{"id":3081,"date":"2025-12-11T14:29:27","date_gmt":"2025-12-11T06:29:27","guid":{"rendered":"https:\/\/www.tiptinker.com\/bert-vs-gpt-the-ultimate-guide-to-encoder-and-decoder-models\/"},"modified":"2025-12-11T14:45:56","modified_gmt":"2025-12-11T06:45:56","slug":"bert-vs-gpt-the-ultimate-guide-to-encoder-and-decoder-models","status":"publish","type":"post","link":"https:\/\/www.tiptinker.com\/ja\/bert-vs-gpt-the-ultimate-guide-to-encoder-and-decoder-models\/","title":{"rendered":"BERT vs. GPT\uff1a\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u30e2\u30c7\u30eb\u3068\u30c7\u30b3\u30fc\u30c0\u30fc\u30e2\u30c7\u30eb\u5b8c\u5168\u30ac\u30a4\u30c9"},"content":{"rendered":"<p>AI\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u3092\u69cb\u7bc9\u3059\u308b\u969b\u3001BERT\u3092\u9078\u3076\u304bGPT\u3092\u9078\u3076\u304b\u306f\u3001\u5358\u306a\u308b\u597d\u307f\u306e\u554f\u984c\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u305d\u308c\u306f\u3001\u3042\u306a\u305f\u306e\u30e2\u30c7\u30eb\u306b\u300c\u8aad\u307e\u305b\u308b\u300d<strong>\u5fc5\u8981\u304c\u3042\u308b\u306e\u304b\u3001\u305d\u308c\u3068\u3082<\/strong>\u300c\u66f8\u304b\u305b\u308b\u300d\u5fc5\u8981\u304c\u3042\u308b\u306e\u304b\u3068\u3044\u3046\u3001\u69cb\u9020\u7684\u306a\u6c7a\u5b9a\u4e8b\u9805\u306a\u306e\u3067\u3059\u3002<\/p>\n<p>\u3069\u3061\u3089\u30822018\u5e74\u306bGoogle\u306b\u3088\u3063\u3066\u767a\u8868\u3055\u308c\u305f\u9769\u65b0\u7684\u306a<strong>Transformer\uff08\u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc\uff09<\/strong>\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u7531\u6765\u3057\u307e\u3059\u304c\u3001\u6839\u672c\u7684\u306b\u7570\u306a\u308b\u554f\u984c\u3092\u89e3\u6c7a\u3059\u308b\u305f\u3081\u306b\u3001\u30a8\u30f3\u30b8\u30f3\u306e\u7570\u306a\u308b\u90e8\u5206\u3092\u4f7f\u7528\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u306e\u9055\u3044\u3092\u8aa4\u89e3\u3059\u308b\u3068\u3001\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u306e\u4f4e\u4e0b\u3084\u8a08\u7b97\u30ea\u30bd\u30fc\u30b9\u306e\u7121\u99c4\u9063\u3044\u306b\u3064\u306a\u304c\u308a\u307e\u3059\u3002<\/p>\n<p>\u672c\u30ac\u30a4\u30c9\u3067\u306f\u3001\u4e21\u8005\u306e\u4ed5\u7d44\u307f\u3001\u30e6\u30fc\u30b9\u30b1\u30fc\u30b9\u3001\u305d\u3057\u3066\u30b3\u30fc\u30c9\u5b9f\u88c5\u306b\u3064\u3044\u3066\u8a73\u3057\u304f\u89e3\u8aac\u3057\u307e\u3059\u3002<\/p>\n<h2>\u4e2d\u6838\u6982\u5ff5\uff1aTransformer\u306e\u5206\u96e2<\/h2>\n<p>\u30aa\u30ea\u30b8\u30ca\u30eb\u306eTransformer\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306f\u30012\u3064\u306e\u30b9\u30bf\u30c3\u30af\uff08\u7a4d\u307f\u91cd\u306d\uff09\u3067\u69cb\u6210\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n<ol>\n<li><strong>\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc (The Encoder):<\/strong> \u5165\u529b\u3092\u51e6\u7406\u3057\u307e\u3059\u3002\u6587\u8108\u3092<strong>\u7406\u89e3<\/strong>\u3059\u308b\u3088\u3046\u306b\u8a2d\u8a08\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/li>\n<li><strong>\u30c7\u30b3\u30fc\u30c0\u30fc (The Decoder):<\/strong> \u51fa\u529b\u3092\u751f\u6210\u3057\u307e\u3059\u3002\u6b21\u306e\u30b9\u30c6\u30c3\u30d7\u3092<strong>\u4e88\u6e2c<\/strong>\u3059\u308b\u3088\u3046\u306b\u8a2d\u8a08\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/li>\n<\/ol>\n<p>\u73fe\u4ee3\u306e\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb\uff08LLM\uff09\u306f\u901a\u5e38\u3001\u3053\u306e\u65b9\u7a0b\u5f0f\u306e\u3069\u3061\u3089\u304b\u534a\u5206\u306b\u7279\u5316\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<h3>BERT (Bidirectional Encoder Representations from Transformers)<\/h3>\n<p>BERT\u306f<strong>\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u306e\u307f\uff08Encoder-only\uff09<\/strong>\u306e\u30e2\u30c7\u30eb\u3067\u3059\u3002<strong>\u53cc\u65b9\u5411\uff08Bi-directional\uff09<\/strong>\u3067\u3042\u308a\u3001\u5358\u8a9e\u306e\u6587\u8108\u3092\u5de6\u3068\u53f3\u306e\u4e21\u65b9\u304b\u3089\u540c\u6642\u306b\u898b\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n<ul>\n<li><strong>\u4f8b\u3048\uff1a<\/strong> \u53e4\u6587\u66f8\u3092\u8aad\u3080\u5b66\u8005\u3002\u524d\u5f8c\u306e\u6587\u8108\u304b\u3089\u66d6\u6627\u306a\u5358\u8a9e\u306e\u610f\u5473\u3092\u7406\u89e3\u3059\u308b\u305f\u3081\u306b\u3001\u6587\u5168\u4f53\u3092\u540c\u6642\u306b\u898b\u6e21\u3057\u307e\u3059\u3002<\/li>\n<li><strong>\u5f37\u307f\uff08\u30b9\u30fc\u30d1\u30fc\u30d1\u30ef\u30fc\uff09\uff1a<\/strong> \u7406\u89e3\u3001\u5206\u985e\u3001\u691c\u7d22\u3002<\/li>\n<\/ul>\n<h3>GPT (Generative Pre-trained Transformer)<\/h3>\n<p>GPT\u306f<strong>\u30c7\u30b3\u30fc\u30c0\u30fc\u306e\u307f\uff08Decoder-only\uff09<\/strong>\u306e\u30e2\u30c7\u30eb\u3067\u3059\u3002<strong>\u81ea\u5df1\u56de\u5e30\uff08Auto-regressive\uff09<\/strong>\u3001\u3064\u307e\u308a\u5358\u65b9\u5411\u3067\u3042\u308a\u3001\u5de6\u304b\u3089\u53f3\u3078\u3068\u8aad\u307f\u9032\u3081\u307e\u3059\u3002\u300c\u672a\u6765\u300d\u306e\u30c8\u30fc\u30af\u30f3\uff08\u5358\u8a9e\uff09\u3092\u898b\u308b\u3053\u3068\u306f\u3067\u304d\u307e\u305b\u3093\u3002\u904e\u53bb\u306e\u5c65\u6b74\u306e\u307f\u306b\u57fa\u3065\u3044\u3066\u3001\u6b21\u306e\u5358\u8a9e\u3092\u4e88\u6e2c\u3057\u307e\u3059\u3002<\/p>\n<ul>\n<li><strong>\u4f8b\u3048\uff1a<\/strong> \u6f14\u8aac\u306e\u539f\u7a3f\u3092\u30e9\u30a4\u30d6\u3067\u66f8\u3044\u3066\u3044\u308b\u30b9\u30d4\u30fc\u30c1\u30e9\u30a4\u30bf\u30fc\u3002\u4e00\u8cab\u6027\u3092\u4fdd\u3064\u305f\u3081\u306b\u3001\u76f4\u524d\u306e\u5358\u8a9e\u306b\u7d9a\u3044\u3066\u3069\u306e\u5358\u8a9e\u304c\u6700\u3082\u81ea\u7136\u306b\u6d41\u308c\u308b\u304b\u3068\u3044\u3046\u70b9\u306b\u5168\u795e\u7d4c\u3092\u96c6\u4e2d\u3055\u305b\u307e\u3059\u3002<\/li>\n<li><strong>\u5f37\u307f\uff08\u30b9\u30fc\u30d1\u30fc\u30d1\u30ef\u30fc\uff09\uff1a<\/strong> \u751f\u6210\u3001\u30c1\u30e3\u30c3\u30c8\u3001\u30c6\u30ad\u30b9\u30c8\u88dc\u5b8c\u3002<\/li>\n<\/ul>\n<h2>\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306e\u53ef\u8996\u5316<\/h2>\n<p>\u9055\u3044\u306f\u3001Attention\uff08\u6ce8\u610f\u6a5f\u69cb\uff09\u3092\u901a\u3058\u3066\u60c5\u5831\u304c\u3069\u306e\u3088\u3046\u306b\u6d41\u308c\u308b\u304b\u306b\u3042\u308a\u307e\u3059\u3002<\/p>\n<div class=\"easy-mermaid-wrapper\">\n<pre><code class=\"language-mermaid\">graph TD\r\n    subgraph \"BERT (\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc)\"\r\n    A[\"\u5165\u529b: 'The bank of the river'\"] --&gt; B[\"Self-Attention (\u53cc\u65b9\u5411)\"]\r\n    B --&gt; C[\"'bank' \u306f 'The', 'of', 'river' \u3092\u53c2\u7167\u3059\u308b\"]\r\n    C --&gt; D[\"\u51fa\u529b: \u6587\u8108\u5316\u3055\u308c\u305f\u57cb\u3081\u8fbc\u307f\u8868\u73fe\"]\r\n    end\r\n\r\n    subgraph \"GPT (\u30c7\u30b3\u30fc\u30c0\u30fc)\"\r\n    E[\"\u5165\u529b: 'The bank of'\"] --&gt; F[\"Masked Self-Attention (\u5358\u65b9\u5411)\"]\r\n    F --&gt; G[\"'of' \u306f 'The', 'bank' \u306e\u307f\u3092\u53c2\u7167\u3059\u308b\"]\r\n    G --&gt; H[\"\u51fa\u529b: \u4e88\u6e2c 'the'\"]\r\n    end\r\n    \r\n    style A fill:#e1f5fe,stroke:#01579b\r\n    style E fill:#fff3e0,stroke:#e65100\r\n<\/code><\/pre>\n<\/div>\n<h2>\u6a5f\u80fd\u6bd4\u8f03\u30de\u30c8\u30ea\u30c3\u30af\u30b9<\/h2>\n<table>\n<thead>\n<tr>\n<th align=\"left\">\u6a5f\u80fd<\/th>\n<th align=\"left\">BERT (\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc)<\/th>\n<th align=\"left\">GPT (\u30c7\u30b3\u30fc\u30c0\u30fc)<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td align=\"left\"><strong>\u65b9\u5411\u6027<\/strong><\/td>\n<td align=\"left\">\u53cc\u65b9\u5411 (\u5de6 &lt;-&gt; \u53f3)<\/td>\n<td align=\"left\">\u5358\u65b9\u5411 (\u5de6 -&gt; \u53f3)<\/td>\n<\/tr>\n<tr>\n<td align=\"left\"><strong>\u4e3b\u306a\u30bf\u30b9\u30af<\/strong><\/td>\n<td align=\"left\">\u7406\u89e3 \/ \u8b58\u5225<\/td>\n<td align=\"left\">\u751f\u6210<\/td>\n<\/tr>\n<tr>\n<td align=\"left\"><strong>\u4e8b\u524d\u5b66\u7fd2\u306e\u76ee\u6a19<\/strong><\/td>\n<td align=\"left\">\u30de\u30b9\u30af\u5316\u8a00\u8a9e\u30e2\u30c7\u30ea\u30f3\u30b0 (\u7a74\u57cb\u3081\u554f\u984c)<\/td>\n<td align=\"left\">\u56e0\u679c\u7684\u8a00\u8a9e\u30e2\u30c7\u30ea\u30f3\u30b0 (\u6b21\u306e\u30c8\u30fc\u30af\u30f3\u4e88\u6e2c)<\/td>\n<\/tr>\n<tr>\n<td align=\"left\"><strong>\u6700\u9069\u306a\u30e6\u30fc\u30b9\u30b1\u30fc\u30b9<\/strong><\/td>\n<td align=\"left\">\u611f\u60c5\u5206\u6790\u3001\u56fa\u6709\u8868\u73fe\u62bd\u51fa(NER)\u3001\u30b9\u30d1\u30e0\u691c\u51fa\u3001\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u691c\u7d22<\/td>\n<td align=\"left\">\u30c1\u30e3\u30c3\u30c8\u30dc\u30c3\u30c8\u3001\u30b3\u30fc\u30c9\u751f\u6210\u3001\u30b9\u30c8\u30fc\u30ea\u30fc\u4f5c\u6210\u3001\u8981\u7d04<\/td>\n<\/tr>\n<tr>\n<td align=\"left\"><strong>\u5165\u529b\u5236\u9650<\/strong><\/td>\n<td align=\"left\">\u56fa\u5b9a (\u901a\u5e38 512 \u30c8\u30fc\u30af\u30f3)<\/td>\n<td align=\"left\">\u67d4\u8edf (\u53ef\u5909\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u30a6\u30a3\u30f3\u30c9\u30a6)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>\u5b9f\u88c5\uff1a\u30b3\u30fc\u30c9\u3067\u898b\u308b\u9055\u3044<\/h2>\n<p>\u5b9f\u969b\u306e\u52d5\u4f5c\u306e\u9055\u3044\u3092\u78ba\u8a8d\u3059\u308b\u305f\u3081\u306b\u3001<a href=\"https:\/\/huggingface.co\/docs\/transformers\/index\">Hugging Face<\/a>\u306ePython\u30e9\u30a4\u30d6\u30e9\u30ea <code>transformers<\/code> \u3092\u4f7f\u7528\u3057\u3066\u307f\u307e\u3057\u3087\u3046\u3002<\/p>\n<h3>1. BERT\u306b\u3088\u308b\u7406\u89e3\uff08\u7279\u5fb4\u91cf\u62bd\u51fa\uff09<\/h3>\n<p>BERT\u3092\u4f7f\u7528\u3057\u3066\u3001\u30c6\u30ad\u30b9\u30c8\u3092\u305d\u306e\u610f\u5473\u3092\u8868\u3059\u30d9\u30af\u30c8\u30eb\uff08\u6570\u5024\uff09\u306b\u5909\u63db\u3057\u307e\u3059\u3002\u3053\u3053\u3067\u306f\u3001\u30e2\u30c7\u30eb\u306b\u300c\u8a71\u3059\u300d\u3053\u3068\u306f\u6c42\u3081\u3066\u3044\u306a\u3044\u70b9\u306b\u6ce8\u76ee\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n<pre><code class=\"language-python\">from transformers import BertTokenizer, BertModel\r\nimport torch\r\n\r\n# 1. BERT\uff08\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\uff09\u306e\u521d\u671f\u5316\r\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\nmodel = BertModel.from_pretrained('bert-base-uncased')\r\n\r\ntext = \"The bank of the river.\"\r\ninputs = tokenizer(text, return_tensors=\"pt\")\r\n\r\n# 2. \u96a0\u308c\u72b6\u614b\uff08Hidden States\uff09\u3092\u53d6\u5f97\u3059\u308b\u305f\u3081\u306e\u30d5\u30a9\u30ef\u30fc\u30c9\u30d1\u30b9\r\nwith torch.no_grad():\r\n    outputs = model(**inputs)\r\n\r\n# 'last_hidden_state' \u306b\u306f\u3001\u5404\u30c8\u30fc\u30af\u30f3\u306e\u6587\u8108\u5316\u3055\u308c\u305f\u57cb\u3081\u8fbc\u307f\u304c\u542b\u307e\u308c\u307e\u3059\r\n# \u5f62\u72b6: [\u30d0\u30c3\u30c1\u30b5\u30a4\u30ba, \u30b7\u30fc\u30b1\u30f3\u30b9\u9577, \u96a0\u308c\u5c64\u30b5\u30a4\u30ba]\r\nembeddings = outputs.last_hidden_state\r\n\r\nprint(f\"Vector Shape: {embeddings.shape}\")\r\n# \u51fa\u529b\u4f8b: torch.Size([1, 7, 768]) \r\n<\/code><\/pre>\n<h3>2. GPT\u306b\u3088\u308b\u751f\u6210<\/h3>\n<p>GPT\u3092\u4f7f\u7528\u3057\u3066\u30c6\u30ad\u30b9\u30c8\u3092\u751f\u6210\u3057\u307e\u3059\u3002\u30e2\u30c7\u30eb\u306f\u4e00\u5ea6\u306b1\u3064\u306e\u30c8\u30fc\u30af\u30f3\u3092\u4e88\u6e2c\u3059\u308b\u305f\u3081\u3001\u30eb\u30fc\u30d7\u51e6\u7406\uff08\u307e\u305f\u306f <code>generate<\/code> \u30e6\u30fc\u30c6\u30a3\u30ea\u30c6\u30a3\uff09\u304c\u5fc5\u8981\u3067\u3059\u3002<\/p>\n<pre><code class=\"language-python\">from transformers import GPT2Tokenizer, GPT2LMHeadModel\r\n\r\n# 1. GPT\uff08\u30c7\u30b3\u30fc\u30c0\u30fc\uff09\u306e\u521d\u671f\u5316\r\ntokenizer = GPT2Tokenizer.from_pretrained('gpt2')\r\nmodel = GPT2LMHeadModel.from_pretrained('gpt2')\r\n\r\nprompt = \"The future of AI is\"\r\ninputs = tokenizer(prompt, return_tensors=\"pt\")\r\n\r\n# 2. \u30c6\u30ad\u30b9\u30c8\u751f\u6210\uff08\u81ea\u5df1\u56de\u5e30\u30eb\u30fc\u30d7\u306f\u5185\u90e8\u3067\u51e6\u7406\u3055\u308c\u307e\u3059\uff09\r\noutput_sequences = model.generate(\r\n    input_ids=inputs['input_ids'],\r\n    max_length=20,\r\n    temperature=0.7,\r\n    num_return_sequences=1,\r\n    do_sample=True\r\n)\r\n\r\ngenerated_text = tokenizer.decode(output_sequences[0], skip_special_tokens=True)\r\nprint(f\"Generated: {generated_text}\")\r\n# \u51fa\u529b\u4f8b: Generated: The future of AI is likely to be shaped by the development of new technologies...\r\n<\/code><\/pre>\n<h2>\u30b9\u30c6\u30c3\u30d7\u30fb\u30d0\u30a4\u30fb\u30b9\u30c6\u30c3\u30d7\uff1a\u9078\u3073\u65b9<\/h2>\n<p>\u30d7\u30ed\u30b8\u30a7\u30af\u30c8\u306b\u6700\u9069\u306a\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u3092\u9078\u3076\u305f\u3081\u306b\u3001\u4ee5\u4e0b\u306e\u30ed\u30b8\u30c3\u30af\u30d5\u30ed\u30fc\u306b\u5f93\u3063\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n<ol>\n<li><strong>\u51fa\u529b\u3092\u5b9a\u7fa9\u3059\u308b:<\/strong>\n<ul>\n<li>\u51fa\u529b\u306f\u300c\u30e9\u30d9\u30eb\u300d\u3067\u3059\u304b\uff1f\uff08\u4f8b\uff1a\u300c\u30dd\u30b8\u30c6\u30a3\u30d6\u300d\u300c\u30b9\u30d1\u30e0\u300d\u300c\u30ab\u30c6\u30b4\u30eaA\u300d\uff09 -&gt; <strong>BERT\u3092\u4f7f\u7528<\/strong><\/li>\n<li>\u51fa\u529b\u306f\u300c\u6570\u5024\u300d\u3067\u3059\u304b\uff1f\uff08\u4f8b\uff1a\u30cb\u30e5\u30fc\u30b9\u306b\u57fa\u3065\u304f\u682a\u4fa1\u4e88\u6e2c\uff09 -&gt; <strong>BERT\u3092\u4f7f\u7528<\/strong><\/li>\n<li>\u51fa\u529b\u306f\u300c\u65b0\u3057\u3044\u30c6\u30ad\u30b9\u30c8\u300d\u3067\u3059\u304b\uff1f -&gt; <strong>GPT\u3092\u4f7f\u7528<\/strong><\/li>\n<\/ul>\n<\/li>\n<li><strong>\u6587\u8108\u306e\u8981\u4ef6\u3092\u8a55\u4fa1\u3059\u308b:<\/strong>\n<ul>\n<li>\u6587\u982d\u306e\u610f\u5473\u304c\u3001\u6587\u672b\u306e\u5185\u5bb9\u306b\u4f9d\u5b58\u3057\u3066\u3044\u307e\u3059\u304b\uff1f\uff08\u4f8b\uff1aDNA\u914d\u5217\u89e3\u6790\u3001\u8907\u96d1\u306a\u6cd5\u7684\u6761\u9805\u306e\u89e3\u91c8\u306a\u3069\uff09<\/li>\n<li><strong>\u30a2\u30af\u30b7\u30e7\u30f3:<\/strong> \u306f\u3044\u306e\u5834\u5408\u3001\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u30e2\u30c7\u30eb\u306e\u6301\u3064\u53cc\u65b9\u5411\u6027\u304c\u512a\u308c\u3066\u3044\u307e\u3059\u3002<\/li>\n<\/ul>\n<\/li>\n<li><strong>\u30ec\u30a4\u30c6\u30f3\u30b7\uff08\u9045\u5ef6\uff09\u3092\u8003\u616e\u3059\u308b:<\/strong>\n<ul>\n<li>\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u30e2\u30c7\u30eb\u306f\u5165\u529b\u3092\u4e00\u5ea6\u306e\u30d1\u30b9\u3067\u51e6\u7406\u3059\u308b\u305f\u3081\u3001\u5206\u985e\u30bf\u30b9\u30af\u306b\u304a\u3044\u3066\u306f\u4e00\u822c\u7684\u306b\u9ad8\u901f\u3067\u3059\u3002<\/li>\n<li>\u30c7\u30b3\u30fc\u30c0\u30fc\u30e2\u30c7\u30eb\u306f\u751f\u6210\u3055\u308c\u308b\u5358\u8a9e\u3054\u3068\u306b\u30e2\u30c7\u30eb\u3092\u9806\u6b21\u5b9f\u884c\u3059\u308b\u5fc5\u8981\u304c\u3042\u308b\u305f\u3081\u3001\u751f\u6210\u30bf\u30b9\u30af\u306b\u304a\u3044\u3066\u306f\u4f4e\u901f\u306b\u306a\u308a\u307e\u3059\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n<h2>\u6570\u5b66\u7684\u76f4\u611f<\/h2>\n<p>\u6838\u5fc3\u7684\u306a\u9055\u3044\u306f\u3001\u78ba\u7387\u8a08\u7b97\u306e\u65b9\u6cd5\u306b\u3042\u308a\u307e\u3059\u3002<\/p>\n<p><strong>GPT (\u81ea\u5df1\u56de\u5e30 \/ Auto-regressive):<\/strong><br \/>\n\u30b7\u30fc\u30b1\u30f3\u30b9 $W$ \u306e\u78ba\u7387\u306f\u3001\u6761\u4ef6\u4ed8\u304d\u78ba\u7387\u306e\u7a4d\u3067\u8868\u3055\u308c\u307e\u3059\u3002<\/p>\n<div class=\"easy-katex-wrapper easy-katex-block\" id=\"katex-1\" data-formula=\"P(W) = \\prod_{i=1}^{n} P(w_i | w_1, ..., w_{i-1})\" data-display=\"true\"><\/div>\n<p><strong>BERT (\u30de\u30b9\u30af\u5316\u30aa\u30fc\u30c8\u30a8\u30f3\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0 \/ Masked Auto-encoding):<\/strong><br \/>\nBERT\u306f\u3001\u30b7\u30fc\u30b1\u30f3\u30b9\u5185\u306e<em>\u305d\u306e\u4ed6\u3059\u3079\u3066\u306e<\/em>\u30c8\u30fc\u30af\u30f3\u3092\u4e0e\u3048\u3089\u308c\u305f\u72b6\u614b\u3067\u3001\u30de\u30b9\u30af\u3055\u308c\u305f\u30c8\u30fc\u30af\u30f3 $w_i$ \u3092\u4e88\u6e2c\u3057\u307e\u3059\u3002<\/p>\n<div class=\"easy-katex-wrapper easy-katex-block\" id=\"katex-2\" data-formula=\"P(w_i | w_1, ..., w_{i-1}, w_{i+1}, ..., w_n)\" data-display=\"true\"><\/div>\n<h2>\u30d1\u30ef\u30fc\u30e6\u30fc\u30b6\u30fc\u5411\u3051\u306e\u30d7\u30ed\u30fb\u30d2\u30f3\u30c8<\/h2>\n<ul>\n<li><strong>\u57cb\u3081\u8fbc\u307f\uff08Embedding\uff09\u306e\u54c1\u8cea:<\/strong> \u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u691c\u7d22\u3084\u30af\u30e9\u30b9\u30bf\u30ea\u30f3\u30b0\u306b\u3001\u751f\u306eGPT\u57cb\u3081\u8fbc\u307f\u3092\u4f7f\u7528<strong>\u3057\u306a\u3044<\/strong>\u3067\u304f\u3060\u3055\u3044\u3002GPT\u306f\u5de6\u5074\u3057\u304b\u898b\u306a\u3044\u305f\u3081\u3001\u6700\u5f8c\u306e\u5358\u8a9e\u306e\u57cb\u3081\u8fbc\u307f\u306f\u76f4\u8fd1\u306e\u6587\u8108\u3092\u904e\u5270\u306b\u8868\u73fe\u3057\u304c\u3061\u3067\u3059\u3002BERT\uff08\u307e\u305f\u306fS-BERT\uff09\u306e\u65b9\u304c\u3001\u5727\u5012\u7684\u306b\u512a\u308c\u305f\u6587\u57cb\u3081\u8fbc\u307f\u3092\u751f\u6210\u3057\u307e\u3059\u3002<\/li>\n<li><strong>\u300c\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u30fb\u30c7\u30b3\u30fc\u30c0\u30fc\u300d\u3068\u3044\u3046\u4e2d\u9593\u5730\u70b9:<\/strong> \u30c6\u30ad\u30b9\u30c8\u3092\u5165\u529b\u3057\u3066\u30c6\u30ad\u30b9\u30c8\u3092\u51fa\u529b\u3059\u308b\u5fc5\u8981\u304c\u3042\u308b\u5834\u5408\uff08\u7ffb\u8a33\u3084\u8981\u7d04\u306a\u3069\uff09\u3001<strong>T5<\/strong>\u3084<strong>BART<\/strong>\u3092\u4f7f\u7528\u3057\u3066\u304f\u3060\u3055\u3044\u3002\u3053\u308c\u3089\u306e\u30e2\u30c7\u30eb\u306f\u3001\u30bd\u30fc\u30b9\u30c6\u30ad\u30b9\u30c8\u3092\u8aad\u3080\u305f\u3081\u306e\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u3068\u3001\u7ffb\u8a33\u3092\u751f\u6210\u3059\u308b\u305f\u3081\u306e\u30c7\u30b3\u30fc\u30c0\u30fc\u3068\u3044\u3046\u3001<em>\u4e21\u65b9<\/em>\u306e\u30b9\u30bf\u30c3\u30af\u3092\u6d3b\u7528\u3057\u307e\u3059\u3002<\/li>\n<li><strong>\u30a4\u30f3\u30b9\u30c8\u30e9\u30af\u30b7\u30e7\u30f3\u30fb\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0:<\/strong> \u73fe\u4ee3\u306e\u300c\u30c1\u30e3\u30c3\u30c8\u300d\u30e2\u30c7\u30eb\uff08ChatGPT\u3084Llama 3\u306a\u3069\uff09\u306f\u3001\u6307\u793a\u3092\u7406\u89e3\u3057\u3066\u3044\u308b\u304b\u306e\u3088\u3046\u306b\u632f\u308b\u821e\u3046\u3088\u3046\u8abf\u6574\uff08Fine-tuned\uff09\u3055\u308c\u305f\u30c7\u30b3\u30fc\u30c0\u30fc\u5c02\u7528\u30e2\u30c7\u30eb\u3067\u3059\u3002\u3053\u308c\u3089\u306f\u30c7\u30b3\u30fc\u30c0\u30fc\u3067\u3059\u304c\u3001\u305d\u306e\u5727\u5012\u7684\u306a\u30b9\u30b1\u30fc\u30eb\u306b\u3088\u308a\u3001\u304b\u3064\u3066\u306f\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u306e\u9818\u57df\u3067\u3042\u3063\u305f\u63a8\u8ad6\u30bf\u30b9\u30af\u3082\u5b9f\u884c\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/li>\n<\/ul>\n<hr \/>\n<p><strong>\u8981\u7d04:<\/strong> \u30de\u30b7\u30f3\u306b\u5206\u6790\u3001\u5206\u985e\u3001\u307e\u305f\u306f\u691c\u7d22\u3092\u3055\u305b\u308b\u5fc5\u8981\u304c\u3042\u308b\u5834\u5408\u306f\u3001<strong>\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc (BERT)<\/strong> \u3092\u4f7f\u7528\u3057\u3066\u304f\u3060\u3055\u3044\u3002\u30de\u30b7\u30f3\u306b\u5275\u9020\u3001\u4f1a\u8a71\u3001\u307e\u305f\u306f\u6587\u7ae0\u306e\u62e1\u5f35\u3092\u3055\u305b\u308b\u5fc5\u8981\u304c\u3042\u308b\u5834\u5408\u306f\u3001<strong>\u30c7\u30b3\u30fc\u30c0\u30fc (GPT)<\/strong> \u3092\u4f7f\u7528\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>AI\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u3092\u69cb\u7bc9\u3059\u308b\u969b\u3001BERT\u3092\u9078\u3076\u304bGPT\u3092\u9078\u3076\u304b\u306f\u3001\u5358\u306a\u308b\u597d\u307f\u306e\u554f\u984c\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u305d\u308c\u306f\u3001\u3042\u306a\u305f\u306e\u30e2\u30c7\u30eb\u306b\u300c\u8aad\u307e\u305b\u308b\u300d\u5fc5\u8981\u304c\u3042\u308b\u306e\u304b\u3001\u305d\u308c\u3068\u3082\u300c\u66f8\u304b\u305b\u308b\u300d\u5fc5\u8981\u304c\u3042\u308b\u306e\u304b\u3068\u3044\u3046\u3001\u69cb\u9020\u7684\u306a\u6c7a\u5b9a\u4e8b\u9805\u306a\u306e\u3067\u3059\u3002 \u3069\u3061\u3089\u30822018\u5e74\u306bGoogle\u306b\u3088\u3063\u3066\u767a\u8868\u3055\u308c\u305f\u9769\u65b0\u7684\u306aTransformer\uff08\u30c8\u30e9\u30f3\u30b9\u30d5\u30a9\u30fc\u30de\u30fc\uff09\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306b\u7531\u6765\u3057\u307e\u3059\u304c\u3001\u6839\u672c\u7684\u306b\u7570\u306a\u308b\u554f\u984c\u3092\u89e3\u6c7a\u3059\u308b\u305f\u3081\u306b\u3001\u30a8\u30f3\u30b8\u30f3\u306e\u7570\u306a\u308b\u90e8\u5206\u3092\u4f7f\u7528\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u306e\u9055\u3044\u3092\u8aa4\u89e3\u3059\u308b\u3068\u3001\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u306e\u4f4e\u4e0b\u3084\u8a08\u7b97\u30ea\u30bd\u30fc\u30b9\u306e\u7121\u99c4\u9063\u3044\u306b\u3064\u306a\u304c\u308a\u307e\u3059\u3002 \u672c\u30ac\u30a4\u30c9\u3067\u306f\u3001\u4e21\u8005\u306e\u4ed5\u7d44\u307f\u3001\u30e6\u30fc\u30b9\u30b1\u30fc\u30b9\u3001\u305d\u3057\u3066\u30b3\u30fc\u30c9\u5b9f\u88c5\u306b\u3064\u3044\u3066\u8a73\u3057\u304f\u89e3\u8aac\u3057\u307e\u3059\u3002 \u4e2d\u6838\u6982\u5ff5\uff1aTransformer\u306e\u5206\u96e2 \u30aa\u30ea\u30b8\u30ca\u30eb\u306eTransformer\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306f\u30012\u3064\u306e\u30b9\u30bf\u30c3\u30af\uff08\u7a4d\u307f\u91cd\u306d\uff09\u3067\u69cb\u6210\u3055\u308c\u3066\u3044\u307e\u3059\u3002 \u30a8\u30f3\u30b3\u30fc\u30c0\u30fc (The Encoder): \u5165\u529b\u3092\u51e6\u7406\u3057\u307e\u3059\u3002\u6587\u8108\u3092\u7406\u89e3\u3059\u308b\u3088\u3046\u306b\u8a2d\u8a08\u3055\u308c\u3066\u3044\u307e\u3059\u3002 \u30c7\u30b3\u30fc\u30c0\u30fc (The Decoder): \u51fa\u529b\u3092\u751f\u6210\u3057\u307e\u3059\u3002\u6b21\u306e\u30b9\u30c6\u30c3\u30d7\u3092\u4e88\u6e2c\u3059\u308b\u3088\u3046\u306b\u8a2d\u8a08\u3055\u308c\u3066\u3044\u307e\u3059\u3002 \u73fe\u4ee3\u306e\u5927\u898f\u6a21\u8a00\u8a9e\u30e2\u30c7\u30eb\uff08LLM\uff09\u306f\u901a\u5e38\u3001\u3053\u306e\u65b9\u7a0b\u5f0f\u306e\u3069\u3061\u3089\u304b\u534a\u5206\u306b\u7279\u5316\u3057\u3066\u3044\u307e\u3059\u3002 BERT (Bidirectional Encoder Representations from Transformers) BERT\u306f\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc\u306e\u307f\uff08Encoder-only\uff09\u306e\u30e2\u30c7\u30eb\u3067\u3059\u3002\u53cc\u65b9\u5411\uff08Bi-directional\uff09\u3067\u3042\u308a\u3001\u5358\u8a9e\u306e\u6587\u8108\u3092\u5de6\u3068\u53f3\u306e\u4e21\u65b9\u304b\u3089\u540c\u6642\u306b\u898b\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002 \u4f8b\u3048\uff1a \u53e4\u6587\u66f8\u3092\u8aad\u3080\u5b66\u8005\u3002\u524d\u5f8c\u306e\u6587\u8108\u304b\u3089\u66d6\u6627\u306a\u5358\u8a9e\u306e\u610f\u5473\u3092\u7406\u89e3\u3059\u308b\u305f\u3081\u306b\u3001\u6587\u5168\u4f53\u3092\u540c\u6642\u306b\u898b\u6e21\u3057\u307e\u3059\u3002 \u5f37\u307f\uff08\u30b9\u30fc\u30d1\u30fc\u30d1\u30ef\u30fc\uff09\uff1a \u7406\u89e3\u3001\u5206\u985e\u3001\u691c\u7d22\u3002 GPT (Generative Pre-trained Transformer) GPT\u306f\u30c7\u30b3\u30fc\u30c0\u30fc\u306e\u307f\uff08Decoder-only\uff09\u306e\u30e2\u30c7\u30eb\u3067\u3059\u3002\u81ea\u5df1\u56de\u5e30\uff08Auto-regressive\uff09\u3001\u3064\u307e\u308a\u5358\u65b9\u5411\u3067\u3042\u308a\u3001\u5de6\u304b\u3089\u53f3\u3078\u3068\u8aad\u307f\u9032\u3081\u307e\u3059\u3002\u300c\u672a\u6765\u300d\u306e\u30c8\u30fc\u30af\u30f3\uff08\u5358\u8a9e\uff09\u3092\u898b\u308b\u3053\u3068\u306f\u3067\u304d\u307e\u305b\u3093\u3002\u904e\u53bb\u306e\u5c65\u6b74\u306e\u307f\u306b\u57fa\u3065\u3044\u3066\u3001\u6b21\u306e\u5358\u8a9e\u3092\u4e88\u6e2c\u3057\u307e\u3059\u3002 \u4f8b\u3048\uff1a \u6f14\u8aac\u306e\u539f\u7a3f\u3092\u30e9\u30a4\u30d6\u3067\u66f8\u3044\u3066\u3044\u308b\u30b9\u30d4\u30fc\u30c1\u30e9\u30a4\u30bf\u30fc\u3002\u4e00\u8cab\u6027\u3092\u4fdd\u3064\u305f\u3081\u306b\u3001\u76f4\u524d\u306e\u5358\u8a9e\u306b\u7d9a\u3044\u3066\u3069\u306e\u5358\u8a9e\u304c\u6700\u3082\u81ea\u7136\u306b\u6d41\u308c\u308b\u304b\u3068\u3044\u3046\u70b9\u306b\u5168\u795e\u7d4c\u3092\u96c6\u4e2d\u3055\u305b\u307e\u3059\u3002 \u5f37\u307f\uff08\u30b9\u30fc\u30d1\u30fc\u30d1\u30ef\u30fc\uff09\uff1a \u751f\u6210\u3001\u30c1\u30e3\u30c3\u30c8\u3001\u30c6\u30ad\u30b9\u30c8\u88dc\u5b8c\u3002 \u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3\u306e\u53ef\u8996\u5316 \u9055\u3044\u306f\u3001Attention\uff08\u6ce8\u610f\u6a5f\u69cb\uff09\u3092\u901a\u3058\u3066\u60c5\u5831\u304c\u3069\u306e\u3088\u3046\u306b\u6d41\u308c\u308b\u304b\u306b\u3042\u308a\u307e\u3059\u3002 graph TD subgraph &#8220;BERT (\u30a8\u30f3\u30b3\u30fc\u30c0\u30fc)&#8221; A[&#8220;\u5165\u529b: &#8216;The bank of the river'&#8221;] &#8211;&gt; B[&#8220;Self-Attention (\u53cc\u65b9\u5411)&#8221;] B &#8211;&gt; C[&#8220;&#8216;bank&#8217; \u306f &#8216;The&#8217;, [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":3077,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[56],"tags":[],"class_list":["post-3081","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tips-tutorials"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/posts\/3081","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/comments?post=3081"}],"version-history":[{"count":0,"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/posts\/3081\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/media\/3077"}],"wp:attachment":[{"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/media?parent=3081"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/categories?post=3081"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tiptinker.com\/ja\/wp-json\/wp\/v2\/tags?post=3081"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}