\ucd5c\uadfc \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP) \ubd84\uc57c\uc5d0\uc11c\uc758 \ud601\uc2e0\uc740 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \ubc1c\uc804 \ub355\ubd84\uc5d0 \uc774\ub8e8\uc5b4\uc84c\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \ud5c8\uae45\ud398\uc774\uc2a4(Hugging Face)\uc5d0\uc11c \uac1c\ubc1c\ud55c \ud2b8\ub80c\uc2a4\ud3ec\uba38(Transformers)<\/strong> \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \uc774\ub7ec\ud55c \ubc1c\uc804\uc758 \uc0c1\uc9d5\uc73c\ub85c \uc790\ub9ac\uc7a1\uc558\uc2b5\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 PEGASUS \ubaa8\ub378\uc744 \uc774\uc6a9\ud558\uc5ec \ud14d\uc2a4\ud2b8 \uc694\uc57d \ud0dc\uc2a4\ud06c\ub97c \uc218\ud589\ud558\uae30 \uc704\ud574 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uc815 \ubc0f \uc0ac\uc804\ud559\uc2b5 \ubaa8\ub378 \ubd88\ub7ec\uc624\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \ub2e4\ub8e8\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ub2e4\uc591\ud55c NLP \ud0dc\uc2a4\ud06c\uc5d0 \uc0ac\uc6a9\ud560 \uc218 \uc788\ub294 \uc0ac\uc804\ud559\uc2b5\ub41c \ubaa8\ub378\uc744 \uc81c\uacf5\ud558\ub294 \ud30c\uc774\uc36c \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. \uc774 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 \ud2b9\ud788 \ud2b8\ub79c\uc2a4\ud3ec\uba38 \uc544\ud0a4\ud14d\ucc98\ub97c \uae30\ubc18\uc73c\ub85c \ub2e4\uc591\ud55c \ubaa8\ub378(BERT, GPT-2, RoBERTa, T5 \ub4f1)\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. PEGASUS\ub294 \uc774\ub7ec\ud55c \ud2b8\ub79c\uc2a4\ud3ec\uba38 \uc544\ud0a4\ud14d\ucc98\ub97c \uae30\ubc18\uc73c\ub85c \ud558\ub294 \ubaa8\ub378 \uc911 \ud558\ub098\ub85c, \uc8fc\ub85c \ud14d\uc2a4\ud2b8 \uc694\uc57d\uc744 \uc704\ud55c \uc0ac\uc804\ud559\uc2b5 \ubaa8\ub378\uc785\ub2c8\ub2e4.<\/p>\n PEGASUS(Pre-trained Text-to-Text Transfer Transformer)\ub294 \uad6c\uae00\uc5d0\uc11c \uac1c\ubc1c\ud55c \ubaa8\ub378\ub85c, \uc790\uc5f0\uc5b4 \ubb38\uc11c\uc5d0\uc11c \uc911\uc694\ud55c \uc815\ubcf4\ub97c \ucd94\ucd9c\ud558\uc5ec \uc694\uc57d\ud558\ub294 \ub370 \ucd5c\uc801\ud654\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4. PEGASUS \ubaa8\ub378\uc758 \ud575\uc2ec \uc544\uc774\ub514\uc5b4\ub294 \uc785\ub825 \ubb38\uc11c\uc5d0\uc11c \ubb38\uc7a5\ub4e4\uc744 \uc784\uc758\ub85c \uc120\ud0dd\ud558\uc5ec \ub9c8\uc2a4\ud0b9\ud558\uace0, \ub9c8\uc2a4\ud0b9\ub41c \ubb38\uc7a5\uc744 \uc608\uce21\ud558\ub294 \ubc29\uc2dd\uc73c\ub85c \uc0ac\uc804\ud559\uc2b5\uc744 \uc218\ud589\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc5d0\uc11c \ubaa8\ub378\uc740 \uae00\uc758 \uc804\ubc18\uc801\uc778 \ub9e5\ub77d\uc744 \uc774\ud574\ud558\uace0 \uc911\uc694\ud55c \uc815\ubcf4\ub97c \uc2dd\ubcc4\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n PEGASUS \ubaa8\ub378\uc744 \uc774\uc6a9\ud558\uae30 \uc704\ud574\uc11c\ub294 \uba3c\uc800 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc5d0\uc11c\ub294 \uc8fc\ub85c \uc124\uce58\uac00 \uc644\ub8cc\ub418\uc5c8\ub2e4\uba74, \ub2e4\uc74c\uacfc \uac19\uc774 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc784\ud3ec\ud2b8\ud569\ub2c8\ub2e4.<\/p>\n \uc774\uc81c PEGASUS \ubaa8\ub378\uacfc \ud1a0\ud06c\ub098\uc774\uc800\ub97c \ubd88\ub7ec\uc62c \ucc28\ub840\uc785\ub2c8\ub2e4. PEGASUS \ubaa8\ub378\uc740 \ud5c8\uae45\ud398\uc774\uc2a4\uc758 \ubaa8\ub378 \ud5c8\ube0c\uc5d0\uc11c \uc27d\uac8c \ubd88\ub7ec\uc62c \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uae30 \uc704\ud574\uc11c\ub294 \uc694\uc57d\ud560 \ud14d\uc2a4\ud2b8\ub97c \uc900\ube44\ud574\uc57c \ud569\ub2c8\ub2e4. \uc544\ub798 \ucf54\ub4dc\uc5d0\uc11c\ub294 \uc0d8\ud50c \ud14d\uc2a4\ud2b8\ub97c \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\n \ud1a0\ud06c\ub098\uc774\uc800\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud14d\uc2a4\ud2b8\ub97c \ud1a0\ud070\ud654 \ud55c \ud6c4, \ubc30\uce58 \ud06c\uae30\ub97c \uc124\uc815\ud558\uace0 \uc694\uc57d\uc744 \uc218\ud589\ud569\ub2c8\ub2e4.<\/p>\n \uc704 \ucf54\ub4dc\uc5d0\uc11c \uc0dd\uc131\ud55c \uc694\uc57d \uacb0\uacfc\ub97c \ud655\uc778\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc0dd\uc131\ub41c \uc694\uc57d \ubb38\uc7a5\uc740 \ubcf8\ub798 \ubb38\uc11c\uc5d0\uc11c \uac00\uc7a5 \uc911\uc694\ud55c \uc815\ubcf4\ub97c \uac04\uacb0\ud558\uac8c \uc804\ub2ec\ud574\uc57c \ud569\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc5d0\uc11c \ubc1c\uc0dd\ud560 \uc218 \uc788\ub294 \ubb38\uc81c\ub85c\ub294 \uc798\ubabb\ub41c \ubb38\uc7a5 \ud615\uc131\uc774\ub098 \ub204\ub77d\ub41c \uc815\ubcf4\uac00 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574\uc11c\ub294 \ud559\uc2b5 \ub370\uc774\ud130\uc758 \uc591\uacfc \uc9c8\uc774 \ub9e4\uc6b0 \uc911\uc694\ud569\ub2c8\ub2e4.<\/p>\n \ubcf8 \uac15\uc88c\ub97c \ud1b5\ud574 PEGASUS \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uc5ec \ud14d\uc2a4\ud2b8 \uc694\uc57d\uc744 \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. PEGASUS \ubaa8\ub378\uc740 \uc0ac\uc804\ud559\uc2b5\ub41c \uac15\ub825\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubaa8\ub378\ub85c, \ud6a8\uacfc\uc801\uc778 \uc694\uc57d \ub3c4\uad6c\ub85c \uc790\ub9ac\uc7a1\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ub354\ubd88\uc5b4 \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud1b5\ud574 \uc27d\uac8c \ubaa8\ub378\uc744 \ubd88\ub7ec\uc640 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc74c\uc744 \ud655\uc778\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n \ud5a5\ud6c4 \uc774 \uac15\uc88c\uc5d0\uc11c\ub294 PEGASUS \ubaa8\ub378\uc744 Fine-tuning\ud558\uc5ec \ud2b9\uc815 \ub3c4\uba54\uc778\uc5d0 \ub9de\uac8c \uc870\uc815\ud558\ub294 \ubc29\ubc95\uacfc, \ub2e4\uc591\ud55c \ud558\uc774\ud37c\ud30c\ub77c\ubbf8\ud130\ub97c \uc870\uc815\ud558\uc5ec \uc131\ub2a5\uc744 \uac1c\uc120\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \ub2e4\ub8e8\uc5b4 \ubcf4\ub3c4\ub85d \ud558\uaca0\uc2b5\ub2c8\ub2e4. NLP\uc758 \uc138\uacc4\ub294 \ub9e4\uc6b0 \ub113\uace0, \ub2e4\uc591\ud55c \uc751\uc6a9 \uac00\ub2a5\uc131\uc744 \uc9c0\ub2c8\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc9c0\uc18d\uc801\uc73c\ub85c \ud559\uc2b5\ud558\uace0 \uc5f0\uad6c\ud558\uc5ec \ub354\uc6b1 \ud6a8\uacfc\uc801\uc778 \ubaa8\ub378\uc744 \uac1c\ubc1c\ud574 \ub098\uac00\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n1. \ud5c8\uae45\ud398\uc774\uc2a4 \ud2b8\ub80c\uc2a4\ud3ec\uba38 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub780?<\/h2>\n
2. PEGASUS \ubaa8\ub378 \uc18c\uac1c<\/h2>\n
2.1. PEGASUS \ubaa8\ub378\uc758 \uc7a5\uc810<\/h3>\n
\n
3. \ud658\uacbd \uc124\uc815<\/h2>\n
transformers<\/code>, torch<\/code> \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uce58\uac00 \ud544\uc694\ud569\ub2c8\ub2e4. \uc544\ub798\ub294 \uc124\uce58 \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/p>\npip install transformers torch<\/code><\/pre>\n3.1. \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac \uc784\ud3ec\ud2b8<\/h3>\n
import torch\nfrom transformers import PegasusForConditionalGeneration, PegasusTokenizer<\/code><\/pre>\n4. PEGASUS \ubaa8\ub378 \ubd88\ub7ec\uc624\uae30<\/h2>\n
model_name = \"google\/pegasus-xsum\"\n\n# \ud1a0\ud06c\ub098\uc774\uc800 \ubc0f \ubaa8\ub378 \ubd88\ub7ec\uc624\uae30\ntokenizer = PegasusTokenizer.from_pretrained(model_name)\nmodel = PegasusForConditionalGeneration.from_pretrained(model_name)<\/code><\/pre>\n4.1 \uc694\uc57d\ud560 \ud14d\uc2a4\ud2b8 \uc900\ube44<\/h3>\n
sample_text = \"\uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub294 \ucef4\ud4e8\ud130\uc640 \uc778\uac04\uc758 \uc5b8\uc5b4 \uc0ac\uc774\uc758 \uc0c1\ud638 \uc791\uc6a9\uc744 \ub2e4\ub8e8\ub294 \ucef4\ud4e8\ud130 \uacfc\ud559\uc758 \ud55c \ubd84\uc57c\uc785\ub2c8\ub2e4. \uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub294 \uc778\uac04\uc758 \uc5b8\uc5b4\ub97c \uc774\ud574\ud558\uace0 \ucc98\ub9ac\ud558\ub294 \ubc29\ubc95\uc744 \uc5f0\uad6c\ud569\ub2c8\ub2e4.\"<\/code><\/pre>\n4.2 \ud14d\uc2a4\ud2b8 \ud1a0\ud070\ud654 \ubc0f \uc694\uc57d<\/h3>\n
# \ud14d\uc2a4\ud2b8 \uc778\ucf54\ub529\ninputs = tokenizer(sample_text, return_tensors=\"pt\", max_length=512, truncation=True)\n\n# \uc694\uc57d \uc0dd\uc131\nsummary_ids = model.generate(inputs[\"input_ids\"], num_beams=4, max_length=50, early_stopping=True)\n\n# \uc694\uc57d \ub514\ucf54\ub529\nsummary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)\nprint(\"Summary:\", summary)<\/code><\/pre>\n5. \uc694\uc57d \uacb0\uacfc \ubd84\uc11d<\/h2>\n
6. \uacb0\ub860<\/h2>\n