\n \uc790\uc5f0\uc5b4 \ucc98\ub9ac(Natural Language Processing, NLP)\ub294 \uc0ac\uc6a9\uc790\uac00 \uc758\ub3c4\ud55c \ubc14\ub97c \uc774\ud574\ud558\uace0, \ubb38\ub9e5\uc5d0 \ub9de\ub294 \uc751\ub2f5\uc744 \uc0dd\uc131\ud558\uba70, \uc5ec\ub7ec \uc5b8\uc5b4\uc801 \uc694\uc18c\ub97c \ubd84\uc11d\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc5d0\uc11c \uc911\uc694\ud55c \uae30\uc220 \uc911 \ud558\ub098\uac00 \uc784\ubca0\ub529(Embedding)<\/strong>\uc785\ub2c8\ub2e4. \uc784\ubca0\ub529\uc740 \ub2e8\uc5b4\ub97c \ubca1\ud130 \uacf5\uac04\uc5d0 \ub9e4\ud551\ud558\uc5ec \uc758\ubbf8\uc801 \uad00\uacc4\ub97c \ub354 \uc218\uce58\uc801\uc73c\ub85c \ud45c\ud604\ud560 \uc218 \uc788\ub3c4\ub85d \ub3c4\uc640\uc90d\ub2c8\ub2e4. \uc624\ub298\uc740 \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub97c \uc704\ud55c \ub2e8\uc5b4 \uc784\ubca0\ub529\uc744 \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n \n \uc784\ubca0\ub529\uc740 \ubcf4\ud1b5 \uace0\ucc28\uc6d0 \ub370\uc774\ud130\ub97c \uc800\ucc28\uc6d0\uc73c\ub85c \ubcc0\ud658\ud558\ub294 \ubc29\ubc95\uc73c\ub85c, \ud2b9\ud788 \ud14d\uc2a4\ud2b8\uc640 \uac19\uc774 \ube44\uc815\ud615 \ub370\uc774\ud130\ub97c \ub2e4\ub8f0 \ub54c \uc911\uc694\ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, ‘\uc0ac\uacfc’, ‘\ubc14\ub098\ub098’, ‘\uc624\ub80c\uc9c0’\ub77c\ub294 \uc138 \ub2e8\uc5b4\ub294 \uac01\uac01\uc758 \uc758\ubbf8\uac00 \uc11c\ub85c \ub2e4\ub974\uc9c0\ub9cc, \uc774 \ub2e8\uc5b4\ub97c \ubca1\ud130\ub85c \ubcc0\ud658\ud560 \uacbd\uc6b0 \ube44\uc2b7\ud55c \uac70\ub9ac\ub85c \ud45c\ud604\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc774 \uc758\ubbf8\ub97c \uc774\ud574\ud558\ub294 \ub370 \ub3c4\uc6c0\uc774 \ub429\ub2c8\ub2e4.\n <\/p>\n \n \uc6d0-\ud56b \uc778\ucf54\ub529\uc740 \uac01 \ub2e8\uc5b4\ub97c \uace0\uc720\ud55c \ubca1\ud130\ub85c \ubcc0\ud658\ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, ‘\uc0ac\uacfc’, ‘\ubc14\ub098\ub098’, ‘\uc624\ub80c\uc9c0’\ub77c\ub294 \ub2e8\uc5b4\ub294 \uac01\uac01 [1, 0, 0], [0, 1, 0], [0, 0, 1]\ub85c \ud45c\ud604\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \uc774 \ubc29\ubc95\uc740 \ub2e8\uc5b4 \uac04\uc758 \uc720\uc0ac\uc131\uc744 \uace0\ub824\ud558\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4.\n <\/p>\n \n \uc6cc\ub4dc \ud22c \ubca1\ud130\ub294 \ub2e8\uc5b4\uc758 \ub9e5\ub77d\uc744 \uace0\ub824\ud558\uc5ec \ubc00\uc9d1 \ubca1\ud130\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4. \uc774 \ubc29\ubc95\uc740 ‘Skip-gram’\uacfc ‘Continuous Bag of Words’ (CBOW) \ubc29\uc2dd\uc73c\ub85c \uad6c\ud604\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uac01 \ub2e8\uc5b4\ub294 \uc8fc\ubcc0 \ub2e8\uc5b4\ub4e4\uc744 \ud1b5\ud574 \ud559\uc2b5\ub418\uba70, \uc758\ubbf8\ub860\uc801 \uac70\ub9ac\ub3c4 \uc720\uc9c0\ud569\ub2c8\ub2e4.\n <\/p>\n \n GloVe\ub294 \ub2e8\uc5b4\uc758 \ub3d9\uc2dc \ubc1c\uc0dd \ud589\ub82c\uc744 \ubd84\ud574\ud558\uc5ec \uc758\ubbf8\uc801 \uc720\uc0ac\uc131\uc744 \ud559\uc2b5\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4. \ubb38\ub9e5 \uc815\ubcf4\ub97c \uae30\ubc18\uc73c\ub85c \uc804 \uc138\uacc4 \ub2e8\uc5b4\uc758 \ud1b5\uacc4\ub97c \ud1b5\ud574 \uc784\ubca0\ub529\uc744 \ubcc0\ud654\uc2dc\ud0b5\ub2c8\ub2e4.\n <\/p>\n \n \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\uc5d0\uc11c \uc81c\uacf5\ud558\ub294 \uc784\ubca0\ub529 \ub808\uc774\uc5b4\ub97c \uc0ac\uc6a9\ud558\uba74 \uc9c1\uc811\uc801\uc73c\ub85c \ub2e8\uc5b4\ub97c \uc800\ucc28\uc6d0 \ubca1\ud130\ub85c \ubcc0\ud658\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc2e4\uc2dc\uac04\uc73c\ub85c \ub370\uc774\ud130\ub97c \ud559\uc2b5\ud558\uba74\uc11c \uc758\ubbf8\uac00 \uc798 \ubc18\uc601\ub41c \ubca1\ud130\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4.\n <\/p>\n \n \uc774\uc81c \ud30c\uc774\ud1a0\uce58(PyTorch)\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc784\ubca0\ub529\uc744 \uc2e4\uc81c\ub85c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uba3c\uc800, \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uac00\uc838\uc624\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n \n Penn Treebank \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uac04\ub2e8\ud55c \uc608\uc81c\ub97c \ub9cc\ub4e4\uc5b4 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \ub370\uc774\ud130\uc14b\uc740 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc5d0\uc11c \ub110\ub9ac \uc0ac\uc6a9\ub429\ub2c8\ub2e4.\n <\/p>\n \n \uc784\ubca0\ub529 \ub808\uc774\uc5b4\ub97c \ud3ec\ud568\ud55c \uac04\ub2e8\ud55c \uc2e0\uacbd\ub9dd \ubaa8\ub378\uc744 \ub9cc\ub4e4\uc5b4 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n \n \uc774\uc81c \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ucf1c \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc190\uc2e4 \ud568\uc218\uc640 \uc635\ud2f0\ub9c8\uc774\uc800\ub97c \uc815\uc758\ud558\uace0 \ud6c8\ub828 \ub8e8\ud504\ub97c \uc791\uc131\ud569\ub2c8\ub2e4.\n <\/p>\n \n \uc784\ubca0\ub529\uc774 \uc798 \ud559\uc2b5\ub418\uc5c8\ub294\uc9c0 \ud655\uc778\ud558\uae30 \uc704\ud574, \ud6c4\ucc98\ub9ac \uacfc\uc815\uc744 \ud1b5\ud574 \ud2b9\uc815 \ub2e8\uc5b4\ub4e4\uc758 \uc784\ubca0\ub529 \ubca1\ud130\ub97c \uc2dc\uac01\ud654\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n \n \uc624\ub298\uc740 \ub525\ub7ec\ub2dd\uacfc \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc790\uc5f0\uc5b4 \ucc98\ub9ac\ub97c \uc704\ud55c \uc784\ubca0\ub529\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \uae30\ubcf8\uc801\uc778 \uc784\ubca0\ub529 \uac1c\ub150\ubd80\ud130 \ub370\uc774\ud130\uc14b \uc900\ube44, \ubaa8\ub378 \uc815\uc758, \ud6c8\ub828, \uc2dc\uac01\ud654\uae4c\uc9c0\uc758 \uc804 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc784\ubca0\ub529\uc740 NLP\uc758 \uc911\uc694\ud55c \uae30\ucd08\uae30\uc220\uc774\uba70, \ub2e4\uc591\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\ub294 \ub370 \uc720\uc6a9\ud558\uac8c \uc0ac\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc2e4\ubb34\uc5d0 \uc801\uc6a9\ud558\uae30 \uc704\ud574 \uc5ec\ub7ec \uac00\uc9c0 \uae30\ubc95\uc744 \uc5f0\uad6c\ud558\ub294 \uac83\uc774 \uc88b\uc2b5\ub2c8\ub2e4.\n <\/p>\n1. \uc784\ubca0\ub529\uc774\ub780?<\/h2>\n
2. \uc784\ubca0\ub529\uc758 \uc885\ub958<\/h2>\n
\n
2.1 \uc6d0-\ud56b \uc778\ucf54\ub529<\/h3>\n
2.2 \uc6cc\ub4dc \ud22c \ubca1\ud130(Word2Vec)<\/h3>\n
2.3 \uae00\ub85c\ubca0(GloVe)<\/h3>\n
2.4 \uc784\ubca0\ub529 \ub808\uc774\uc5b4(Embeddings Layer)<\/h3>\n
3. \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud55c \uc784\ubca0\ub529<\/h2>\n
python\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torchtext.datasets import PennTreebank\nfrom torchtext.data import Field, TabularDataset, BucketIterator\nimport numpy as np\nimport random\nimport spacy\nnlp = spacy.load('en_core_web_sm')\n <\/code><\/pre>\n3.1 \ub370\uc774\ud130 \uc900\ube44<\/h3>\n
python\nTEXT = Field(tokenize='spacy', lower=True)\ntrain_data, valid_data, test_data = PennTreebank.splits(TEXT)\n\nTEXT.build_vocab(train_data, max_size=10000, min_freq=2)\nvocab_size = len(TEXT.vocab)\n <\/code><\/pre>\n3.2 \uc784\ubca0\ub529 \ubaa8\ub378 \uc815\uc758<\/h3>\n
python\nclass EmbeddingModel(nn.Module):\n def __init__(self, vocab_size, embedding_dim):\n super(EmbeddingModel, self).__init__()\n self.embedding = nn.Embedding(vocab_size, embedding_dim)\n self.fc = nn.Linear(embedding_dim, vocab_size)\n\n def forward(self, x):\n embedded = self.embedding(x)\n return self.fc(embedded)\n <\/code><\/pre>\n3.3 \ubaa8\ub378 \ud6c8\ub828<\/h3>\n
python\ndef train(model, iterator, optimizer, criterion):\n model.train()\n epoch_loss = 0\n\n for batch in iterator:\n optimizer.zero_grad()\n output = model(batch.text)\n loss = criterion(output.view(-1, vocab_size), batch.target.view(-1))\n loss.backward()\n optimizer.step()\n epoch_loss += loss.item()\n\n return epoch_loss \/ len(iterator)\n\nembedding_dim = 100\nmodel = EmbeddingModel(vocab_size, embedding_dim)\noptimizer = optim.Adam(model.parameters())\ncriterion = nn.CrossEntropyLoss()\n\n# Iterators\ntrain_iterator, valid_iterator, test_iterator = BucketIterator.splits(\n (train_data, valid_data, test_data), \n batch_size=64,\n device=device\n)\n\n# Training\nfor epoch in range(10):\n train_loss = train(model, train_iterator, optimizer, criterion)\n print(f'Epoch {epoch + 1}, Train Loss: {train_loss:.3f}')\n <\/code><\/pre>\n4. \ub2e8\uc5b4 \uc784\ubca0\ub529 \uc2dc\uac01\ud654<\/h2>\n
python\ndef visualize_embeddings(model, word):\n embedding_matrix = model.embedding.weight.data.numpy()\n word_index = TEXT.vocab.stoi[word]\n word_embedding = embedding_matrix[word_index]\n\n # \uc720\uc0ac\ud55c \ub2e8\uc5b4 \ud0d0\uc0c9\n similarities = np.dot(embedding_matrix, word_embedding)\n similar_indices = np.argsort(similarities)[-10:]\n similar_words = [TEXT.vocab.itos[idx] for idx in similar_indices]\n \n return similar_words\n\nprint(visualize_embeddings(model, 'apple'))\n <\/code><\/pre>\n5. \uacb0\ub860<\/h2>\n
6. \ucc38\uace0 \uc790\ub8cc<\/h2>\n
\n