\uc790\uc5f0\uc5b4 \ucc98\ub9ac(Natural Language Processing, NLP)\ub294 \uc778\uacf5\uc9c0\ub2a5 \ubd84\uc57c\uc5d0\uc11c \uc0ac\ub78c\uc758 \uc5b8\uc5b4\ub97c \uc774\ud574\ud558\uace0 \ud574\uc11d\ud558\ub294 \ubc29\ubc95\uc744 \uc5f0\uad6c\ud558\ub294 \ubd84\uc57c\uc785\ub2c8\ub2e4. NLP\ub294 \uc815\ubcf4 \uac80\uc0c9, \uae30\uacc4 \ubc88\uc5ed, \uac10\uc815 \ubd84\uc11d \ub4f1 \ub2e4\uc591\ud55c \uc560\ud50c\ub9ac\ucf00\uc774\uc158\uc5d0 \ud65c\uc6a9\ub418\uba70, \uadf8 \uc911 \uac1c\uccb4\uba85 \uc778\uc2dd(Named Entity Recognition, NER)\uc740 \ud14d\uc2a4\ud2b8\uc5d0\uc11c \uace0\uc720 \uba85\uc0ac(\uc608: \uc778\uba85, \uc9c0\uba85, \ub0a0\uc9dc \ub4f1)\ub97c \uc2dd\ubcc4\ud558\ub294 \uc911\uc694\ud55c \uc791\uc5c5\uc785\ub2c8\ub2e4. \ucd5c\uadfc \ub525 \ub7ec\ub2dd \uae30\uc220\uc758 \ubc1c\uc804\uc73c\ub85c NER \uc131\ub2a5\uc774 \ud06c\uac8c \ud5a5\uc0c1\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \uc591\ubc29\ud5a5 LSTM(Bidirectional Long Short-Term Memory) \ubaa8\ub378\uc740 \uc774\ub7ec\ud55c \uc791\uc5c5\uc5d0 \ud6a8\uacfc\uc801\uc774\uba70, \ubcf8 \uae00\uc5d0\uc11c\ub294 Bi-LSTM\uc744 \ud65c\uc6a9\ud55c NER \uc2dc\uc2a4\ud15c\uc758 \uc774\ub860\uacfc \uad6c\ud604\uc744 \uc790\uc138\ud788 \uc124\uba85\ud569\ub2c8\ub2e4.<\/p>\n
\uac1c\uccb4\uba85 \uc778\uc2dd(NER)\uc740 \uc8fc\uc5b4\uc9c4 \ud14d\uc2a4\ud2b8\uc5d0\uc11c \uc0ac\ub78c, \uc7a5\uc18c, \uae30\uad00, \ub0a0\uc9dc \ub4f1\uc758 \uac1c\uccb4\ub97c \uc2dd\ubcc4\ud558\uace0 \uc774\ub97c \ubd84\ub958\ud558\ub294 \uc791\uc5c5\uc785\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, “\uc774\ud0c8\ub9ac\uc544\uc758 \ub85c\ub9c8\ub294 \uc544\ub984\ub2e4\uc6b4 \ub3c4\uc2dc\uc785\ub2c8\ub2e4.”\ub77c\ub294 \ubb38\uc7a5\uc5d0\uc11c “\uc774\ud0c8\ub9ac\uc544”\ub294 \uc7a5\uc18c(\uc9c0\uba85)\ub85c, “\ub85c\ub9c8”\ub294 \ud2b9\uc815 \ub3c4\uc2dc\ub85c \uc778\uc2dd\ub429\ub2c8\ub2e4. NER\uc740 \uc815\ubcf4 \ucd94\ucd9c, \uc9c8\uc758 \uc751\ub2f5 \uc2dc\uc2a4\ud15c, \uae30\uacc4 \ubc88\uc5ed \ub4f1 \ub2e4\uc591\ud55c NLP \uc751\uc6a9 \ud504\ub85c\uadf8\ub7a8\uc5d0\uc11c \ud575\uc2ec\uc801\uc778 \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4.<\/p>\n
NER\uc740 \uae30\uc874\uc5d0 \uaddc\uce59 \uae30\ubc18 \uc811\uadfc \ubc29\uc2dd \ubc0f \ud1b5\uacc4\uc801 \uae30\uacc4 \ud559\uc2b5 \ubc29\ubc95\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc218\ud589\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uaddc\uce59 \uae30\ubc18 \uc811\uadfc\ubc95\uc740 \uc804\ubb38\uac00\uc758 \uc9c0\uc2dd\uc744 \ubc14\ud0d5\uc73c\ub85c \uc218\uc791\uc5c5\uc73c\ub85c \uaddc\uce59\uc744 \uc124\uc815\ud558\uc5ec \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\ub294 \ubc29\uc2dd\uc785\ub2c8\ub2e4. \ubc18\uba74 \ud1b5\uacc4\uc801 \uae30\uacc4 \ud559\uc2b5 \ubc29\ubc95\uc740 \ub300\ub7c9\uc758 \ub370\uc774\ud130\uc5d0\uc11c \ud328\ud134\uc744 \ud559\uc2b5\ud558\uc5ec \uac1c\uccb4\ub97c \uc778\uc2dd\ud558\uc9c0\ub9cc, \ubb38\ub9e5\uc744 \ud30c\uc545\ud558\ub294 \ub370 \ud55c\uacc4\uac00 \uc788\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n
\ucd5c\uadfc \uba87 \ub144\uac04 \ub525 \ub7ec\ub2dd\uc740 \uc774\ubbf8\uc9c0 \uc778\uc2dd, \uc74c\uc131 \uc778\uc2dd, \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ub4f1 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud601\uc2e0\uc801\uc778 \uacb0\uacfc\ub97c \uac00\uc838\uc654\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubcc0\ud654\ub294 \uc8fc\ub85c \uc2ec\uce35 \uc2e0\uacbd\ub9dd(Deep Neural Networks)\uc758 \ubc1c\uc804\uacfc \ub300\ub7c9\uc758 \ub370\uc774\ud130, \uac15\ub825\ud55c \ucef4\ud4e8\ud305 \ud30c\uc6cc\uc758 \uacb0\ud569 \ub355\ubd84\uc785\ub2c8\ub2e4. \ub525 \ub7ec\ub2dd\uc740 \ud2b9\ud788 \ube44\uc815\ud615 \ub370\uc774\ud130\uc5d0\uc11c \ud6a8\uacfc\uc801\uc73c\ub85c \ud2b9\uc9d5\uc744 \ucd94\ucd9c\ud560 \uc218 \uc788\uc5b4, NER\uacfc \uac19\uc740 \ubcf5\uc7a1\ud55c \ud0dc\uc2a4\ud06c\uc5d0 \uc801\ud569\ud569\ub2c8\ub2e4.<\/p>\n
\uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c \uc21c\ud658 \uc2e0\uacbd\ub9dd(Recurrent Neural Networks, RNN)\uc740 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \uc720\uc6a9\ud55c \uad6c\uc870\uc785\ub2c8\ub2e4. \uadf8\ub7ec\ub098 RNN\uc740 \uae34 \uc2dc\ud000\uc2a4\ub97c \ud559\uc2b5\ud560 \ub54c \uae30\uc6b8\uae30 \uc18c\uc2e4(Vanishing Gradient) \ubb38\uc81c\uc5d0 \uc2dc\ub2ec\ub9bd\ub2c8\ub2e4. \uc774\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 LSTM(Long Short-Term Memory) \ub124\ud2b8\uc6cc\ud06c\uac00 \uac1c\ubc1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4. LSTM\uc740 \uba54\ubaa8\ub9ac \uc140\uacfc \uac8c\uc774\ud2b8 \uad6c\uc870\ub97c \ub3c4\uc785\ud558\uc5ec \uc815\ubcf4\ub97c \ub354 \uc624\ub798 \uae30\uc5b5\ud558\uace0 \uc78a\ub294 \uac83\uc744 \uc870\uc808\ud569\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ud2b9\uc131 \ub355\ubd84\uc5d0 LSTM\uc740 \ubb38\ub9e5\uc744 \uc798 \uc774\ud574\ud558\uace0 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc758 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0b5\ub2c8\ub2e4.<\/p>\n
\uc77c\ubc18\uc801\uc778 LSTM\uc740 \uc2dc\ud000\uc2a4\ub97c \ud55c \ubc29\ud5a5\uc73c\ub85c\ub9cc \ucc98\ub9ac\ud569\ub2c8\ub2e4. \ud558\uc9c0\ub9cc Bi-LSTM\uc740 \uc2dc\ud000\uc2a4\ub97c \ub450 \ubc29\ud5a5\uc5d0\uc11c \ucc98\ub9ac\ud558\uc5ec, \ud604\uc7ac \ub2e8\uc5b4\uc5d0 \ub300\ud55c \uc774\uc804 \ub2e8\uc5b4\ubfd0\ub9cc \uc544\ub2c8\ub77c \uc774\ud6c4 \ub2e8\uc5b4\uc5d0 \ub300\ud55c \uc815\ubcf4\ub3c4 \ud559\uc2b5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \ubb38\ub9e5\uc744 \ub354 \uc798 \ud30c\uc545\ud558\uc5ec \ub354 \uc815\ud655\ud55c \uac1c\uccb4\uba85 \uc778\uc2dd\uc774 \uac00\ub2a5\ud569\ub2c8\ub2e4.<\/p>\n
NER \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\uae30 \uc704\ud574\uc11c\ub294 \uace0\uc720 \uba85\uc0ac\uac00 \ub808\uc774\ube14\ub9c1\ub41c \ud559\uc2b5 \ub370\uc774\ud130\uac00 \ud544\uc694\ud569\ub2c8\ub2e4. \ub300\ud45c\uc801\uc73c\ub85c CoNLL-2003 \ub370\uc774\ud130\uc14b\uc774 \ub110\ub9ac \uc0ac\uc6a9\ub418\uba70, \uc774 \ub370\uc774\ud130\uc14b\uc740 \uc0ac\ub78c, \uc7a5\uc18c, \uae30\uad00, \uae30\ud0c0\ub85c \ub098\ub204\uc5b4 \ub808\uc774\ube14\uc774 \ubd99\uc740 \ud14d\uc2a4\ud2b8\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4. \ub370\uc774\ud130\uc14b\uc744 \ubd88\ub7ec\uc624\uace0 \uc804\ucc98\ub9ac\ud558\ub294 \uacfc\uc815\uc774 \ubaa8\ub378 \uc131\ub2a5\uc5d0 \ud070 \uc601\ud5a5\uc744 \ubbf8\uce58\ubbc0\ub85c, \uc2e0\uc911\ud788 \uc218\ud589\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
\uc774\uc81c \ub370\uc774\ud130 \uc900\ube44\uac00 \ub05d\ub0ac\ub2e4\uba74 Bi-LSTM \ubaa8\ub378\uc744 \uad6c\ucd95\ud560 \ucc28\ub840\uc785\ub2c8\ub2e4. TensorFlow \ubc0f Keras \uac19\uc740 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc190\uc27d\uac8c \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \uad6c\ucd95\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 Bi-LSTM \ubaa8\ub378\uc758 \uc77c\ubc18\uc801\uc778 \uad6c\uc870\ub97c \ub098\ud0c0\ub0c5\ub2c8\ub2e4:<\/p>\n
\nimport numpy as np\nfrom keras.models import Sequential\nfrom keras.layers import Embedding, LSTM, Dense, Dropout, Bidirectional\n\n# \ubaa8\ub378 \ucd08\uae30\ud654\nmodel = Sequential()\n\n# \uc784\ubca0\ub529 \ub808\uc774\uc5b4\nmodel.add(Embedding(input_dim=vocab_size, output_dim=embedding_dim, input_length=max_length))\n\n# Bi-LSTM \ub808\uc774\uc5b4 \ucd94\uac00\nmodel.add(Bidirectional(LSTM(units=hidden_units, return_sequences=True)))\n\n# \ub4dc\ub86d\uc544\uc6c3 \ub808\uc774\uc5b4\nmodel.add(Dropout(rate=dropout_rate))\n\n# \ucd9c\ub825 \ub808\uc774\uc5b4\nmodel.add(Dense(units=num_classes, activation='softmax'))\n\n# \ubaa8\ub378 \ucef4\ud30c\uc77c\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n4.4 \ud559\uc2b5 \ubc0f \ud3c9\uac00<\/h3>\n
\ubaa8\ub378\uc774 \uad6c\ucd95\ub418\uba74 \ud559\uc2b5\uc744 \uc9c4\ud589\ud569\ub2c8\ub2e4. \uc801\uc808\ud55c \ud558\uc774\ud37c\ud30c\ub77c\ubbf8\ud130(learning rate, batch size \ub4f1)\ub97c \uc124\uc815\ud558\uace0, fit \uba54\uc18c\ub4dc\ub97c \ud1b5\ud574 \ud559\uc2b5\uc744 \uc218\ud589\ud569\ub2c8\ub2e4. \ud559\uc2b5\uc774 \uc644\ub8cc\ub418\uba74 \uac80\uc99d \ub370\uc774\ud130\ub97c \ud1b5\ud574 \ubaa8\ub378 \uc131\ub2a5\uc744 \ud3c9\uac00\ud569\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \ud559\uc2b5 \ucf54\ub4dc\uc758 \uc608\uc2dc\uc785\ub2c8\ub2e4:<\/p>\n
\nhistory = model.fit(train_X, train_Y, validation_data=(val_X, val_Y), epochs=epochs, batch_size=batch_size)\n<\/code><\/pre>\n4.5 \ubaa8\ub378 \uc131\ub2a5 \ud5a5\uc0c1<\/h3>\n
\ubaa8\ub378 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud574 \uc5ec\ub7ec \uac00\uc9c0 \uae30\ubc95\uc744 \uc801\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \ub370\uc774\ud130 \uc99d\uac15(Data Augmentation), \ub354 \uae4a\uc740 \ub124\ud2b8\uc6cc\ud06c \uad6c\uc131, \uc804\uc774 \ud559\uc2b5(Transfer Learning) \uae30\ubc95 \ub4f1\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. Pre-trained \ubaa8\ub378\uc744 \ud65c\uc6a9\ud558\uba74 \uc801\uc740 \uc591\uc758 \ub370\uc774\ud130\ub85c\ub3c4 \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \uc5bb\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
5. \uacb0\ub860<\/h2>\n
Bi-LSTM\uc740 \uac1c\uccb4\uba85 \uc778\uc2dd(NER) \uc791\uc5c5\uc5d0 \ud6a8\uacfc\uc801\uc778 \ub525 \ub7ec\ub2dd \ubaa8\ub378\uc785\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \ubb38\ub9e5\uc744 \uc798 \uc774\ud574\ud558\uace0 \ub2e4\uc591\ud55c \uac1c\uccb4\ub97c \uc815\ud655\ud788 \uc778\uc2dd\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \ube14\ub85c\uadf8\uc5d0\uc11c\ub294 NER\uc758 \uac1c\ub150, Bi-LSTM\uc758 \uc774\ub860\uacfc \uad6c\ud604 \uacfc\uc815\uc744 \uc0c1\uc138\ud788 \uc124\uba85\ud558\uc5ec \ub3c5\uc790\ub4e4\uc774 NER \uc2dc\uc2a4\ud15c\uc744 \uac1c\ubc1c\ud558\ub294 \ub370 \ud544\uc694\ud55c \uae30\ubcf8\uc801\uc778 \uc9c0\uc2dd\uc744 \uac16\ucd9c \uc218 \uc788\ub3c4\ub85d \ub3d5\uace0\uc790 \ud588\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c\ub3c4 NLP \ubd84\uc57c\ub294 \uc131\uc7a5\uc744 \uac70\ub4ed\ud560 \uac83\uc774\uba70, \ub354\uc6b1 \ubc1c\uc804\ub41c \uae30\ubc95\ub4e4\uc774 \uc9c0\uc18d\uc801\uc73c\ub85c \uc5f0\uad6c\ub418\uace0 \uc801\uc6a9\ub420 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"