alpha=0.5)\nplt.scatter(tsne_results[len(moderna_df):, 0], tsne_results[len(moderna_df):, 1], label='Pfizer', alpha=0.5)\nplt.title('t-SNE Visualization of Moderna vs Pfizer')\nplt.xlabel('t-SNE Component 1')\nplt.ylabel('t-SNE Component 2')\nplt.legend()\nplt.show()<\/code><\/pre>\n6. \uacb0\uacfc \ubd84\uc11d<\/h2>\n
t-SNE\uc758 \uacb0\uacfc\ub97c \ubd84\uc11d\ud558\uc5ec \ubaa8\ub354\ub098\uc640 \ud654\uc774\uc790 \uac04\uc758 \uad00\uacc4\uc131\uc744 \uc774\ud574\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub370\uc774\ud130 \ud3ec\uc778\ud2b8\uac04\uc758 \ubd84\ud3ec\uac00 \uc5b4\ub5bb\uac8c \ud615\uc131\ub418\uc5c8\ub294\uc9c0 \uc2dc\uac01\uc801\uc73c\ub85c \ud655\uc778\ud558\uba74, \uac01 \ubc31\uc2e0\uc774 \uac00\uc9c0\ub294 \ud2b9\uc131 \ubc0f \ucc28\ubcc4\uc810\uc5d0 \ub300\ud574\uc11c\ub3c4 \uc54c \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \ubd84\uc11d\uc744 \ud1b5\ud574 \uacfc\ud559\uc801 \uc5f0\uad6c\uc640 \ub9c8\ucf00\ud305 \uc804\ub7b5 \uc218\ub9bd\uc5d0 \uae30\uc5ec\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\uacb0\ub860<\/h2>\n
\ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uba74 \ubcf5\uc7a1\ud55c \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc784\ubca0\ub529\uc744 \uc27d\uac8c \uc0dd\uc131\ud558\uace0, \uc774\ub97c \ud1b5\ud574 \ub2e4\uc591\ud55c \uc2dc\uac01\ud654 \uae30\ubc95\uc73c\ub85c \ub370\uc774\ud130\ub97c \ubd84\uc11d\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\ub97c \ud1b5\ud574 \ubc30\uc6b4 \ub0b4\uc6a9\uc740 \ubc14\uc774\uc624 \ub370\uc774\ud130 \ubd84\uc11d, \ud2b9\ud788 \ubc31\uc2e0 \ub370\uc774\ud130\uc640 \uac19\uc740 \ubbfc\uac10\ud55c \uc8fc\uc81c\ub97c \ub2e4\ub8e8\ub294 \ub370\uc5d0 \ud070 \ub3c4\uc6c0\uc744 \uc904 \uac83\uc785\ub2c8\ub2e4. \ud5a5\ud6c4\uc5d0\ub294 \ub2e4\ub978 \ubaa8\ub378\uacfc \uae30\ubc95\uc744 \ud65c\uc6a9\ud558\uc5ec \ub354\uc6b1 \uae4a\uc774 \uc788\ub294 \ubd84\uc11d\uc744 \uc9c4\ud589\ud560 \uc218 \uc788\uc744 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0 \uc790\ub8cc<\/h2>\n