\ucd5c\uadfc \uba87 \ub144\uac04 \uc778\uacf5\uc9c0\ub2a5(AI) \uae30\uc220\uc758 \ubc1c\ub2ec\uc740 \uc6b0\ub9ac \uc0dd\ud65c\uc758 \ub9ce\uc740 \ubd80\ubd84\uc5d0 \uae4a\uc219\uc774 \uce68\ud22c\ud558\uace0 \uc788\uc73c\uba70, \ud2b9\ud788 \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP) \ubd84\uc57c\uc5d0\uc11c\uc758 \ud601\uc2e0\uc740 \uc5c4\uccad\ub09c \ubc1c\uc804\uc744 \ubcf4\uc600\uc2b5\ub2c8\ub2e4. \uadf8\uc911\uc5d0\uc11c\ub3c4 \uc21c\ud658 \uc2e0\uacbd\ub9dd(Recurrent Neural Network, RNN)\uc740 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc5d0\uc11c \ub9e4\uc6b0 \uc911\uc694\ud55c \uc5ed\ud560\uc744 \uc218\ud589\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 RNN\uc744 \uae30\ubc18\uc73c\ub85c \ud55c \uc5b8\uc5b4 \ubaa8\ub378, \uc989 RNN \uc5b8\uc5b4 \ubaa8\ub378(Recurrent Neural Network Language Model, RNNLM)\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP, Natural Language Processing)\ub294 \ucef4\ud4e8\ud130\uc640 \uc778\uac04\uc758 \uc5b8\uc5b4 \uac04\uc758 \uc0c1\ud638\uc791\uc6a9\uc744 \ub2e4\ub8e8\ub294 \uc778\uacf5\uc9c0\ub2a5\uc758 \ud55c \ubd84\uc57c\uc785\ub2c8\ub2e4. NLP\uc758 \ubaa9\uc801\uc740 \ucef4\ud4e8\ud130\uac00 \uc790\uc5f0\uc5b4(\uc778\uac04\uc774 \uc0ac\uc6a9\ud558\ub294 \uc5b8\uc5b4)\ub97c \uc774\ud574\ud558\uace0, \ud574\uc11d\ud558\uace0, \uc0dd\uc131\ud558\ub3c4\ub85d \ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \uc74c\uc131 \uc778\uc2dd, \uae30\uacc4 \ubc88\uc5ed, \uac10\uc815 \ubd84\uc11d \ub4f1 \ub2e4\uc591\ud55c \uc751\uc6a9 \ud504\ub85c\uadf8\ub7a8\uc5d0\uc11c \ud65c\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\ub525 \ub7ec\ub2dd(Deep Learning)\uc740 \uba38\uc2e0\ub7ec\ub2dd\uc758 \ud55c \ubd84\uc57c\ub85c, \uc2e0\uacbd\ub9dd\uc744 \ud1b5\ud574 \ub370\uc774\ud130\uc5d0\uc11c \ud328\ud134\uc744 \ud559\uc2b5\ud558\ub294 \uc54c\uace0\ub9ac\uc998\uc785\ub2c8\ub2e4. NLP\uc5d0\uc11c \ub525 \ub7ec\ub2dd \uae30\uc220\uc758 \ub3c4\uc785\uc740 \uc804\ud1b5\uc801\uc778 \ubc29\ubc95\ub4e4\ubcf4\ub2e4 \ub354 \ub192\uc740 \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc8fc\uace0 \uc788\uc73c\uba70, \ud2b9\ud788 \ub300\ub7c9\uc758 \ub370\uc774\ud130\uc640 \uac15\ub825\ud55c \ucef4\ud4e8\ud305 \ud30c\uc6cc\uc758 \uacb0\ud569\uc73c\ub85c \uc778\ud574 \uac00\uc18d\ud654\ub418\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n
RNN \uc5b8\uc5b4 \ubaa8\ub378\uc740 \ud14d\uc2a4\ud2b8\uc5d0\uc11c \ub2e8\uc5b4\uc758 \ubc1c\uc0dd \ud655\ub960\uc744 \ubaa8\ub378\ub9c1\ud558\ub294 \ub370 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uc804\ud1b5\uc801\uc778 \uc5b8\uc5b4 \ubaa8\ub378(\uc608: n-\uadf8\ub7a8 \ubaa8\ub378)\uc740 \uc9c1\uc811\uc801\uc73c\ub85c \ub2e8\uc5b4\uc758 \ubc1c\uc0dd \ud655\ub960\uc744 \ucd94\uc815\ud558\ub294 \ub370 \ud55c\uacc4\uac00 \uc788\uc9c0\ub9cc, RNN\uc740 \uc21c\ucc28 \ub370\uc774\ud130\uc758 \ud328\ud134\uc744 \ud559\uc2b5\ud568\uc73c\ub85c\uc368 \uc774\ub7ec\ud55c \ud55c\uacc4\ub97c \uadf9\ubcf5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
RNN\uc740 \uc785\ub825\uac12\uc744 \uc21c\uc11c\ub300\ub85c \ud558\ub098\uc529 \ucc98\ub9ac\ud558\uba70, \uc774\uc804 \uc0c1\ud0dc(hidden state)\ub97c \ub2e4\uc74c \uc0c1\ud0dc\ub85c \uc804\ub2ec\ud558\ub294 \uad6c\uc870\ub97c \uac00\uc9d1\ub2c8\ub2e4. \uc774 \uad6c\uc870 \ub355\ubd84\uc5d0 RNN\uc740 \uc2dc\uac04\uc5d0 \ub530\ub978 \uc815\ubcf4\uc758 \ud750\ub984\uc744 \ubaa8\ub378\ub9c1\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uae30\ubcf8\uc801\uc778 RNN \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n# \uae30\ubcf8 RNN \uc140 \uad6c\uc870 pseudocode\nfor t in range(1, T):\n h[t] = f(W * h[t-1] + U * x[t])\n y[t] = g(V * h[t])\n<\/code><\/pre>\n\uc5ec\uae30\uc11c h[t]<\/code>\ub294 t\ubc88\uc9f8 \uc2dc\uac04\uc5d0\uc11c\uc758 hidden state, x[t]<\/code>\ub294 t\ubc88\uc9f8 \uc785\ub825 \ub370\uc774\ud130, \uadf8\ub9ac\uace0 y[t]<\/code>\ub294 t\ubc88\uc9f8 \ucd9c\ub825 \ub370\uc774\ud130\uc785\ub2c8\ub2e4. W<\/code>, U<\/code>, V<\/code>\ub294 \ud559\uc2b5 \uac00\ub2a5\ud55c \ud30c\ub77c\ubbf8\ud130\uc785\ub2c8\ub2e4.<\/p>\n3.2 RNN\uc758 \ud55c\uacc4<\/h3>\n
RNN\uc740 \uc7a5\uae30 \uc758\uc874\uc131(long-term dependency) \ubb38\uc81c\ub97c \uac00\uc9c0\ub294\ub370, \uc774\ub294 \uc624\ub79c \uc2dc\uac04 \uac04\uaca9\uc744 \ub450\uace0 \uc788\ub294 \uc785\ub825 \uac04\uc758 \uad00\uacc4\ub97c \ud559\uc2b5\ud558\ub294 \ub370 \uc5b4\ub824\uc6c0\uc744 \uacaa\ub294\ub2e4\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 LSTM\uacfc GRU\uc640 \uac19\uc740 \uac1c\uc120\ub41c RNN \uad6c\uc870\ub4e4\uc774 \uac1c\ubc1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n
4. RNN \uc5b8\uc5b4 \ubaa8\ub378\uc758 \uad6c\ucd95<\/h2>\n
RNN \uc5b8\uc5b4 \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\ub294 \ud750\ub984\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \ub370\uc774\ud130 \uc218\uc9d1:<\/strong> \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\uc14b\uc744 \uc218\uc9d1\ud569\ub2c8\ub2e4.<\/li>\n
- \ub370\uc774\ud130 \uc804\ucc98\ub9ac:<\/strong> \uc218\uc9d1\ub41c \ub370\uc774\ud130\ub97c \uc815\uc81c\ud558\uc5ec \ub2e8\uc5b4\uc758 \ub9ac\uc2a4\ud2b8\ub85c \ubcc0\ud658\ud558\uace0, \uc815\uc218 \uc778\ucf54\ub529\uc744 \uc218\ud589\ud569\ub2c8\ub2e4.<\/li>\n
- \ubaa8\ub378 \uc124\uacc4:<\/strong> RNN \uad6c\uc870\ub97c \uc124\uacc4\ud569\ub2c8\ub2e4.<\/li>\n
- \ubaa8\ub378 \ud6c8\ub828:<\/strong> \uc190\uc2e4 \ud568\uc218\ub97c \ucd5c\uc18c\ud654\ud558\ub3c4\ub85d \ubaa8\ub378\uc744 \ud6c8\ub828\ud569\ub2c8\ub2e4.<\/li>\n
- \ubaa8\ub378 \ud3c9\uac00:<\/strong> \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\ub97c \ud1b5\ud574 \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud569\ub2c8\ub2e4.<\/li>\n<\/ol>\n
4.1 \ub370\uc774\ud130 \uc804\ucc98\ub9ac<\/h3>\n
\ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\ub294 \uc77c\ubc18\uc801\uc73c\ub85c \ub2e4\uc74c\uacfc \uac19\uc740 \uc804\ucc98\ub9ac \uacfc\uc815\uc744 \uac70\uce69\ub2c8\ub2e4:<\/p>\n
\n- HTML \ud0dc\uadf8 \uc81c\uac70<\/li>\n
- \uc18c\ubb38\uc790 \ubcc0\ud658<\/li>\n
- \ud2b9\uc218\ubb38\uc790 \uc81c\uac70<\/li>\n
- \ud1a0\ud070\ud654(tokenization)<\/li>\n
- \uc815\uc218 \uc778\ucf54\ub529(Integer Encoding)<\/li>\n<\/ul>\n
\uc608\ub97c \ub4e4\uc5b4, \ub2e4\uc74c \ubb38\uc7a5\uc744 \uace0\ub824\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4:<\/p>\n
\n\"\ub525 \ub7ec\ub2dd\uc740 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc758 \uc911\uc694\ud55c \ubc29\ubc95\uc785\ub2c8\ub2e4.\"\n<\/code><\/pre>\n\uc774 \ubb38\uc7a5\uc740 \ub2e4\uc74c\uacfc \uac19\uc774 \uc804\ucc98\ub9ac\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \ud1a0\ud070\ud654: [“\ub525”, “\ub7ec\ub2dd”, “\uc740”, “\uc790\uc5f0\uc5b4”, “\ucc98\ub9ac”, “\uc758”, “\uc911\uc694\ud55c”, “\ubc29\ubc95”, “\uc785\ub2c8\ub2e4”]<\/li>\n
- \uc815\uc218 \uc778\ucf54\ub529: [1, 2, 3, 4, 5, 6, 7, 8, 9]<\/li>\n<\/ul>\n
4.2 \ubaa8\ub378 \uc124\uacc4<\/h3>\n
\ubaa8\ub378\uc740 \uc77c\ubc18\uc801\uc73c\ub85c \ub2e4\uc74c\uacfc \uac19\uc740 \uad6c\uc131 \uc694\uc18c\ub97c \uac00\uc9d1\ub2c8\ub2e4:<\/p>\n
\n- \uc784\ubca0\ub529 \uce35(Embedding Layer)<\/li>\n
- RNN \uce35(Recurrent Layer)<\/li>\n
- \ucd9c\ub825 \uce35(Output Layer)<\/li>\n<\/ul>\n
\ub2e4\uc74c\uc740 TensorFlow\ub97c \uc0ac\uc6a9\ud55c RNNLM\uc758 \uc608\uc2dc \ucf54\ub4dc\uc785\ub2c8\ub2e4:<\/p>\n
\nimport tensorflow as tf\n\nmodel = tf.keras.Sequential([\n tf.keras.layers.Embedding(input_dim=vocab_size, output_dim=embedding_dim, input_length=max_length),\n tf.keras.layers.SimpleRNN(units=hidden_units, return_sequences=True),\n tf.keras.layers.Dense(vocab_size, activation='softmax')\n])\n<\/code><\/pre>\n4.3 \ubaa8\ub378 \ud6c8\ub828<\/h3>\n
\ubaa8\ub378 \ud6c8\ub828\uc740 \ub370\uc774\ud130\ub97c \ub124\ud2b8\uc6cc\ud06c\ub97c \ud1b5\ud574 \uc804\ub2ec\ud558\uace0, \uc190\uc2e4 \ud568\uc218\ub97c \ucd5c\uc18c\ud654\ud558\ub3c4\ub85d \ud30c\ub77c\ubbf8\ud130\ub97c \uc870\uc815\ud558\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4. \uc190\uc2e4 \ud568\uc218\ub85c\ub294 \uc8fc\ub85c \uad50\ucc28 \uc5d4\ud2b8\ub85c\ud53c(cross-entropy) \ubc29\uc2dd\uc774 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n
4.4 \ubaa8\ub378 \ud3c9\uac00<\/h3>\n
\ud6c8\ub828\uc774 \uc644\ub8cc\ub41c \ubaa8\ub378\uc740 \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc5d0 \ub300\ud574 \ud3c9\uac00\ub429\ub2c8\ub2e4. \uc774\ub294 \ubaa8\ub378\uc774 \uc2e4\uc81c \ub370\uc774\ud130\uc5d0 \uc5bc\ub9c8\ub098 \uc798 \uc77c\ubc18\ud654\ub418\ub294\uc9c0\ub97c \uce21\uc815\ud558\ub294 \ub370 \uc911\uc694\ud569\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc73c\ub85c \uc815\ud655\ub3c4(accuracy), \uc815\ubc00\ub3c4(precision), \uc7ac\ud604\uc728(recall) \ub4f1\uc758 \ub2e4\uc591\ud55c \uba54\ud2b8\ub9ad\uc774 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n
5. RNN \uc5b8\uc5b4 \ubaa8\ub378\uc758 \uc751\uc6a9<\/h2>\n
RNN \uc5b8\uc5b4 \ubaa8\ub378\uc740 \ub2e4\uc591\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uc751\uc6a9 \ud504\ub85c\uadf8\ub7a8\uc5d0\uc11c \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n
\n- \uae30\uacc4 \ubc88\uc5ed<\/li>\n
- \uc74c\uc131 \uc778\uc2dd<\/li>\n
- \ub300\ud654\ud615 AI<\/li>\n
- \ud14d\uc2a4\ud2b8 \uc0dd\uc131<\/li>\n<\/ul>\n
\uc608\ub97c \ub4e4\uc5b4, \ud14d\uc2a4\ud2b8 \uc0dd\uc131\uc5d0\uc11c\ub294 \uc8fc\uc5b4\uc9c4 \uc2dc\ud000\uc2a4\uc5d0 \ub530\ub77c \ub2e4\uc74c \ub2e8\uc5b4\ub97c \uc608\uce21\ud558\ub294 \ubc29\uc2dd\uc73c\ub85c \ud65c\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n
6. \uacb0\ub860<\/h2>\n
RNN \uc5b8\uc5b4 \ubaa8\ub378\uc740 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc758 \uc911\uc694\ud55c \ubd80\ubd84\uc73c\ub85c \uc790\ub9ac\uc7a1\uc558\uace0, \ud604\ub300 AI \uae30\uc220\uc758 \ubc1c\uc804\uacfc \ud568\uaed8 \uadf8 \ud65c\uc6a9 \ubc94\uc704\uac00 \ub354\uc6b1 \ub113\uc5b4\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\ub97c \ud1b5\ud574 RNN \uc5b8\uc5b4 \ubaa8\ub378\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \uad6c\ucd95 \ubc29\ubc95\uc5d0 \ub300\ud574 \ubc30\uc6b8 \uc218 \uc788\uc5c8\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub354 \ubc1c\uc804\ub41c \ub525\ub7ec\ub2dd \uae30\ubc18 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uae30\uc220\uc5d0 \ub300\ud574 \uc9c0\uc18d\uc801\uc73c\ub85c \uad00\uc2ec\uc744 \uac00\uc9c0\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0 \uc790\ub8cc<\/h2>\n