{"id":276,"date":"2020-08-25T17:00:56","date_gmt":"2020-08-25T08:00:56","guid":{"rendered":"https:\/\/staka.jp\/wordpress\/?p=276"},"modified":"2020-08-25T17:00:56","modified_gmt":"2020-08-25T08:00:56","slug":"%e7%bf%bb%e8%a8%b3%e3%82%a8%e3%83%b3%e3%82%b8%e3%83%b3%e3%81%ae%e6%a7%8b%e7%af%89%ef%bc%88marian-nmt%ef%bc%89","status":"publish","type":"post","link":"https:\/\/staka.jp\/wordpress\/?p=276","title":{"rendered":"\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u306e\u69cb\u7bc9\uff08Marian-NMT\uff09"},"content":{"rendered":"\n<p>\u590f\u4f11\u307f\u306e\u5bbf\u984c\uff08\uff1f\uff09\u3068\u3057\u3066<a href=\"https:\/\/marian-nmt.github.io\/\">Marian-NMT<\/a>\u3092\u4f7f\u3063\u3066\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u3092\u69cb\u7bc9\u3057\u3066\u307f\u305f\u3002\u69cb\u7bc9\u3057\u305f\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u306f\u300c<a href=\"https:\/\/devneko.jp\/demo\/\">\u82f1\u8a9e\u2192\u65e5\u672c\u8a9e\u7ffb\u8a33\uff08https:\/\/devneko.jp\/demo\/\uff09<\/a>\u300d\u304b\u3089\u8a66\u3059\u3053\u3068\u304c\u3067\u304d\u308b\u3002\u8a33\u3059\u3079\u304d\u5358\u8a9e\u3092\u98db\u3070\u3059\u306a\u3069\u6df1\u5c64\u5b66\u7fd2\u306a\u6a5f\u68b0\u7ffb\u8a33\u7279\u6709\u306e\u30df\u30b9\u3092\u3059\u308b\u3053\u3068\u3082\u591a\u3044\u304c\u3001\u500b\u4eba\u304c\u8a66\u4f5c\u3057\u305f\u3082\u306e\u3068\u3057\u3066\u306f\u76f8\u5fdc\u306e\u6027\u80fd\u306a\u6c17\u304c\u3059\u308b\u3002\u5272\u3068\u8a33\u305b\u3066\u3044\u3066\u6b63\u76f4\u9a5a\u3044\u305f\u3002\u305f\u3060\u3001\u5165\u529b\u6587\u306e\u9069\u5207\u306a\u5206\u5272\u306a\u3069\u524d\u51e6\u7406\u7684\u306a\u4e8b\u3092\u307b\u3068\u3093\u3069\u3084\u3063\u3066\u3044\u306a\u3044\u306e\u3067\u5909\u306a\u7d50\u679c\u306b\u306a\u308b\u3053\u3068\u3082\u591a\u3044\u3002<\/p>\n\n\n\n<p>\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306f\u300cWEB \u30af\u30ed\u30fc\u30eb + Universal Sentence Encoder\u3067\u53ce\u96c6\u3057\u305f\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u300d+\u300cFree(Creative Commons\u30e9\u30a4\u30bb\u30f3\u30b9\u306a\u3069)\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u300d\u3001\u4f7f\u7528\u3057\u305f\u624b\u6cd5\u306fTransformer + sentence piece\u3067\u3042\u308a\u3001\u30d0\u30ea\u30d0\u30ea\u306eDeep Learning\u3067\u73fe\u6642\u70b9\u3067\u3082\u672c\u683c\u7684\u3067\u3042\u308b\u3002\u305f\u3060\u3057\u3001\u74b0\u5883\u5236\u7d04\uff08\u3068\u3044\u3046\u304b\u6642\u9593\u5236\u7d04\uff06\u4e88\u7b97\u5236\u7d04\uff08\u5f8c\u8ff0\uff09\u304b\u3089Back Translation\u306f\u4f7f\u3048\u3066\u3044\u306a\u3044\u3002\uff09<\/p>\n\n\n\n<p>\u4e0a\u8a18\u30a8\u30f3\u30b8\u30f3\u3067\u306f300\u6587\u5b57\u4ee5\u5185\u306e\u82f1\u8a9e\u6587\uff08\u8907\u6570\u6587\u306e\u5834\u5408\u6027\u80fd\u306f\u843d\u3061\u308b\uff09 [1] \u3092\u65e5\u672c\u8a9e\u306b\u7ffb\u8a33\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u3002\u8a33\u629c\u3051\u3092\u9632\u6b62\u3059\u308b\u30e2\u30fc\u30c9\u3082\u3042\u308b\u304c\u3001\u30ab\u30f3\u30de\u3084\u8a18\u53f7\u3067\u6587\u3092\u5206\u3051\u3066\u3044\u308b\u3060\u3051\u306a\u306e\u3067\u8a33\u629c\u3051\u9632\u6b62\u30e2\u30fc\u30c9\u306e\u6027\u80fd\u306f\u3042\u307e\u308a\u3088\u308d\u3057\u304f\u306a\u3044\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u3092\u4f5c\u3063\u305f\u7406\u7531<\/h3>\n\n\n\n<p>\u672c\u5f53\u306f\u65e5\u672c\u8a9e\u3067GPT-2\u8fba\u308a\u306efine tuning\u3092\u8a66\u305d\u3046\u3068\u601d\u3063\u3066\u3044\u3066\u300c<a href=\"https:\/\/medium.com\/@pierre_guillou\/faster-than-training-from-scratch-fine-tuning-the-english-gpt-2-in-any-language-with-hugging-f2ec05c98787\">Faster than training from scratch \u2014 Fine-tuning the English GPT-2 in any language with Hugging Face and fastai v2 (practical case with Portuguese)<\/a>\u300d\u3068\u3044\u3046\u7d20\u6674\u3089\u3057\u3044\u8a18\u4e8b\u3092\u8aad\u3093\u3067\u3044\u305f\u3002\u305d\u306e\u4e2d\u306b\u300cFor example, to obtain a Portuguese GPT-2, we could download from the Transformers library of Hugging Face the OpenAI GPT-2 pre-trained in English and the MarianMT translator\u300d\u3068\u3044\u3046\u8a18\u8f09\u304c\u3042\u3063\u305f\u3082\u306e\u306e\u3001\u6b8b\u5ff5\u306a\u304c\u3089\u300c\u65e5\u672c\u8a9e\u2192\u82f1\u8a9e\u300d\u306e\u30e2\u30c7\u30eb\u306f\u516c\u958b\u3055\u308c\u3066\u3044\u308b\u304c\u300c\u82f1\u8a9e\u2192\u65e5\u672c\u8a9e\u300d \u306e\u30e2\u30c7\u30eb\u306f\u516c\u958b\u3055\u308c\u3066\u3044\u306a\u304b\u3063\u305f[2]\u3002<\/p>\n\n\n\n<p>\u81ea\u7531\u306b\u4f7f\u3048\u308b\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u306f\u5f79\u306b\u7acb\u3061\u305d\u3046[3]\u306a\u306e\u3067\u81ea\u5206\u3067\u69cb\u7bc9\u3059\u308b\u3053\u3068\u306b\u3057\u305f\u3002\u8eca\u8f2a\u306e\u518d\u767a\u660e\u3067\u306f\u3042\u308b\u304c\u3001\u8272\u3005\u3068\u826f\u3044\u7d4c\u9a13\u306b\u306a\u3063\u305f\u3068\u601d\u3046\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u306e\u4f5c\u308a\u65b9<\/h3>\n\n\n\n<p>\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u3092\u4f5c\u308b\u306b\u306f\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3001\u5b66\u7fd2\u7528\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u3001\u5b66\u7fd2\u74b0\u5883\u304c\u5fc5\u8981\u3067\u3042\u308b\u3002\u4eca\u56de\u306f\u4e0b\u8a18\u3092\u7528\u3044\u305f\u3002<\/p>\n\n\n\n<ol class=\"wp-block-list\"><li>\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8: WEB\u3092\u30af\u30ed\u30fc\u30ea\u30f3\u30b0\u3057\u3066\u53ce\u96c6[4]\uff0bFree\u3067\u516c\u958b\u3055\u308c\u3066\u3044\u308b\u3082\u306e\u3092\u8ffd\u52a0 [5]<\/li><li>\u5b66\u7fd2\u7528\u306e\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2: <a href=\"https:\/\/marian-nmt.github.io\/\">Marian-NMT (transformer) <\/a>+ <a href=\"https:\/\/github.com\/google\/sentencepiece\">sentencepiece  <\/a>[6] <\/li><li>\u5b66\u7fd2\u74b0\u5883: AWS p3.2xlarge \u30a4\u30f3\u30b9\u30bf\u30f3\u30b9 [7]<\/li><\/ol>\n\n\n\n<!--more-->\n\n\n\n<h4 class=\"wp-block-heading\">\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u4f5c\u6210<\/h4>\n\n\n\n<p>\u4ee5\u4e0b\u306e\u30d5\u30ed\u30fc\u3067\u5bfe\u8a33\u30c7\u30fc\u30bf\u3092\u4f5c\u6210\u3057\u305f\u3002<a href=\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder\/3\">Universal Sentence Encoder<\/a>\u306f\u6975\u3081\u3066\u5f37\u529b\u3067\u3042\u308b\u3002<\/p>\n\n\n\n<ol class=\"wp-block-list\"><li>\u2460\u82f1\u8a9e\u306e\u539f\u6587\u304b\u3089\u4f5c\u6210\u3055\u308c\u305f\u3063\u307d\u3044\u65e5\u672c\u8a9e\u306e\u30da\u30fc\u30b8[8]\u3001\u2461\u5bfe\u8a33\u304c\u8f09\u3063\u3066\u3044\u305d\u3046\u306a\u65e5\u672c\u8a9e\u306e\u30da\u30fc\u30b8[9] \u3092\u30af\u30ed\u30fc\u30eb\u3059\u308b\u3002<\/li><li>\u2460\u306e\u539f\u6587\u30da\u30fc\u30b8\u3092\u62bd\u51fa\u3001\u53d6\u5f97\u3059\u308b\u3002<\/li><li>\u2460\u3068\u539f\u6587\u30da\u30fc\u30b8\u304b\u3089\u65e5\u672c\u8a9e\u6587\u30fb\u82f1\u8a9e\u6587\u3092\u62bd\u51fa\u3059\u308b\u3002<\/li><li>Universal Sentence Encoder\u3067embed\u3057\u305f\u30d9\u30af\u30c8\u30eb\u306ecos\u8ddd\u96e2\uff06\u6587\u306e\u4f4d\u7f6e\u304c\u4e00\u5b9a\u6c34\u6e96\u3067\u8fd1\u3044\u6587\u3092\u30da\u30a2\u306b\u3059\u308b\u30022\u6587\u7d44\u307f\u5408\u308f\u305b\u305f\u307b\u3046\u304c\u826f\u3044\u5834\u5408\u3082\u3042\u308b\u305f\u3081\u30012\u6587\u305a\u3064\u3092\u30da\u30a2\u306b\u3057\u305fembed\u3082\u884c\u3046\u3002\uff081\u6587\u306e\u82f1\u8a9e\u30922\u6587\u306b\u5206\u3051\u3066\u548c\u8a33\u3059\u308b\u4f8b\u3082\u591a\u3044\u305f\u3081\u3002\uff09<\/li><li>\u6587\u306e\u30da\u30a2\u306e\u3046\u3061\u6700\u3082\u8ddd\u96e2\u304c\u8fd1\u304f\u5f62\u614b\u7d20\u6570\u3082\u8fd1\u3044\u3082\u306e\u3092\u5bfe\u8a33\u3068\u8003\u3048\u308b\u3002\uff08\u672c\u5f53\u306f\u65e5\u672c\u8a9e\u2192\u82f1\u8a9e\u7ffb\u8a33\u3057\u305f\u3046\u3048\u3067\u5358\u8a9e\u6570\u3092\u6bd4\u8f03\u3057\u305f\u304b\u3063\u305f\u304c\u8a08\u7b97\u6642\u9593\u304c\u304b\u304b\u308b\u305f\u3081\u8ae6\u3081\u305f\uff09<\/li><li>\u2461\u306b\u3064\u3044\u3066\u30013.4.5.\u76f8\u5f53\u306e\u51e6\u7406\u3092\u884c\u3046\u3002<\/li><\/ol>\n\n\n\n<p>\u91cd\u8981\u306a\u51e6\u7406\u306f4.\u3067\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u5b9f\u884c\u3057\u305f\u3002\u6587\u306e\u4f4d\u7f6e\u306b\u6ce8\u76ee\u3059\u308b\u305f\u3081\u6700\u521d\u306boffset\u3092\u6c42\u3081\u3066\u3044\u308b\u3002\u3053\u306e\u8fba\u308a\u306f\u753b\u50cf\u306e\u30de\u30c3\u30c1\u30f3\u30b0\u306b\u8fd1\u3044\u3002\u300cTwitter\u306b\u6295\u7a3f\u300d\u306e\u3088\u3046\u306a\u8a18\u8f09\u304c\u3042\u308b\u3068offset\u304c\u5909\u306b\u306a\u308b\u306e\u3067\u6700\u521d\u306b\u30d5\u30a3\u30eb\u30bf\u30ea\u30f3\u30b0\u3059\u308b\u306a\u3069\u5de5\u592b\u3092\u3057\u3066\u3044\u308b\u3002\uff08\u5b8c\u5168\u306a\u30bd\u30fc\u30b9\u306f\u5f8c\u65e5\u516c\u958b\u3059\u308b\uff08\u304b\u3082\u3057\u308c\u306a\u3044\uff09\uff09<\/p>\n\n\n<pre class=\"wp-block-preformatted\">embed = hub.load(\"https:\/\/tfhub.dev\/google\/universal-sentence-encoder-multilingual\/3\")\n\ndef get_align(ja_sents, en_sents, threshold=0.4, decay=0.1):\n    ja_vec = embed(ja_sents)\n    en_vec = embed(en_sents)\n    offsets = []\n    for jidx, jv in enumerate(ja_vec):\n        max_sim, ofs = max([(cos_sim(jv, ev), eidx - jidx) for eidx, ev in enumerate(en_vec)])\n        if max_sim &gt; threshold:\n            offsets.append(ofs)\n    if len(offsets):\n        offset = sum(offsets) \/ len(offsets)\n    else:\n        offset = 0\n    ret = []\n    jidx = 0\n    while jidx &lt; len(ja_vec):\n        jv = ja_vec[jidx]\n        max_sim, eidx = max(\n            [((1.0 - min(abs(eidx - (jidx + offset)) * decay, 1.0)) * cos_sim(jv, ev), eidx) for eidx, ev in\n             enumerate(en_vec)])\n        if max_sim &gt; threshold:\n            ret.append((max_sim, ja_sents[jidx], en_sents[eidx]))\n        jidx += 1\n    return ret<\/pre>\n\n\n<p>\u3053\u308c\u3089\u306e\u51e6\u7406\u306b\u3088\u3063\u3066\u3001200\u4e07\u30da\u30a2\u7a0b\u5ea6\u304c\u53ce\u96c6\u3067\u304d\u305f\u3002\u4e0a\u8a18\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306b\u300cJapanese-English Subtitle Corpus [10] \u30924. 5.\u3068\u540c\u69d8\u306b\u30af\u30ea\u30fc\u30cb\u30f3\u30b0\u3057\u305f\u3082\u306e\u300d\u300c\u4eac\u90fd\u30d5\u30ea\u30fc\u7ffb\u8a33\u30bf\u30b9\u30af (KFTT) [11]\u306e\u30c7\u30fc\u30bf \u300d\u300cTanaka Corpus[12]\u300d\u3092\u52a0\u3048\u3001\u5b66\u7fd2\u30c7\u30fc\u30bf [13] \u306b\u306f\u5168400\u4e07\u30da\u30a2\uff08\u65e5\u672c\u8a9e420MB, \u82f1\u8a9e380MB\uff09\u3092\u7528\u3044\u305f\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u5b66\u7fd2\uff08Marian-NMT\uff09<\/h4>\n\n\n\n<p>Marian-NMT\u306f\u975e\u5e38\u306b\u826f\u3044\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u3067\u3042\u308a\u3001\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3068Example\u304c\u5145\u5b9f\u3057\u3066\u3044\u308b\u3002\u305d\u306e\u305f\u3081\u5b66\u7fd2\u306f\u30b5\u30f3\u30d7\u30eb\u306b\u6cbf\u3063\u3066\u884c\u3048\u3070\u554f\u984c\u306a\u3044\u3002<\/p>\n\n\n\n<ol class=\"wp-block-list\"><li><a href=\"https:\/\/marian-nmt.github.io\/docs\/\">Marian-NMT\u306e\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8<\/a>\u3001Installation\u306b\u305d\u3063\u3066\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3092\u884c\u3046\u3002\u3053\u3053\u3067\u3001sentencepiece\u3092\u4f7f\u3046\u305f\u3081\u306b\u300c-DUSE_SENTENCEPIECE=on\u300d\u3092\u8a2d\u5b9a\u3059\u308b\u5fc5\u8981\u304c\u3042\u308b\u3002\u307e\u305f\u3001CPU\u30d0\u30fc\u30b8\u30e7\u30f3\u3067\u30c6\u30b9\u30c8\u3059\u308b\u5834\u5408\u306f\u300c-DCOMPILE_CPU=on -DCOMPILE_CUDA=off -DUSE_SENTENCEPIECE=on -DCOMPILE_SERVER=on\u300d\u3068\u3044\u3046\u611f\u3058\u3067marian-server\u3092\u542b\u3081\u3066\u304a\u304f\u3068\u5f8c\u3005\u4fbf\u5229\u3067\u3042\u308b\u3002<\/li><li>\u30e2\u30c7\u30eb\u3092\u4f5c\u6210\u3059\u308b\u305f\u3081\u3001<a href=\"https:\/\/github.com\/marian-nmt\/marian-examples\">marian-example<\/a>\u3092\u53c2\u8003\u306b\u8af8\u8a2d\u5b9a\u3092\u884c\u3046\u3002\u4f5c\u308a\u305f\u3044\u30e2\u30c7\u30eb\u306b\u8fd1\u3044example\u3092\u30b3\u30d4\u30fc\u3057\u3001\u30c7\u30fc\u30bf\u3092\u914d\u7f6e\u3001run-me.sh\u3092\u4fee\u6b63\u3059\u308b\u306e\u304c\u697d\u3067\u3042\u308b\u3002\uff08\u4eca\u56de\u4f5c\u6210\u3057\u305f\uff09\u30e2\u30c7\u30eb\u306e\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u306f<a href=\"https:\/\/github.com\/Helsinki-NLP\/OPUS-MT-train\">Helsinki-NLP\/OPUS-MT-train<\/a> [14]\u3092\u53c2\u8003\u306b\u3057\u305f\u3002<\/li><li>.\/run-me.sh \u3057\u3066\u5f85\u3064\u3002<\/li><\/ol>\n\n\n\n<p>AWS p3.2xlarge \u30a4\u30f3\u30b9\u30bf\u30f3\u30b9 \u3092\u7528\u3044\u308b\u3068\u300c18,000 words \/ s\u300d\u7a0b\u5ea6\u306e\u901f\u5ea6\u3067\u5b66\u7fd2\u304c\u9032\u3080\u3002400\u4e07\u6587\u30da\u30a2\u3060\u30681 epoch 1\u6642\u9593\u7a0b\u5ea6\u3067\u3042\u308b\u3002\u300c\u82f1\u8a9e\u2192\u65e5\u672c\u8a9e\u300d\u30e2\u30c7\u30eb\u306e\u69cb\u7bc9\u5168\u4f53\u306b\u306f20\u6642\u9593\u7a0b\u5ea6\u304b\u304b\u3063\u305f\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u8cbb\u7528<\/h4>\n\n\n\n<p>\u4eca\u56de\u306e\u8a66\u884c\u306b\u304b\u304b\u3063\u305f\u5fc5\u8981\u306f\u30af\u30ed\u30fc\u30eb&amp;\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u4f5c\u6210\u306b100 USD\u3001\u30e2\u30c7\u30eb\u69cb\u7bc9\u306b240 USD\u3001\u8a08340 USD\uff08\u7d0436,000\u5186\uff09\u7a0b\u5ea6\u3067\u3042\u3063\u305f\u3002\u3002\u3002[15]<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u6027\u80fd\u8a55\u4fa1<\/h3>\n\n\n\n<p><a href=\"https:\/\/github.com\/mjpost\/sacrebleu\">sacrebleu[ja]<\/a>\u3067\u8a08\u6e2c\u3057\u3066\u307f\u305f\u3068\u3053\u308d\u3001out sample\u306eKFTT\uff08test\u30bb\u30c3\u30c8\uff09\u306b\u5bfe\u3057\u3066\u3001\u672c\u4f53:23.3\u3001KFTT train\u30bb\u30c3\u30c8\u3067fine tuning\u3059\u308b\u306831.0\u3067\u3042\u3063\u305f\u3002<\/p>\n\n\n\n<pre class=\"wp-block-preformatted\">\u3010fine tuning \u524d\u3011\nBLEU+case.mixed+numrefs.1+smooth.exp+tok.ja-mecab-0.996-IPA+version.1.4.13 = 23.3 59.0\/31.8\/18.5\/11.0 (BP = 0.938 ratio = 0.940 hyp_len = 24790 ref_len = 26370)\n\u3010fine tuning \u5f8c\u3011\nBLEU+case.mixed+numrefs.1+smooth.exp+tok.ja-mecab-0.996-IPA+version.1.4.13 = 31.0 65.9\/43.5\/30.6\/21.9 (BP = 0.834 ratio = 0.846 hyp_len = 22322 ref_len = 26370)<\/pre>\n\n\n\n<p>\u30b9\u30b3\u30a2\u306f\u305d\u3053\u307e\u3067\u9ad8\u3044\u3068\u306f\u8a00\u3048\u306a\u3044\u304c\u3001\u9069\u5f53\u306a\u82f1\u6587\u3092\u5165\u308c\u308b\u3068\u305d\u3053\u305d\u3053\u306e\u7cbe\u5ea6\u3067\u65e5\u672c\u8a9e\u6587\u304c\u8fd4\u3063\u3066\u304f\u308b\u30021\u9031\u9593\u3067\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u4f5c\u6210\uff5e\u5b66\u7fd2\u3001WEB\u5316\u307e\u3067\u5b9f\u65bd\u3057\u305f\u7d50\u679c\u3068\u3057\u3066\u306f\u60aa\u304f\u306a\u3044\u306e\u3067\u306f\u306a\u304b\u308d\u3046\u304b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u6240\u611f<\/h3>\n\n\n\n<p> \u79c1\u3068\u3057\u3066\u306f\u500b\u4eba\u3067\u305d\u3053\u305d\u3053\u306e\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u3092\u4f5c\u308c\u308b\u3053\u3068\u81ea\u4f53\u304c\u9a5a\u304d\u3067\u3042\u3063\u305f\u3002\u3053\u308c\u306f\u7814\u7a76\u6210\u679c\u3084\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u3092\u516c\u958b\u3057\u3066\u3044\u308b\u7814\u7a76\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u306e\u304a\u304b\u3052\u3067\u3042\u308b\u3002 \u7814\u7a76\u8005\u30fb\u30bd\u30d5\u30c8\u30a6\u30a7\u30a2\u958b\u767a\u8005\u306b\u611f\u8b1d\u3067\u3042\u308b\u3002<\/p>\n\n\n\n<p>\u307e\u305f\u3001\u6642\u9593\u5358\u4f4d\u3067\u4f7f\u7528\u53ef\u80fd\u306aAWS\u306a\u3069\u30af\u30e9\u30a6\u30c9\u74b0\u5883\u304c\u6574\u5099\u3055\u308c\u305f\u3053\u3068\u3082\u5927\u304d\u3044\u3002\u975e\u5e38\u306b\u826f\u3044\u6642\u4ee3\u306b\u306a\u3063\u305f\u3068\u601d\u3046\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u811a\u6ce8<\/h4>\n\n\n\n<p>[1] \u5236\u9650\u306fVPS\u7684\u306b\u91cd\u304f\u306a\u3044\u3088\u3046\u306b\u3066\u304d\u3068\u30fc\u306b\u6c7a\u3081\u3066\u3044\u308b\u3002\u305d\u306e\u3046\u3061\u30e2\u30c7\u30eb\u81ea\u4f53\u3092\u516c\u958b\u3057\u3088\u3046\u3068\u601d\u3063\u3066\u3044\u308b\u306e\u3067\u7121\u5236\u9650\u306b\u4f7f\u3044\u305f\u3044\u65b9\u306f\u305d\u308c\u3092\u304a\u5f85\u3061\u4e0b\u3055\u3044\u3002<br \/>[2] \u5f8c\u3067\u77e5\u3063\u305f\u304c\u300c<a href=\"http:\/\/www.kecl.ntt.co.jp\/icl\/lirg\/jparacrawl\/\">JParaCrawl<\/a>\u300d\u306f\u300c\u65e5\u672c\u8a9e\u21d4\u82f1\u8a9e\u300d\u3067\u69cb\u7bc9\u3057\u305f\u30e2\u30c7\u30eb\u3092\u516c\u958b\u3057\u3066\u3044\u308b\u3002\u975e\u5546\u7528\u3067\u3042\u308c\u3070\u5229\u7528\u53ef\u80fd\u3002<br \/>[3] \u4f8b\u3048\u3070\u81ea\u7136\u8a00\u8a9e\u51e6\u7406\u306b\u304a\u3051\u308b\u516c\u5e73\u6027\u306f\u91cd\u8981\u3067\u3042\u308a\u3001\u81ea\u7531\u306b\u89e3\u6790\u3067\u304d\u308b\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u304c\u3042\u308b\u3068\u7814\u7a76\u4e0a\u8272\u3005\u4fbf\u5229\u3067\u3042\u308b \u3002  \uff08[2]\u3067\u826f\u3044\u3058\u3083\u3093\u3068\u3044\u3046\u610f\u898b\u3082\u3042\u308b\u304c\u30fb\u30fb\u30fb\uff09 \u305d\u306e\u4ed6\u306b\u3082\u82f1\u8a9e\u306e\u8a00\u8a9e\u8cc7\u6e90\u3092\u65e5\u672c\u8a9e\u5316\u3057\u305f\u3044\u307f\u305f\u3044\u306a\u6642\u306b\u3082\u4fbf\u5229\u304b\u306a\u3068\u601d\u3046\u3002\uff08\u73fe\u72b6\u3060\u3068\u6027\u80fd\u304c\u8db3\u308a\u3066\u3044\u306a\u3044\u3068\u3044\u3046\u610f\u898b\u3082\u3042\u308b\u304c\u30fb\u30fb\u30fb\uff09<br \/>[4]  \u300c<a href=\"https:\/\/h-bank.nict.go.jp\/seminars\/download\/20190306\/taichikakimuma190306.pdf\">\u8457\u4f5c\u6a29\u6cd5\u6539\u6b63\u304cAI\u958b\u767a\u306b\u4e0e\u3048\u308b\u300c\u885d\u6483\u300d<\/a>\u300d\u306b\u3042\u308b\u901a\u308a\u3001\u65e5\u672c\u306e\u8457\u4f5c\u6a29\u6cd5\u3067\u306fWEB\u30af\u30ed\u30fc\u30ea\u30f3\u30b0\u3092\u884c\u3063\u3066\u30c7\u30fc\u30bf\u3092\u96c6\u3081\u308b\u884c\u70ba\u306f\uff08\u9650\u5ea6\u306f\u3042\u308b\u304c\uff09\u9055\u6cd5\u3067\u306f\u306a\u3044\u3002 <br \/>[5] \u8ffd\u52a0\u3057\u305f\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306f\u300c<a href=\"http:\/\/www.phontron.com\/japanese-translation-data.php?lang=ja\">\u65e5\u672c\u8a9e\u5bfe\u8a33\u30c7\u30fc\u30bf<\/a>\u300d\u306b\u8a18\u8f09\u306e\u3042\u3063\u305f\u3082\u306e\u306e\u3046\u3061\u3001\u6761\u4ef6\u304c\u975e\u5546\u7528\u306b\u9650\u3089\u308c\u3066\u3044\u306a\u3044\u3082\u306e\u3068\u3057\u305f\u3002<br \/>[6] \u3082\u3068\u3082\u3068\u8aad\u3093\u3067\u3044\u305f\u8a18\u4e8b\u3067\u7d39\u4ecb\u3055\u308c\u3066\u3044\u305f\u305f\u3081\u3002\u4ed6\u306b\u3082\u300c<a href=\"https:\/\/opennmt.net\/\">OpenNMT<\/a>\u300d\u306a\u3069\u69d8\u3005\u306a\u9078\u629e\u80a2\u304c\u3042\u308b\u3002<br \/>[7] \u305d\u306e\u4ed6\u3001\u30af\u30ed\u30fc\u30eb\u306bt2.medium\u3001\u5bfe\u8a33\u30da\u30a2\u306e\u5224\u65ad\u306br5a.2xlarge \u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u306a\u3069\u3092\u4f7f\u7528\u3057\u3066\u3044\u308b\u3002<br \/>[8] \u539f\u6587\u3078\u306e\u30ea\u30f3\u30af\u304c\u3042\u308b\u30cb\u30e5\u30fc\u30b9\u30b5\u30a4\u30c8\u306a\u3069<br \/>[9] \u30a4\u30e1\u30fc\u30b8\u3068\u3057\u3066<a href=\"https:\/\/www.ppc.go.jp\/files\/pdf\/gdpr-preface-ja.pdf\">\u6cd5\u4ee4\u95a2\u9023\u306e\u5bfe\u8a33<\/a>\u306e\u3088\u3046\u306a\u3082\u306e\u304c\u8f09\u3063\u3066\u3044\u308b\u30b5\u30a4\u30c8<br \/>[10] <a href=\"https:\/\/nlp.stanford.edu\/projects\/jesc\/\">JESC: Japanese-English Subtitle Corpus, 2018<\/a><br \/>[11] <a href=\"http:\/\/www.phontron.com\/kftt\/index-ja.html\">The Kyoto Free Translation Task, 2011<\/a><br \/>[12] <a href=\"http:\/\/www.edrdg.org\/wiki\/index.php\/Tanaka_Corpus\">Tanaka Corpus<\/a><br \/>[13] \u8ffd\u52a0\u3057\u305f\u30c7\u30fc\u30bf\u306fCC\u30e9\u30a4\u30bb\u30f3\u30b9\u3067NC\u304c\u3064\u3044\u3066\u3044\u306a\u3044\u305f\u3081\u3001\u304a\u305d\u3089\u304f\u5546\u7528\u5229\u7528\u3082\u53ef\u80fd\u306a\u30c7\u30fc\u30bf\u306b\u306a\u3063\u3066\u3044\u308b\u3068\u601d\u308f\u308c\u308b\u3002\uff08\u4f5c\u6210\u3057\u305f\u30a8\u30f3\u30b8\u30f3\u306e\u30e9\u30a4\u30bb\u30f3\u30b9\u304c\u30c7\u30fc\u30bf\u3068\u3069\u3046\u95a2\u4fc2\u3059\u308b\u304b\u306f\u5225\u306e\u8b70\u8ad6\u304c\u5fc5\u8981\uff09<br \/>[14] <a href=\"https:\/\/huggingface.co\/\">Hugging Face<\/a>\u3067Helsinki-NLP\/opus-mt-ja-en\u3068\u3044\u3046\u65e5\u672c\u8a9e\u2192\u82f1\u8a9e\u306e\u7ffb\u8a33\u30e2\u30c7\u30eb\u3092\u516c\u958b\u3057\u3066\u3044\u308b\u65b9\u306e\u30ea\u30dd\u30b8\u30c8\u30ea\u3067\u3042\u308b\u3002<br \/>[15] \u4eca\u56de\u306f\u300c\u2460 \u82f1\u8a9e\u2192\u65e5\u672c\u8a9e\u300d\u300c\u2461 \u65e5\u672c\u8a9e\u2192\u82f1\u8a9e\u300d\u300c\u2462 \u2460\u3092KFTT\u7528\u306bfine tuning\u300d\u306e3\u30e2\u30c7\u30eb\u3092\u4f5c\u6210\u3001\u74b0\u5883\u69cb\u7bc9\u3092\u542b\u308160\u6642\u9593\u7a0b\u5ea6\u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u3092\u4f7f\u3063\u3066\u3044\u308b\u3002GPU\u30a4\u30f3\u30b9\u30bf\u30f3\u30b9\u306f$4.194\/\u6642\u9593\u3068\uff08\u30b3\u30b9\u30c8\u30d1\u30d5\u30a9\u30fc\u30de\u30f3\u30b9\u306f\u826f\u3044\u304c\uff09\u7d50\u69cb\u9ad8\u3044\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u305d\u306e\u4ed6<\/h4>\n\n\n\n<p>\u590f\u4f11\u307f\u3068\u3044\u3046\u3053\u3068\u3067\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u4f5c\u6210\u304b\u3089\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u69cb\u7bc9\uff0bWEB\u5316\u3092\u4e00\u901a\u308a\u5b9f\u65bd\u3057\u3066\u307f\u305f\u3002\u624b\u3092\u52d5\u304b\u3059\u3068\u697d\u3057\u3044\u3002\u6027\u80fd\u3068\u3057\u3066\u3082\u307e\u305a\u307e\u305a\u3067\u30c9\u30e1\u30a4\u30f3\u306b\u3088\u3063\u3066\u306f\u4ed6\u306e\u30a8\u30f3\u30b8\u30f3\u3068\u3082\u826f\u3044\u52dd\u8ca0\u304c\u3067\u304d\u305d\u3046\u3067\u3042\u308b\u3002\u307e\u305f\u3001fine tuning\u3092\u884c\u3048\u3070\u300c<a href=\"https:\/\/cloud.google.com\/translate\/automl\/docs\/evaluate?hl=ja#interpretation\">\u9069\u5ea6\u306a\u54c1\u8cea\u306e\u7ffb\u8a33<\/a>\u300d\u30ec\u30d9\u30eb\u307e\u3067\u884c\u3051\u305d\u3046\u3002\uff08\u305d\u3082\u305d\u3082BLEU\u304c\u4fe1\u983c\u3067\u304d\u308b\u304b\u3001\u3068\u3044\u3046\u8b70\u8ad6\u306f\u3042\u308b\u306e\u3060\u304c\u30fb\u30fb\u30fb\u3002\uff09<\/p>\n\n\n\n<p>\u73fe\u72b6\u30af\u30ed\u30fc\u30eb\u3057\u305f\u5148\u306f\u6280\u8853\u7cfb\u306b\u504f\u3063\u3066\u304a\u308a\u3001\u4f01\u696d\u306e\u30d7\u30ec\u30b9\u30ea\u30ea\u30fc\u30b9\u3084\u30de\u30cb\u30e5\u30a2\u30eb\u3001\u516c\u7684\u6a5f\u95a2\u306e\u30b5\u30a4\u30c8\u304c\u5c11\u306a\u3044\u3002\u3053\u306e\u8fba\u308a\u3092\u62e1\u5145\uff06\u5bfe\u8a33\u30da\u30a2\u3092\u53d6\u308b\u30d7\u30ed\u30b0\u30e9\u30e0\u3092\u9ad8\u5ea6\u5316\u3059\u308c\u3070\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u30922-3\u500d\u306b\u62e1\u5f35\u3059\u308b\u3053\u3068\u306f\u53ef\u80fd\u306a\u6c17\u304c\u3057\u3066\u3044\u308b\u3002\uff08\u3068\u3044\u3046\u304bJParaCrawl\u304c\u5b9f\u73fe\u3057\u3066\u3044\u308b\u3002\uff09 \u305d\u3046\u306a\u308b\u3068\u30d5\u30ea\u30fc\u306e\u30a8\u30f3\u30b8\u30f3\u3068\u3057\u3066\u306f\u5b58\u5728\u611f\u304c\u51fa\u308b\u306e\u304b\u306a\u3068\u601d\u308f\u306a\u304f\u3082\u306a\u3044\u3002<\/p>\n\n\n\n<p>\u6280\u8853\u7684\u306b\u3082\u3067\u304d\u308b\u3053\u3068\u306f\u591a\u3044\u3002\u5192\u982d\u306eBack Translation\u4ee5\u524d\u306b\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3082\u5341\u5206\u3067\u306f\u306a\u3044\u3002\u3068\u306f\u3044\u3048\u30011\u56de1\u4e07\u5186\u304b\u304b\u308b\u3068\u601d\u3046\u3068\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3092\u30b4\u30ea\u30b4\u30ea\u3084\u308b\u306e\u306f\u6c17\u304c\u5f15\u3051\u308b\u30fb\u30fb\u30fb\u3002<\/p>\n\n\n\n<p>\u3044\u3064\u3082\u901a\u308a\u3001\u4eca\u5f8c\u30e2\u30c7\u30eb\u306e\u6539\u5584\u3092\u5b9f\u65bd\u3057\u3066\u3044\u304f\u304b\u306f\u60c5\u71b1\u6b21\u7b2c\u3068\u3044\u3063\u305f\u3068\u3053\u308d\u3002<\/p>\n\n\n\n<p>\u81ea\u5206\u3067\u66f8\u3044\u305f\u30b3\u30fc\u30c9\u3068\u30e2\u30c7\u30eb\u306f\u305d\u306e\u3046\u3061\u516c\u958b\u3057\u3088\u3046\u304b\u306a\u3001\u3068\u601d\u3063\u3066\u3044\u308b\u304c\u5ff5\u306e\u305f\u3081\u30e9\u30a4\u30bb\u30f3\u30b9\u95a2\u9023\u3092\u78ba\u8a8d\u4e2d\u3002\u6cd5\u7684\u306b\u306f\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3082\u516c\u958b\u53ef\u80fd\u306a\u306e\u3067\u306f\u306a\u304b\u308d\u3046\u304b\u3068\u601d\u3063\u3066\u3044\u308b\u3082\u306e\u306e\u3001\u305d\u3053\u307e\u3067\u516c\u958b\u3059\u308b\u304b\u306f\u6c7a\u3081\u3066\u3044\u306a\u3044\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u590f\u4f11\u307f\u306e\u5bbf\u984c\uff08\uff1f\uff09\u3068\u3057\u3066Marian-NMT\u3092\u4f7f\u3063\u3066\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u3092\u69cb\u7bc9\u3057\u3066\u307f\u305f\u3002\u69cb\u7bc9\u3057\u305f\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u306f\u300c\u82f1\u8a9e\u2192\u65e5\u672c\u8a9e\u7ffb\u8a33\uff08https:\/\/devneko.jp\/demo\/\uff09\u300d\u304b\u3089\u8a66\u3059\u3053\u3068\u304c\u3067\u304d\u308b\u3002\u8a33\u3059\u3079\u304d\u5358\u8a9e\u3092\u98db\u3070\u3059\u306a\u3069 &hellip; <a href=\"https:\/\/staka.jp\/wordpress\/?p=276\" class=\"more-link\"><span class=\"screen-reader-text\">&#8220;\u7ffb\u8a33\u30a8\u30f3\u30b8\u30f3\u306e\u69cb\u7bc9\uff08Marian-NMT\uff09&#8221; \u306e<\/span>\u7d9a\u304d\u3092\u8aad\u3080<\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2,8,11],"tags":[27,45,83],"class_list":["post-276","post","type-post","status-publish","format-standard","hentry","category-ai","category-python","category-11","tag-deeplearning","tag-marian","tag-83"],"_links":{"self":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/276","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=276"}],"version-history":[{"count":0,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/276\/revisions"}],"wp:attachment":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=276"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=276"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=276"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}