\n\ubcf8 \uae00\uc5d0\uc11c\ub294 T5 \ubaa8\ub378(TexT-to-Text Transfer Transformer)\uc744 \ud65c\uc6a9\ud558\uc5ec \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP)\uc5d0\uc11c \uc694\uc57d \uc0dd\uc131\uae30\ub97c \ub9cc\ub4dc\ub294 \uc2e4\uc2b5\uc744 \ub2e4\ub8f0 \uac83\uc785\ub2c8\ub2e4. T5\ub294 \ud14d\uc2a4\ud2b8 \uc785\ub825\uc744 \ud14d\uc2a4\ud2b8 \ucd9c\ub825\uc73c\ub85c \ubcc0\ud658\ud558\ub294 \ubaa8\ub378\ub85c\uc11c \ub2e4\uc591\ud55c NLP \uc791\uc5c5\uc744 \uc218\ud589\ud560 \uc218 \uc788\ub294 \uac15\ub825\ud55c \ub3c4\uad6c\uc785\ub2c8\ub2e4. \uc774 \uae00\uc744 \ud1b5\ud574 T5 \ubaa8\ub378\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \uc694\uc57d \uc0dd\uc131 \uacfc\uc815\uc744 \uc790\uc138\ud788 \uc124\uba85\ud558\uace0, \uc2e4\uc81c \ud30c\uc778 \ud29c\ub2dd \uacfc\uc815\uc744 \uc2dc\uc5f0\ud558\uba70, \uacb0\uacfc\ub97c \ud3c9\uac00\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n<\/header>\n\n1. T5 \ubaa8\ub378 \uc18c\uac1c<\/h2>\n
T5\ub294 Google\uc5d0\uc11c \uac1c\ubc1c\ud55c Transformer \uae30\ubc18\uc758 \ubaa8\ub378\ub85c, \ub2e4\uc591\ud55c \ud14d\uc2a4\ud2b8 \ubcc0\ud658 \uc791\uc5c5\uc744 \ucc98\ub9ac\ud558\uae38 \ubaa9\uc801\uc73c\ub85c \uc124\uacc4\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 “\ubaa8\ub4e0 NLP \ubb38\uc81c\ub97c \ud14d\uc2a4\ud2b8 \ubcc0\ud658 \ubb38\uc81c\ub85c \uc7ac\uad6c\uc131\ud55c\ub2e4”\ub294 \ucca0\ud559\uc744 \uae30\ubc18\uc73c\ub85c \ud569\ub2c8\ub2e4. T5\ub294 \ub2e4\uc591\ud55c \ub370\uc774\ud130\uc14b\uc5d0\uc11c \uc0ac\uc804 \ud6c8\ub828\ub41c \ud6c4, \ud2b9\uc815 \uc791\uc5c5\uc5d0 \ub300\ud574 \ud30c\uc778 \ud29c\ub2dd\uc744 \ud1b5\ud574 \uc131\ub2a5\uc744 \uadf9\ub300\ud654\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\ubaa8\ub378\uc758 \uad6c\uc870\ub294 \uc778\ucf54\ub354-\ub514\ucf54\ub354 \uad6c\uc870\ub97c \uae30\ubc18\uc73c\ub85c \ud558\uba70, \uba40\ud2f0 \ud5e4\ub4dc \uc140\ud504 \uc5b4\ud150\uc158(self-attention) \uba54\ucee4\ub2c8\uc998\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uc774 \ub355\ubd84\uc5d0 \ubaa8\ub378\uc740 \ubb38\ub9e5\uc744 \uc774\ud574\ud558\uace0, \ud750\ub984\uc5d0 \ub9de\ub294 \ud14d\uc2a4\ud2b8 \uc0dd\uc131\uc744 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n\n2. \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc5d0\uc11c \uc694\uc57d \uc0dd\uc131\uc758 \uc911\uc694\uc131<\/h2>\n
\uc694\uc57d \uc0dd\uc131(summary generation)\uc740 \uc6d0\ubb38\uc744 \ucd95\uc57d\ud558\uc5ec \uc911\uc694\ud55c \uc815\ubcf4\ub9cc\uc744 \uc804\ub2ec\ud558\ub294 \uc791\uc5c5\uc785\ub2c8\ub2e4. \uc774\ub294 \uc815\ubcf4\uc758 \uc591\uc774 \ubc29\ub300\ud574\uc9c0\ub294 \ud604\ub300 \uc0ac\ud68c\uc5d0\uc11c \ud575\uc2ec\uc801\uc778 \uae30\uc220\ub85c \uc790\ub9ac\uc7a1\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ud6a8\uc728\uc801\uc778 \uc694\uc57d \uc0dd\uc131\uae30\ub97c \ud1b5\ud574 \uc6b0\ub9ac\ub294 \ub354 \ube60\ub974\uac8c \ud544\uc694\ud55c \uc815\ubcf4\ub97c \uc5bb\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\ub525 \ub7ec\ub2dd \uae30\ubc95\uc774 \ubc1c\uc804\ud558\uba74\uc11c, \uc778\uacf5\uc9c0\ub2a5(AI) \uae30\ubc18 \uc694\uc57d \uc0dd\uc131\uae30\uac00 \ub4f1\uc7a5\ud588\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uae30\ubc95\ub4e4\uc740 \ub192\uc740 \uc815\ud655\ub3c4\ub97c \ubc14\ud0d5\uc73c\ub85c \uc694\uc57d\uc758 \ud488\uc9c8\uc744 \ud06c\uac8c \ud5a5\uc0c1\uc2dc\ucf1c \uc654\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n\n3. T5 \ubaa8\ub378 \uc0ac\uc6a9\uc744 \uc704\ud55c \uc900\ube44<\/h2>\n
T5 \ubaa8\ub378\uc744 \uc0ac\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. Hugging Face\uc758 Transformers \ub77c\uc774\ube0c\ub7ec\ub9ac\ub294 T5 \ubaa8\ub378\uacfc \uadf8 \uc678 \ub2e4\uc591\ud55c NLP \ubaa8\ub378\uc744 \uc27d\uac8c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d \ub3c4\uc640\uc90d\ub2c8\ub2e4.<\/p>\n
pip install transformers datasets<\/code><\/pre>\n\uc774\ud6c4, \ub370\uc774\ud130\uc14b\uc744 \uc120\ud0dd\ud558\uace0 \uc694\uc57d \uc791\uc5c5\uc744 \uc704\ud55c \ub370\uc774\ud130 \uc804\ucc98\ub9ac\ub97c \uc9c4\ud589\ud574\uc57c \ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, CNN\/Daily Mail \ub370\uc774\ud130\uc14b\uc740 \ub274\uc2a4 \uae30\uc0ac\ub97c \uc694\uc57d\ud558\ub294 \ub370 \uc801\ud569\ud55c \ub370\uc774\ud130\uac00 \ub420 \uac83\uc785\ub2c8\ub2e4.<\/p>\n<\/section>\n\n4. \ub370\uc774\ud130\uc14b \ub85c\ub4dc \ubc0f \uc804\ucc98\ub9ac<\/h2>\n
\ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud558\uace0 \uc804\ucc98\ub9ac\ud558\ub294 \uacfc\uc815\uc740 \ubaa8\ub378 \ud6c8\ub828\uc758 \ud575\uc2ec \ub2e8\uacc4\uc785\ub2c8\ub2e4. \uc544\ub798\ub294 Hugging Face\uc758 datasets \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud558\uc5ec CNN\/Daily Mail \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud558\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4.<\/p>\n
\nfrom datasets import load_dataset\ndataset = load_dataset('cnn_dailymail', '3.0.0')\n <\/code><\/pre>\n\ub370\uc774\ud130\uc14b\uc774 \ub85c\ub4dc\ub418\uba74, \uc694\uc57d \uc0dd\uc131\uc744 \uc704\ud55c \uc785\ub825 \ub370\uc774\ud130\uc640 \ucd9c\ub825 \ub370\uc774\ud130\ub97c \ub098\ub204\ub294 \uacfc\uc815\uc774 \ud544\uc694\ud569\ub2c8\ub2e4.<\/p>\n<\/section>\n\n5. T5 \ubaa8\ub378 \ud30c\uc778 \ud29c\ub2dd<\/h2>\n
T5 \ubaa8\ub378\uc744 \ud30c\uc778 \ud29c\ub2dd\ud558\ub294 \uacfc\uc815\uc740 \ud06c\uac8c \ub370\uc774\ud130 \uc900\ube44, \ubaa8\ub378 \ucd08\uae30\ud654, \ud6c8\ub828, \ud3c9\uac00\ub85c \ub098\ub20c \uc218 \uc788\uc2b5\ub2c8\ub2e4. Hugging Face\uc758 Trainer API\ub97c \ud65c\uc6a9\ud558\uba74 \uc774 \uacfc\uc815\uc744 \uc27d\uac8c \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\nfrom transformers import T5Tokenizer, T5ForConditionalGeneration, Trainer, TrainingArguments\n\ntokenizer = T5Tokenizer.from_pretrained(\"t5-base\")\nmodel = T5ForConditionalGeneration.from_pretrained(\"t5-base\")\n\n# \ub370\uc774\ud130\uc14b \uc804\ucc98\ub9ac \ubc0f \ud150\uc11c \ubcc0\ud658 \uacfc\uc815 \uc0dd\ub7b5\ntraining_args = TrainingArguments(\n output_dir=\".\/results\",\n per_device_train_batch_size=8,\n num_train_epochs=3,\n logging_dir='.\/logs',\n)\ntrainer = Trainer(\n model=model,\n args=training_args,\n train_dataset=train_dataset,\n eval_dataset=eval_dataset\n)\ntrainer.train()\n <\/code><\/pre>\n\ud6c8\ub828\uc774 \uc644\ub8cc\ub41c \ud6c4\uc5d0\ub294 \ubaa8\ub378\uc744 \ud3c9\uac00\ud558\uc5ec \uc2e4\uc81c \uc131\ub2a5\uc744 \uce21\uc815\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n\n6. \uacb0\uacfc \ud3c9\uac00 \ubc0f \ud65c\uc6a9<\/h2>\n
\ud6c8\ub828\ub41c \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\uae30 \uc704\ud574 ROUGE \uc810\uc218\uc640 \uac19\uc740 \uba54\ud2b8\ub9ad\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uc810\uc218\ub294 \uc0dd\uc131\ub41c \uc694\uc57d\uacfc \uc2e4\uc81c \uc694\uc57d \uac04\uc758 \uc720\uc0ac\uc131\uc744 \uce21\uc815\ud569\ub2c8\ub2e4.<\/p>\n
\nfrom datasets import load_metric\nmetric = load_metric(\"rouge\")\n\npredictions = trainer.predict(eval_dataset)\nresults = metric.compute(predictions=predictions.predictions, references=predictions.label_ids)\n <\/code><\/pre>\n\ud3c9\uac00 \uacb0\uacfc\ub97c \ubc14\ud0d5\uc73c\ub85c \ubaa8\ub378\uc744 \uac1c\uc120\ud558\uac70\ub098, \ucd94\uac00\uc801\uc778 \ub370\uc774\ud130\ub85c \uc7ac\ud6c8\ub828\uc744 \uace0\ub824\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n