{"id":4529,"date":"2024-02-15T12:39:19","date_gmt":"2024-02-15T03:39:19","guid":{"rendered":"https:\/\/blog.since2020.jp\/?p=4529"},"modified":"2024-02-15T12:44:31","modified_gmt":"2024-02-15T03:44:31","slug":"bert%e3%81%ab%e3%82%88%e3%82%8b%e3%80%81%e6%ad%8c%e8%a9%9e%e3%81%ab%e5%9f%ba%e3%81%a5%e3%81%8fmr-children%e3%81%ae%e6%9b%b2%e3%81%ae%e5%88%86%e9%a1%9e%e3%81%ab%e6%8c%91%e6%88%a6%ef%bc%81%e3%80%90part","status":"publish","type":"post","link":"https:\/\/since2020.jp\/media\/bert%e3%81%ab%e3%82%88%e3%82%8b%e3%80%81%e6%ad%8c%e8%a9%9e%e3%81%ab%e5%9f%ba%e3%81%a5%e3%81%8fmr-children%e3%81%ae%e6%9b%b2%e3%81%ae%e5%88%86%e9%a1%9e%e3%81%ab%e6%8c%91%e6%88%a6%ef%bc%81%e3%80%90part\/","title":{"rendered":"BERT\u306b\u3088\u308b\u3001\u6b4c\u8a5e\u306b\u57fa\u3065\u304fMr.Children\u306e\u66f2\u306e\u5206\u985e\u306b\u6311\u6226\uff01\u3010Part\u2460\u3011"},"content":{"rendered":"\n

Mr.Children\u306e\u66f2\u309220\u66f2\u53d6\u308a\u4e0a\u3052\u3001BERT\u3092\u5229\u7528\u3057\u3066\u3001\u6b4c\u8a5e\u306b\u57fa\u3065\u3044\u305f\u5206\u985e\u3092\u5b9f\u884c\u3057\u3066\u307f\u307e\u3057\u305f\u3002<\/p>\n\n\n

\u306f\u3058\u3081\u306b<\/h2>\n

BERT\uff08Bidirectional Encoder Representations from Transformers\uff09\u3068\u3044\u3046\u30e2\u30c7\u30eb\u3092\u3054\u5b58\u77e5\u3067\u3057\u3087\u3046\u304b\uff1fBERT\u306f\u3001Google\u304c\u958b\u767a\u3057\u305f\u81ea\u7136\u8a00\u8a9e\u51e6\u7406\u306e\u305f\u3081\u306e\u6df1\u5c64\u5b66\u7fd2\u30e2\u30c7\u30eb<\/strong>\u3067\u3059\u3002\u3053\u308c\u306f\u3001\u9ad8\u5ea6\u306a\u8a00\u8a9e\u7406\u89e3\u30bf\u30b9\u30af\u3092\u5b9f\u884c\u3059\u308b\u305f\u3081\u306b\u958b\u767a\u3055\u308c\u3001\u6587\u7ae0\u306e\u8981\u7d04\u3001\u6587\u66f8\u5206\u985e\u3001\u8cea\u554f\u5fdc\u7b54\u3001\u8a00\u8a9e\u7ffb\u8a33\u306a\u3069\u3001\u69d8\u3005\u306a\u3068\u3053\u308d\u3067\u5fdc\u7528\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\r\n

\u672c\u8a18\u4e8b\u3067\u306f\u3001BERT\u3092\u7528\u3044\u3066\u3001\u6b4c\u8a5e\u306b\u57fa\u3065\u3044\u305fMr.Children\u306e\u66f2\u306e\u5206\u985e\u3092\u8a66\u3057\u3066\u307f\u3088\u3046\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\r\n

\u5bfe\u8c61\u3068\u3059\u308b\u306e\u306f\u3001\u6b21\u306e20\u66f2\u3067\u3059\u3002<\/p>\r\n

\u300c365\u65e5\u300d\u300c\u3057\u308b\u3057\u300d\u300c\u7d42\u308f\u308a\u306a\u304d\u65c5\u300d\u300cSign\u300d\u300cGIFT\u300d\u300c\u540d\u3082\u306a\u304d\u8a69\u300d\u300cTomorrow never knows\u300d\u300cHERO\u300d\u300c\u30a8\u30bd\u30e9\u300d\u300cinnocent world\u300d\u300c\u304f\u308b\u307f\u300d\u300c\u62b1\u304d\u3057\u3081\u305f\u3044\u300d\u300c\u65c5\u7acb\u3061\u306e\u5504\u300d\u300c\u541b\u304c\u597d\u304d\u300d\u300c\u30b7\u30fc\u30bd\u30fc\u30b2\u30fc\u30e0 \uff5e\u52c7\u6562\u306a\u604b\u306e\u6b4c\uff5e\u300d\u300c\u7948\u308a \uff5e\u6d99\u306e\u8ecc\u9053\u300d\u300cyouthful days\u300d\u300c\u638c\u300d\u300cMarshmallow day\u300d<\/p>\r\n

\u5206\u985e\u306f\u6b21\u306e\u624b\u9806\u3067\u884c\u3044\u307e\u3059\u3002<\/p>\r\n

    \r\n\t
  1. \u6b4c\u8a5e\uff08\u6587\u5b57\u5217\uff09\u3092\u6570\u5024\u5316<\/li>\r\n\t
  2. \u6570\u5024\u5316\u3057\u305f\u30d9\u30af\u30c8\u30eb\u3092\u30af\u30e9\u30b9\u30bf\u30fc\u5206\u6790\uff08K\u5e73\u5747\u6cd5\uff09<\/li>\r\n<\/ol>\r\n

    <\/p>\n\n

    \u6e96\u5099<\/h2>\n

    \u306f\u3058\u3081\u306b\u3001\u8af8\u3005\u306e\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u3092\u6e08\u307e\u305b\u307e\u3059\u3002<\/p>\r\n

    \u8aad\u307f\u8fbc\u3080csv\u30d5\u30a1\u30a4\u30eb\u306b\u306f\u30011\u5217\u76ee\u306b\u300c\u66f2\u540d\u300d\u30012\u5217\u76ee\u306b\u300c\u6b4c\u8a5e\u300d\u306e\u60c5\u5831\u304c\u5165\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\r\n

    <\/p>\r\n

    \r\n
    !pip install japanize_matplotlib\r\n!pip install fugashi\r\n!pip install ipadic\r\n\r\nimport numpy as np\r\nimport pandas as pd\r\nimport matplotlib.pyplot as plt\r\nimport torch\r\nfrom sklearn.feature_extraction.text import TfidfVectorizer\r\nfrom sklearn.manifold import TSNE\r\nfrom transformers import BertJapaneseTokenizer, BertModel\r\nfrom sklearn.cluster import KMeans\r\n\r\ndf= pd.read_csv(\"Mr.Children.csv\")\r\ndf= df.iloc[1:20]#\u4eca\u56de\u306f\u6b4c\u8a5e\u95b2\u89a7\u30e9\u30f3\u30ad\u30f3\u30b0\u4e0a\u4f4d20\u500b\u3092\u5bfe\u8c61\u3068\u3057\u307e\u3059<\/code><\/pre>\r\n<\/div>\r\n
    \u6b21\u306b\u3001BERT\u306e\u65e5\u672c\u8a9e\u30e2\u30c7\u30eb\u3092\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9\u3057\u307e\u3059\u3002<\/p>\r\n
    \r\n
    # BERT\u306e\u65e5\u672c\u8a9e\u30e2\u30c7\u30eb\u3001tokenizer\u3001model\u306e\u30c0\u30a6\u30f3\u30ed\u30fc\u30c9 \r\nBERT_model_jp = 'cl-tohoku\/bert-base-japanese-whole-word-masking' \r\ntokenizer = BertJapaneseTokenizer.from_pretrained(BERT_model_jp) \r\nmodel = BertModel.from_pretrained(BERT_model_jp)<\/code><\/pre>\r\n<\/div>\r\n
    \u30c7\u30fc\u30bf\u30d5\u30ec\u30fc\u30e0\u306e\u4e2d\u8eab\u306f\u4e0b\u306e\u753b\u50cf\u306e\u3088\u3046\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\r\n
    \u3053\u308c\u3067\u6e96\u5099\u5b8c\u4e86\u3067\u3059\u3002
    \r\n<\/p>\r\n
     <\/p>\n\n
    \u6b4c\u8a5e\u3092\u6570\u5024\u5316<\/h2>\n
    \u65e9\u901f\u3001\u6b4c\u8a5e\u3092\u6570\u5024\u30c7\u30fc\u30bf\u306b\u5909\u63db\u3057\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002\u6587\u5b57\u5217\u3092\u6570\u5024\u306b\u5909\u63db\u3059\u308b\u5f79\u5272\u3092\u62c5\u3046\u306e\u304c\u3001\u5148\u307b\u3069\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3057\u305f\u3001BERT\u30e2\u30c7\u30eb\u3067\u3059\u3002<\/p>\r\n
    \r\n
    #\u6b4c\u8a5e\u306e\u30d9\u30af\u30c8\u30eb\u5316\r\nmax_length = 256\r\nsentence_vec= []\r\n\r\nfor i in range(len(df)):\r\nlines = df.iloc[i, 1].splitlines()\r\ntext = '\\n'.join(lines)\r\nencoding = tokenizer(\r\ntext,\r\nmax_length=max_length,\r\npadding='max_length',\r\ntruncation=True,\r\nreturn_tensors='pt'\r\n)\r\nwith torch.no_grad():\r\noutput = model(**encoding)\r\nlast_hidden_state = output.last_hidden_state\r\nattention_mask = encoding['attention_mask']\r\nnormalized_hidden_state = (last_hidden_state * attention_mask.unsqueeze(-1)).sum(1) \/ attention_mask.sum(1, keepdim=True)\r\n\r\nsentence_vec.append(normalized_hidden_state.numpy())\r\n\r\n#\u30ea\u30b9\u30c8\u306e\u5f62\u5f0f\u3092\u5909\u66f4\r\nsentence_vectors = np.vstack(sentence_vectors)\r\n\r\n<\/code><\/pre>\r\n
    \u6700\u5f8c\u306e\u3001sentence_vectors\u306e\u6b21\u5143\u306f\u3001<\/p>\r\n
    sentence_vectors.shape\r\n<\/code><\/pre>\r\n
    1\u66f2\u306e\u6b4c\u8a5e\u304c\u3001768\u6b21\u5143\u306e\u6570\u5024\u30d9\u30af\u30c8\u30eb\u306b\u5909\u63db\u3055\u308c\u3066\u304a\u308a\u3001\u305d\u308c\u304c20\u66f2\u5206\u4e26\u3093\u3060\u3082\u306e\u304c\u3001sentence_vectors\u306b\u683c\u7d0d\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u3053\u3053\u307e\u3067\u3067\u3001\u6b4c\u8a5e\uff08\u6587\u5b57\uff09\u306e\u60c5\u5831\u3092\u6570\u5024\u306b\u5909\u63db\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3057\u305f\u3002<\/code><\/p>\r\n<\/div>\n\n
    \u6b4c\u8a5e\u306b\u57fa\u3065\u3044\u3066\u3001\u66f2\u3092\u30b0\u30eb\u30fc\u30d7\u5206\u3051<\/h2>\n
    \u4eca\u56de\u306f\u3001K\u5e73\u5747\u6cd5\u3068\u3044\u3046\u624b\u6cd5\u3092\u7528\u3044\u3066\u30af\u30e9\u30b9\u30bf\u30ea\u30f3\u30b0\u3092\u884c\u3044\u307e\u3059\u3002K\u5e73\u5747\u6cd5\u306e\u5b9f\u884c\u306f\u3001\u6b21\u306e2\u884c\u306e\u30b3\u30fc\u30c9\u3067\u884c\u3046\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002\u30e2\u30c7\u30eb\u3092\u4f5c\u308b\u969b\u306b\u3001\u30b0\u30eb\u30fc\u30d7\u306e\u6570\uff08n_clusters\u306e\u90e8\u5206\uff09\u3092\u89e3\u6790\u8005\u304c\u9078\u3076\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u304c\u3001\u4eca\u56de\u306f4\u3064\u3068\u3057\u307e\u3057\u305f\u3002<\/p>\r\n
    \r\n
    kmeans = KMeans(n_clusters=4, random_state=22) \r\nkmeans.fit(sentence_vectors)<\/code><\/pre>\r\n<\/div>\r\n
    \u6700\u5f8c\u306b\u3001K\u5e73\u5747\u6cd5\u306b\u3088\u308b\u30af\u30e9\u30b9\u30bf\u30ea\u30f3\u30b0\u306e\u7d50\u679c\u3092\u898b\u3066\u307f\u307e\u3059\u3002<\/p>\r\n
    \r\n
    from collections import defaultdict #\u30af\u30e9\u30b9\u30bf\u30fc\u3054\u3068\u306b\u6240\u5c5e\u3059\u308b\u30c7\u30fc\u30bf\u30dd\u30a4\u30f3\u30c8\u306e\u884c\u3092\u307e\u3068\u3081\u308b\u8f9e\u66f8\u3092\u4f5c\u6210 \r\ncluster_dict = defaultdict(list) \r\n\r\n#\u5404\u30c7\u30fc\u30bf\u30dd\u30a4\u30f3\u30c8\u3092\u30af\u30e9\u30b9\u30bf\u30fc\u3054\u3068\u306b\u5206\u985e \r\nfor i, label in enumerate(kmeans.labels_): cluster_dict[label].append(df.iloc[i]) \r\nsorted_clusters = sorted(cluster_dict.keys()) \r\n\r\n#\u5404\u30af\u30e9\u30b9\u30bf\u30fc\u306b\u6240\u5c5e\u3059\u308b\u30af\u30e9\u30b9\u30bf\u30fc\u3092\u8868\u793a \r\nfor cluster in sorted_clusters: \r\n   print(f\"\u30af\u30e9\u30b9\u30bf\u30fc {cluster} \u306b\u6240\u5c5e\u3059\u308b\u66f2\u540d:\") \r\n   for data_point in cluster_dict[cluster]: \r\n       print(data_point.iloc[0]) \r\n   print(\" \")<\/code><\/pre>\r\n<\/div>\r\n
    <\/p>\n\n
    \u7d50\u679c\u767a\u8868\uff01<\/h2>\n
    \u5148\u307b\u3069\u306e\u30af\u30e9\u30b9\u30bf\u30ea\u30f3\u30b0\u306e\u7d50\u679c\u3001\u6b21\u306e\u3088\u3046\u306b\u30b0\u30eb\u30fc\u30d7\u5206\u3051\u3055\u308c\u307e\u3057\u305f\u3002<\/p>\r\n
      \r\n\t
    • \u30af\u30e9\u30b9\u30bf\u30fc\uff10\uff1a \u300c\u3057\u308b\u3057\u300d\u300c\u7d42\u308f\u308a\u306a\u304d\u65c5\u300d\u300c\u540d\u3082\u306a\u304d\u8a69\u300d\u300c\u304f\u308b\u307f\u300d\u300c\u62b1\u304d\u3057\u3081\u305f\u3044\u300d\u300c\u65c5\u7acb\u3061\u306e\u5504\u300d\u300c\u541b\u304c\u597d\u304d\u300d\u300c\u30b7\u30fc\u30bd\u30fc\u30b2\u30fc\u30e0 \uff5e\u52c7\u6562\u306a\u604b\u306e\u6b4c\uff5e\u300d\u300c\u7948\u308a \uff5e\u6d99\u306e\u8ecc\u9053\u300d\u300c\u638c\u300d<\/li>\r\n\t
    • \u30af\u30e9\u30b9\u30bf\u30fc 1 \uff1a \u300c365\u65e5\u300d\u300cSign\u300d\u300cTomorrow never knows\u300d\u300cinnocent world\u300d\u300cyouthful days\u300d \u300cMarshmallow day\u300d<\/li>\r\n\t
    • \u30af\u30e9\u30b9\u30bf\u30fc \uff12 \uff1a\u300c\u30a8\u30bd\u30e9\u300d<\/li>\r\n\t
    • \u30af\u30e9\u30b9\u30bf\u30fc\uff13\uff1a\u300cHANABI\u300d\u300cGIFT\u300d\u300cHERO\u300d<\/li>\r\n<\/ul>\r\n
      <\/p>\n\n
      \u6700\u5f8c\u306b<\/h2>\n
      \u30af\u30e9\u30b9\u30bf\u30fc\u5206\u6790\u306b\u3088\u3063\u3066\u3001\u6b4c\u8a5e\u306b\u57fa\u3065\u3044\u3066\u66f2\u3092\u30014\u3064\u306e\u30b0\u30eb\u30fc\u30d7\u306b\u5206\u985e\u3092\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3057\u305f\u3002<\/p>\r\n
      \u3057\u304b\u3057\u306a\u304c\u3089\u3001\u3053\u3053\u3067\u3001\u91cd\u8981\u306a\u904e\u5931\u306b\u6c17\u304c\u3064\u304d\u307e\u3057\u305f\u3002\u300c\u7b46\u8005\u306f\u3001\u4eca\u56de\u6271\u3063\u305f20\u66f2\u306eMr.Children\u306e\u6b4c\u8a5e\u3001\u5168\u90e8\u628a\u63e1\u3067\u304d\u3066\u3044\u307e\u305b\u3093\u3002\u300d\u305d\u306e\u305f\u3081\u3001\u30af\u30e9\u30b9\u30bf\u30fc\u5206\u6790\u306e\u7d50\u679c\u304c\u3001\u975e\u5e38\u306b\u826f\u3044\u306e\u304b\u3001\u3042\u308b\u7a0b\u5ea6\u826f\u3044\u306e\u304b\u3001\u4eca\u3072\u3068\u3064\u306a\u306e\u304b\u3001\u306f\u3063\u304d\u308a\u3057\u305f\u3053\u3068\u306f\u8a00\u3048\u307e\u305b\u3093\u2026\u3002<\/p>\r\n
      BERT\u3092\u7528\u3044\u308b\u3053\u3068\u3067\u3001\u6587\u7ae0\u30c7\u30fc\u30bf\u3092\u6570\u5024\u30c7\u30fc\u30bf\u306b\u5909\u63db\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3057\u305f\u3002\u6587\u7ae0\u306e\u307e\u307e\u3060\u3068\u6271\u3044\u306b\u304f\u3044\u30c7\u30fc\u30bf\u3082\u3001\u6570\u5024\u30c7\u30fc\u30bf\u306b\u5909\u63db\u3059\u308b\u3053\u3068\u3067\u3001\u8272\u3005\u306a\u5206\u6790\u304c\u53ef\u80fd\u306b\u306a\u308a\u307e\u3059\u3002\u4eca\u56de\u6271\u3063\u305fK\u5e73\u5747\u6cd5\u4ee5\u5916\u306b\u3082\u30af\u30e9\u30b9\u30bf\u30ea\u30f3\u30b0\u624b\u6cd5\u306f\u3042\u308a\u307e\u3059\u3057\u3001\u6b21\u5143\u524a\u6e1b\u306a\u3069\u4ed6\u306e\u65b9\u6cd5\u3092\u8a66\u3057\u3066\u3082\u9762\u767d\u305d\u3046\u3067\u3059\u3002<\/p>\r\n
      \u6700\u5f8c\u307e\u3067\u3001\u8aad\u3093\u3067\u3044\u305f\u3060\u304d\u3042\u308a\u304c\u3068\u3046\u3054\u3056\u3044\u307e\u3057\u305f\uff01<\/p>\r\n
      <\/p>","protected":false},"excerpt":{"rendered":"
      Mr.Children\u306e\u66f2\u309220\u66f2\u53d6\u308a\u4e0a\u3052\u3001BERT\u3092\u5229\u7528\u3057\u3066\u3001\u6b4c\u8a5e\u306b\u57fa\u3065\u3044\u305f\u5206\u985e\u3092\u5b9f\u884c\u3057\u3066\u307f\u307e\u3057\u305f\u3002 \u306f\u3058\u3081\u306b BERT\uff08Bidirectional Encoder Representations from Trans […]<\/p>\n","protected":false},"author":16,"featured_media":4532,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"content-type":"","swell_btn_cv_data":"","footnotes":"","_wp_rev_ctl_limit":""},"categories":[1246],"tags":[],"class_list":["post-4529","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-data-infrastructure"],"_links":{"self":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts\/4529","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/users\/16"}],"replies":[{"embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/comments?post=4529"}],"version-history":[{"count":0,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/posts\/4529\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/media\/4532"}],"wp:attachment":[{"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/media?parent=4529"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/categories?post=4529"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/since2020.jp\/media\/wp-json\/wp\/v2\/tags?post=4529"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}