\uc790\uc5f0\uc5b4 \ucc98\ub9ac(Natural Language Processing, NLP)\ub294 \ucef4\ud4e8\ud130 \uacfc\ud559\uacfc \uc5b8\uc5b4\ud559\uc758 \uc811\ubaa9 \ubd84\uc57c\ub85c, \uc778\uac04 \uc5b8\uc5b4\ub97c \uc774\ud574\ud558\uace0 \ucc98\ub9ac\ud558\ub294 \uae30\ubc95\uc744 \uc5f0\uad6c\ud569\ub2c8\ub2e4. \ucd5c\uadfc \uba87 \ub144\uac04 \ub525 \ub7ec\ub2dd \uae30\uc220\uc774 \ubc1c\uc804\ud558\uba74\uc11c \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c\ub3c4 \ud070 \ubcc0\ud654\uc640 \ubc1c\uc804\uc774 \uc788\uc5c8\uc2b5\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 \ub525 \ub7ec\ub2dd\uc744 \ud65c\uc6a9\ud55c \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc758 \ud55c \uc608\ub85c \ud0dc\uae45 \uc791\uc5c5, \ud2b9\ud788 \uac1c\uccb4\uba85 \uc778\uc2dd(Named Entity Recognition, NER)\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub294 \ucef4\ud4e8\ud130\uac00 \uc778\uac04\uc758 \uc5b8\uc5b4\ub97c \uc774\ud574\ud558\uace0 \ucc98\ub9ac\ud558\ub294 \ub2a5\ub825\uc744 \uc758\ubbf8\ud569\ub2c8\ub2e4. \uc774\ub294 \uc74c\uc131 \uc778\uc2dd, \ud14d\uc2a4\ud2b8 \ubd84\uc11d, \uae30\uacc4 \ubc88\uc5ed \ub4f1 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uace0\uc804\uc801\uc778 NLP \uae30\ubc95\uc740 \uc8fc\ub85c \uaddc\uce59 \uae30\ubc18 \uc811\uadfc \ubc29\uc2dd\uc5d0 \uc758\uc874\ud588\uc9c0\ub9cc, \ucd5c\uadfc\uc5d0\ub294 \ub525 \ub7ec\ub2dd \uae30\ubc18\uc758 \uc811\uadfc \ubc29\uc2dd\uc774 \ub110\ub9ac \uc0ac\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\ud0dc\uae45 \uc791\uc5c5\uc740 \ud14d\uc2a4\ud2b8\uc758 \uac01 \uc694\uc18c\uc5d0 \ub808\uc774\ube14\uc744 \ubd99\uc774\ub294 \uacfc\uc815\uc73c\ub85c, \uc774\ub294 \ud14d\uc2a4\ud2b8 \ubd84\uc11d\uc5d0\uc11c \ub9e4\uc6b0 \uc911\uc694\ud55c \uc791\uc5c5\uc785\ub2c8\ub2e4. \ub300\ud45c\uc801\uc778 \ud0dc\uae45 \uc791\uc5c5\uc73c\ub85c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uac83\ub4e4\uc774 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
NER\uc740 \ud14d\uc2a4\ud2b8\uc5d0\uc11c \uc0ac\ub78c, \uc7a5\uc18c, \uc870\uc9c1 \ub4f1\uc758 \uace0\uc720 \uba85\uc0ac\ub97c \ucc3e\uc544\ub0b4\uace0 \uc774\ub97c \ub808\uc774\ube14\ub85c \ubd84\ub958\ud558\ub294 \uc791\uc5c5\uc785\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, “Steve Jobs\ub294 Apple\uc758 CEO\uc600\ub2e4.”\ub77c\ub294 \ubb38\uc7a5\uc5d0\uc11c “Steve Jobs”\ub294 PERSON, “Apple”\uc740 ORGANIZATION\uc73c\ub85c \ud0dc\uae45\ub429\ub2c8\ub2e4.<\/p>\n
\ud488\uc0ac \ud0dc\uae45\uc740 \ud14d\uc2a4\ud2b8\uc5d0\uc11c \uac01 \ub2e8\uc5b4\uc758 \ud488\uc0ac\ub97c \uc2dd\ubcc4\ud558\ub294 \uc791\uc5c5\uc785\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, “\ub098\ub294 \ud559\uad50\uc5d0 \uac04\ub2e4.”\ub77c\ub294 \ubb38\uc7a5\uc5d0\uc11c ‘\ub098’\ub294 \ub300\uba85\uc0ac, ‘\ud559\uad50’\ub294 \uba85\uc0ac\ub85c \ud0dc\uae45\ub429\ub2c8\ub2e4.<\/p>\n
\uc804\ud1b5\uc801\uc778 \ud0dc\uae45 \uc791\uc5c5\uc740 \uaddc\uce59 \uae30\ubc18 \uc2dc\uc2a4\ud15c\uc774\ub098 \ud1b5\uacc4\uc801 \ubaa8\ub378\uc5d0 \uc758\uc874\ud588\uc2b5\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \ub525 \ub7ec\ub2dd\uc758 \ubc1c\uc804\uc73c\ub85c \uc778\ud574 \ub354 \ubcf5\uc7a1\ud558\uace0 \uc815\uad50\ud55c \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uac8c \ub418\uc5c8\uc2b5\ub2c8\ub2e4. \ud2b9\ud788 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \uc801\ud569\ud55c RNN(Recurrent Neural Network) \ubc0f \uadf8 \ubcc0\ud615\uc778 LSTM(Long Short-Term Memory) \ub124\ud2b8\uc6cc\ud06c\uac00 \ub9ce\uc774 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n
RNN\uc740 \uc2dc\ud000\uc2a4 \ub370\uc774\ud130\uc778 \ud14d\uc2a4\ud2b8\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \uc720\uc6a9\ud558\uc9c0\ub9cc, \uae34 \uc2dc\ud000\uc2a4\ub97c \ucc98\ub9ac\ud560 \ub54c\ub294 \uc815\ubcf4\uac00 \uc18c\uc2e4\ub418\ub294 \ub2e8\uc810\uc774 \uc788\uc2b5\ub2c8\ub2e4. LSTM\uc740 \uc774\ub7ec\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 \uc124\uacc4\ub41c \uad6c\uc870\ub85c, \uc7a5\uae30 \uc758\uc874\uc131\uc744 \ud559\uc2b5\ud560 \uc218 \uc788\ub3c4\ub85d \ub3d5\uc2b5\ub2c8\ub2e4. LSTM\uc740 \uc140 \uc0c1\ud0dc(cell state)\uc640 \uc5ec\ub7ec \uac1c\uc758 \uac8c\uc774\ud2b8(gate)\ub97c \uc774\uc6a9\ud558\uc5ec \uc815\ubcf4\ub97c \uc800\uc7a5\ud558\uace0 \uad00\ub9ac\ud569\ub2c8\ub2e4.<\/p>\n
import numpy as np\nfrom keras.models import Sequential\nfrom keras.layers import LSTM, Embedding, Dense, TimeDistributed, Dropout, Bidirectional\nfrom keras.preprocessing.sequence import pad_sequences\n\n# \ub370\uc774\ud130 \uc900\ube44\nX = # \uc785\ub825 \ub370\uc774\ud130 (\ub2e8\uc5b4 \uc778\ub371\uc2a4)\ny = # \ucd9c\ub825 \ub370\uc774\ud130 (One-hot \uc778\ucf54\ub529\ub41c \ub808\uc774\ube14)\n\n# \ud328\ub529\nX_pad = pad_sequences(X, padding='post')\n\n# \ubaa8\ub378 \uc815\uc758\nmodel = Sequential()\nmodel.add(Embedding(input_dim=vocab_size, output_dim=embedding_dim, input_length=max_length))\nmodel.add(Bidirectional(LSTM(units=64, return_sequences=True)))\nmodel.add(TimeDistributed(Dense(num_classes, activation='softmax')))\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n\n# \ubaa8\ub378 \ud559\uc2b5\nmodel.fit(X_pad, y, batch_size=32, epochs=10)<\/code><\/pre>\n5. \ud0dc\uae45 \uc791\uc5c5\uc5d0\uc11c\uc758 \ud3c9\uac00<\/h2>\n
\ud0dc\uae45 \uc791\uc5c5\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\ub294 \ubc29\ubc95\uc73c\ub85c\ub294 \uc815\ubc00\ub3c4(Precision), \uc7ac\ud604\uc728(Recall), F1-score \ub4f1\uc774 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uc9c0\ud45c\ub4e4\uc740 \ubaa8\ub378\uc774 \uc5bc\ub9c8\ub098 \uc815\ud655\ud558\uac8c \ud0dc\uae45\ud588\ub294\uc9c0\ub97c \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n
5.1 \uc815\ubc00\ub3c4(Precision)<\/h3>\n
\uc815\ubc00\ub3c4\ub294 \ubaa8\ub378\uc774 \uc608\uce21\ud55c \ud0dc\uadf8 \uc911\uc5d0\uc11c \uc2e4\uc81c\ub85c \ub9de\uc740 \ube44\uc728\uc785\ub2c8\ub2e4. \uacf5\uc2dd\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
Precision = True Positives \/ (True Positives + False Positives)<\/code><\/pre>\n5.2 \uc7ac\ud604\uc728(Recall)<\/h3>\n
\uc7ac\ud604\uc728\uc740 \uc2e4\uc81c \ud0dc\uadf8 \uc911\uc5d0\uc11c \ubaa8\ub378\uc774 \ub9de\ucd98 \ube44\uc728\uc785\ub2c8\ub2e4. \uacf5\uc2dd\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
Recall = True Positives \/ (True Positives + False Negatives)<\/code><\/pre>\n5.3 F1-score<\/h3>\n
F1-score\ub294 \uc815\ubc00\ub3c4\uc640 \uc7ac\ud604\uc728\uc758 \uc870\ud654 \ud3c9\uade0\uc73c\ub85c, \ub450 \uc9c0\ud45c \uac04\uc758 \uade0\ud615\uc744 \uce21\uc815\ud569\ub2c8\ub2e4. \uacf5\uc2dd\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
F1 = 2 * (Precision * Recall) \/ (Precision + Recall)<\/code><\/pre>\n6. \uacb0\ub860<\/h2>\n
\ub525 \ub7ec\ub2dd \uae30\uc220\uc740 \ud0dc\uae45 \uc791\uc5c5\uc5d0\uc11c \ud070 \ubc1c\uc804\uc744 \uc774\ub8e8\uc5c8\uc73c\uba70, \ud2b9\ud788 LSTM \ub124\ud2b8\uc6cc\ud06c\ub294 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c \uac15\ub825\ud55c \ub3c4\uad6c\ub85c \uc790\ub9ac\uc7a1\uc558\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c\uc758 \uc5f0\uad6c\uc640 \ubc1c\uc804\uc744 \ud1b5\ud574 \ub354\uc6b1 \uc815\uad50\ud558\uace0 \ud6a8\uc728\uc801\uc778 \uc790\uc5f0\uc5b4 \ucc98\ub9ac \uae30\uc220\uc774 \ub4f1\uc7a5\ud560 \uac83\uc73c\ub85c \uae30\ub300\ub429\ub2c8\ub2e4. \ud0dc\uae45 \uc791\uc5c5\uc740 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc758 \ud575\uc2ec\uc801\uc778 \uae30\uc220 \uc911 \ud558\ub098\ub85c, \ub2e4\uc591\ud55c \uc560\ud50c\ub9ac\ucf00\uc774\uc158\uc5d0\uc11c \ud65c\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
7. \ucc38\uace0 \ubb38\ud5cc<\/h2>\n