{"id":8892,"date":"2025-10-10T09:26:35","date_gmt":"2025-10-10T09:26:35","guid":{"rendered":"https:\/\/pokecon.jp\/job\/?p=8892"},"modified":"2025-10-10T09:26:35","modified_gmt":"2025-10-10T09:26:35","slug":"%e3%83%99%e3%82%af%e3%83%88%e3%83%ab%e6%a4%9c%e7%b4%a2%e3%81%aetext-embedding-%e3%81%ae%e5%9f%ba%e7%a4%8e%e3%82%92%e7%90%86%e8%a7%a3%e3%81%99%e3%82%8b-%e3%83%95%e3%83%ac%e3%82%af%e3%83%88%e3%81%ae","status":"publish","type":"post","link":"https:\/\/pokecon.jp\/job\/8892\/","title":{"rendered":"\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306eText Embedding \u306e\u57fa\u790e\u3092\u7406\u89e3\u3059\u308b – \u30d5\u30ec\u30af\u30c8\u306e\u30af\u30e9\u30a6\u30c9blog"},"content":{"rendered":"\n<\/p>\n
\n

\u3053\u3093\u306b\u3061\u306f\u3002\u30a8\u30f3\u30b8\u30cb\u30a2\u306e\u5927\u6a4b\u3067\u3059\u3002<\/p>\n

\u5148\u65e5\u3001\u3068\u3042\u308b\u6280\u8853\u691c\u8a3c\u306b\u3066AI\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u306eRAG\u69cb\u7bc9\u3092\u62c5\u5f53\u3057\u305f\u4eba\u304c\u3001\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306e\u7cbe\u5ea6\u304c\u30a4\u30de\u30a4\u30c1\u3068\u3044\u3046\u3053\u3068\u3067\u60a9\u3093\u3067\u3044\u307e\u3057\u305f\u3002\u8a73\u3057\u304f\u8a71\u3092\u805e\u304f\u3068\u3001\u3069\u3046\u3084\u3089\u7279\u5b9a\u306e\u30ad\u30fc\u30ef\u30fc\u30c9\u3067\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u3057\u305f\u5834\u5408\u306b\u3001\u3042\u307e\u308a\u826f\u3044\u691c\u7d22\u7d50\u679c\u304c\u5f97\u3089\u308c\u306a\u304b\u3063\u305f\u3088\u3046\u3067\u3059\u3002\u305d\u306e\u5f8c\u3082\u3046\u4e00\u5ea6\u8a71\u3092\u805e\u3044\u3066\u307f\u308b\u3068\u3001\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306eEmbedding\u30e2\u30c7\u30eb\u3092\u6587\u5358\u4f4d\u304b\u3089\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u306b\u5909\u3048\u3066\u307f\u308b\u3068\u3001\u304b\u306a\u308a\u7cbe\u5ea6\u304c\u4e0a\u304c\u3063\u305f\u3001\u3068\u3044\u3046\u306e\u3067\u3059\u3002\u3053\u306e\u4e00\u9023\u306e\u4f1a\u8a71\u304b\u3089\u3001\u79c1\u81ea\u8eab\u3053\u306e\u8fba\u308a\u306eEmbedding\u306b\u3064\u3044\u3066\u306e\u7406\u89e3\u3092\u518d\u78ba\u8a8d\u3057\u3066\u304a\u3053\u3046\u3068\u601d\u3044\u307e\u3057\u305f\u3002<\/p>\n

\u3053\u3046\u3044\u3063\u305f\u80cc\u666f\u3082\u3042\u308a\u3001\u4eca\u56de\u306f\u3001RAG\u306e\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306b\u5fc5\u8981\u306aText Embedding\uff08\u30c6\u30ad\u30b9\u30c8\u57cb\u3081\u8fbc\u307f\uff09\u306b\u3064\u3044\u3066\u66f8\u304d\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002\u3053\u306e\u6a5f\u4f1a\u306bText Embedding\u306b\u3064\u3044\u3066\u5b66\u3093\u3060\u30d7\u30ed\u30bb\u30b9\u306e\u4e2d\u3067\u7591\u554f\u306b\u601d\u3063\u305f\u3053\u3068\u3084\u7406\u89e3\u3057\u305f\u3053\u3068\u3092\u5171\u6709\u3057\u3066\u307f\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n

\u305d\u3082\u305d\u3082Embedding\u306f\u306a\u305c\u5fc5\u8981\u306a\u306e\u304b<\/h3>\n

\u305d\u3082\u305d\u3082\u3001Embedding\u306f\u306a\u305c\u5fc5\u8981\u306a\u306e\u3067\u3057\u3087\u3046\u304b\u3002\u305d\u308c\u306f\u3001RAG\u306e\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306b\u304a\u3051\u308bEmbedding\u306e\u4f4d\u7f6e\u4ed8\u3051\u304b\u3089\u8003\u3048\u308b\u5fc5\u8981\u304c\u3042\u308a\u305d\u3046\u3067\u3059\u3002<\/p>\n

AI\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u958b\u767a\u3067\u306f\u3001\u591a\u304f\u306e\u5834\u5408\u3001RAG\u306b\u3088\u308aLLM\u306b\u4f55\u3089\u304b\u306e\u5916\u90e8\u30c7\u30fc\u30bf\u3092\u53c2\u7167\u3055\u305b\u307e\u3059\u3002RAG\u306e\u76ee\u7684\u306f\u3001\u300c\u30e6\u30fc\u30b6\u30fc\u306e\u8cea\u554f\u610f\u56f3\u306b\u5408\u3063\u305f\u30ce\u30a4\u30ba\u306e\u5c11\u306a\u3044\u7684\u78ba\u306a\u30b3\u30f3\u30c6\u30ad\u30b9\u30c8\u3092LLM\u306b\u63d0\u4f9b\u3059\u308b\u3053\u3068\u300d\u3067\u3059\u3002\u3053\u3053\u3067\u30e6\u30fc\u30b6\u30fc\u306e\u8cea\u554f\u610f\u56f3\u306b\u5408\u3063\u305f\u60c5\u5831\u3092\u53d6\u5f97\u3059\u308b\u306b\u306f\u3001\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30b5\u30fc\u30c1\uff08\u610f\u5473\u691c\u7d22\uff09\u3068\u306e\u76f8\u6027\u304c\u826f\u3044\u3068\u8003\u3048\u3089\u308c\u3066\u3044\u307e\u3059\u3002\u305d\u306e\u305f\u3081\u3001RAG\u306e\u691c\u7d22\u624b\u6cd5\u306b\u306f\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u304c\u3088\u304f\u4f7f\u7528\u3055\u308c\u307e\u3059\u3002<\/p>\n

\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306f\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30b5\u30fc\u30c1\u306e\u624b\u6cd5\u3067\u3059\u304c\u3001\u5185\u90e8\u7684\u306b\u306f\u3001\u6587\u7ae0\u4e2d\u306e\u5358\u8a9e\u306e\u610f\u5473\u7684\u306a\u95a2\u4fc2\u6027\u3092\u8e0f\u307e\u3048\u305f\u30d9\u30af\u30c8\u30eb\u8868\u73fe\u306e\u985e\u4f3c\u5ea6\u3092\u8a08\u7b97\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u3053\u3067\u767b\u5834\u3057\u305f\u30d9\u30af\u30c8\u30eb\u8868\u73fe\u304cEmbedding\uff08\u610f\u5473\u30d9\u30af\u30c8\u30eb\uff09\u3067\u3059\u3002\u3064\u307e\u308a\u3001\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306b\u306f\u4e8b\u524d\u306bEmbedding\u751f\u6210\u304c\u5fc5\u8981\u3068\u306a\u308b\u308f\u3051\u3067\u3059\u3002<\/p>\n

Embedding\u30e2\u30c7\u30eb\u3068\u306f<\/h3>\n

\u3053\u306eEmbedding\u751f\u6210\u306b\u306f\u3001Embedding\u30e2\u30c7\u30eb\u304c\u4f7f\u7528\u3055\u308c\u3066\u3044\u307e\u3059\u3002Text Embedding\u30e2\u30c7\u30eb\u306e\u5834\u5408\u3001\u5927\u91cf\u306e\u30c6\u30ad\u30b9\u30c8\u30c7\u30fc\u30bf\u3092\u7528\u3044\u3066\u4e8b\u524d\u5b66\u7fd2\u3055\u308c\u3066\u304a\u308a\u3001\u3053\u306e\u5b66\u7fd2\u30d7\u30ed\u30bb\u30b9\u306b\u3088\u3063\u3066\u3001\u5358\u8a9e\u3084\u6587\u7ae0\u306e\u610f\u5473\u7684\u306a\u95a2\u4fc2\u6027\u3092\u6570\u5024\u30d9\u30af\u30c8\u30eb\u3068\u3057\u3066\u8868\u73fe\u3059\u308b\u80fd\u529b\u3092\u7372\u5f97\u3057\u3066\u3044\u308b\u306e\u3067\u3059\u3002<\/p>\n

Embedding\u30e2\u30c7\u30eb\u306e\u7a2e\u985e<\/h3>\n

\u3067\u306f\u3001Embedding\u30e2\u30c7\u30eb\u306f\u3069\u306e\u3088\u3046\u306a\u5206\u985e\u306b\u5206\u304b\u308c\u308b\u306e\u3067\u3057\u3087\u3046\u304b\u3002\u3044\u308d\u3044\u308d\u306a\u5206\u985e\u8ef8\u306f\u3042\u308b\u3068\u601d\u3044\u307e\u3059\u304c\u3001\u305d\u306e\u4e00\u3064\u306bEmbedding\u306e\u751f\u6210\u5358\u4f4d\u304c\u3042\u308a\u307e\u3059\u3002\u3053\u308c\u306b\u306f\u3001\u300cDocument Embedding\u300d\u300cParagraph Embedding\u300d\u300cSentence Embedding\u300d\u300cToken Embedding\u300d\u300cCharacter Embedding\u300d\u304c\u3042\u308a\u307e\u3059\u3002\u4eca\u56de\u306f\u3001\u5192\u982d\u306e\u7591\u554f\u306b\u3082\u51fa\u3066\u304d\u305f\u300cSentence Embedding\u300d\u3068\u300cToken Embedding\u300d\u306b\u3064\u3044\u3066\u8003\u3048\u3066\u3044\u304d\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n

Sentence Embedding<\/h3>\n

Sentence Embedding\u30e2\u30c7\u30eb\u306e\u5834\u5408\u3001\u6587\u5358\u4f4d\u306b\u5bfe\u3059\u308b\u5358\u4e00\u306eEmbedding\u304c\u751f\u6210\u3055\u308c\u307e\u3059\u3002\u3053\u306e\u6642\u751f\u6210\u3055\u308c\u308bEmbedding\u306f\u3001\u6587\u7ae0\u5168\u4f53\u306e\u610f\u5473\u3092\u8868\u73fe\u3057\u305f\u3082\u306e\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002\u306a\u308b\u307b\u3069\u3001\u3068\u601d\u3046\u4e00\u65b9\u3067\u3001\u3067\u306f\u3001\u4e00\u4f53\u3069\u3046\u3084\u3063\u3066\u6587\u7ae0\u5168\u4f53\u306e\u610f\u5473\u3092\u8868\u73fe\u3057\u305fEmbedding\u3092\u751f\u6210\u3057\u3066\u3044\u308b\u306e\u304b\u3001\u3068\u7591\u554f\u304c\u6e67\u304d\u307e\u3059\u3002\u6587\u5168\u4f53\u306b\u5bfe\u3057\u3066\u4e00\u62ec\u3067Embedding\u751f\u6210\u3059\u308b\u4ed5\u7d44\u307f\u304c\u3042\u308b\u306e\u3067\u3057\u3087\u3046\u304b\u3002<\/p>\n

\u5b9f\u969b\u306f\u3001\u6700\u521d\u306b\u6587\u3092\u69cb\u6210\u3059\u308b\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067Embedding\u3092\u751f\u6210\u3057\u3001\u305d\u308c\u3089\u3092\u4f55\u304b\u3057\u3089\u306e\u65b9\u6cd5\u3067\u30d7\u30fc\u30ea\u30f3\u30b0\uff08\u96c6\u7d04\uff09\u3057\u3066\u3044\u308b\u3088\u3046\u3067\u3059\u3002\u3053\u306e\u3088\u3046\u306a\u65b9\u6cd5\u3067\u6587\u7ae0\u5168\u4f53\u306eEmbedding\u751f\u6210\u304c\u3067\u304d\u308b\u306e\u306f\u3001Sentence Embedding\u30e2\u30c7\u30eb\u306e\u4e8b\u524d\u5b66\u7fd2\u6642\u306b\u79d8\u5bc6\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n

\u3053\u306e\u90e8\u5206\u3092\u7406\u89e3\u3059\u308b\u306b\u306f\u3001Sentence Embedding\u30e2\u30c7\u30eb\u306eSBERT\uff08Sentence BERT\uff09\u306e\u958b\u767a\u7d4c\u7def\u3092\u8ffd\u3046\u306e\u304c\u5206\u304b\u308a\u3084\u3059\u3044\u3067\u3059\u3002\u8ad6\u6587\u306e\u8a73\u7d30\u306a\u5185\u5bb9\u306f\u4ed6\u306e\u8a18\u4e8b\u3067\u3082\u7d39\u4ecb\u3055\u308c\u3066\u3044\u308b\u3068\u601d\u3046\u306e\u3067\u3001\u3053\u3053\u3067\u306f\u7c21\u5358\u306b\u79c1\u306e\u7406\u89e3\u3067\u8aac\u660e\u3057\u307e\u3059\u3002SBERT\u306e\u8ad6\u6587<\/a>\u306b\u8a18\u8f09\u3055\u308c\u3066\u3044\u308b\u3088\u3046\u306b\u3001\u5143\u3005\u3001BERT\u51fa\u529b\u3067\u751f\u6210\u3055\u308c\u305f\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u306eEmbedding\u3092\u30d7\u30fc\u30ea\u30f3\u30b0\u3059\u308b\u3060\u3051\u3067\u306f\u3001\u6587\u7ae0\u5168\u4f53\u306eEmbedding\u3068\u3057\u3066\u306f\u4e0d\u5341\u5206\u3067\u3057\u305f\u3002<\/p>\n

\n

The most<\/a> commonlyused approach is to average the BERT output layer (known as BERT embeddings) or by using the output of the first token (the [CLS] token). As we will show, this common practice yields rather bad sentence embeddings, often worse than averaging GloVe embeddings (Pennington et al., 2014).<\/p>\n<\/blockquote>\n

\u305d\u3053\u3067\u3001\u8ffd\u52a0\u306e\u30d5\u30a1\u30a4\u30f3\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\uff08Siamese Network \u3068 Triplet Network Structure\uff09\u3092\u884c\u3046\u3053\u3068\u3067\u3001\u6587\u7ae0\u5168\u4f53\u306e\u610f\u5473\u3092\u8868\u73fe\u3057\u305fEmbedding\u751f\u6210\u304c\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u3063\u305f\u3001\u3068\u3044\u3046\u308f\u3051\u3067\u3059\u3002<\/p>\n

\n

we present Sentence-BERT (SBERT), a modification of the pretrained BERT network that use siamese and triplet network structures to derive semantically meaningful sentence embeddings that can be compared using cosine<\/a>-similarity.<\/p>\n<\/blockquote>\n

\u300cSiamese Network\u300d\u3068\u300cTriplet Network Structure\u300d\u306e\u8a73\u7d30\u3092\u7406\u89e3\u3059\u308b\u3053\u3068\u3082\u91cd\u8981\u3067\u3059\u304c\u3001\u3053\u3053\u3067\u306f\u300c\u6587\u5168\u4f53\u306eEmbedding\u751f\u6210\u306b\u306f\u3001\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u306eEmbedding\u3092\u751f\u6210\u3057\u3001\u305d\u308c\u3089\u3092\u30d7\u30fc\u30ea\u30f3\u30b0\u3057\u3066\u3044\u308b\u300d\u3053\u3068\u3092\u62bc\u3055\u3048\u3066\u304a\u304d\u305f\u3044\u3068\u3053\u308d\u3067\u3059\u3002<\/p>\n

\u306a\u304a\u3001\u30d7\u30fc\u30ea\u30f3\u30b0\u6226\u7565\u306b\u306f\u300cCLS-token\u306e\u5229\u7528\u300d\u300c\u6700\u5927\u30d7\u30fc\u30ea\u30f3\u30b0\u300d\u300c\u91cd\u307f\u4ed8\u3051\u5e73\u5747\u30d7\u30fc\u30ea\u30f3\u30b0\u300d\u304c\u7d39\u4ecb\u3055\u308c\u3066\u304a\u308a\u3001\u30c7\u30d5\u30a9\u30eb\u30c8\u3067\u306f\u5e73\u5747\u30d7\u30fc\u30ea\u30f3\u30b0\u304c\u63a1\u7528\u3055\u308c\u3066\u3044\u308b\u3088\u3046\u3067\u3059\u3002<\/p>\n

\u3053\u306e\u3088\u3046\u306b\u3001Sentence Embedding\u30e2\u30c7\u30eb\u306f\u3001\u4e8b\u524d\u5b66\u7fd2\u306e\u6642\u70b9\u3067\u6587\u7ae0\u5168\u4f53\u306eEmbedding\u751f\u6210\u306b\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3055\u308c\u3066\u3044\u308b\u305f\u3081\u3001\u79c1\u305f\u3061\u306f\u305d\u306e\u30e2\u30c7\u30eb\u306b\u6587\u7ae0\u3092\u5165\u529b\u3059\u308b\u3060\u3051\u3067\u3001\u6587\u5168\u4f53\u306e\u610f\u5473\u3092\u8868\u73fe\u3059\u308bEmbedding\u304c\u751f\u6210\u3067\u304d\u308b\u3001\u3068\u3044\u3046\u308f\u3051\u3067\u3059\u306d\u3002<\/p>\n

\u3067\u306f\u3001Sentence Embedding\u306f\u3069\u306e\u304f\u3089\u3044\u4f7f\u7528\u3055\u308c\u3066\u3044\u308b\u304b\u3068\u3044\u3046\u3068\u3001RAG\u306e\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u3067\u306f\u30c7\u30d5\u30a1\u30af\u30c8\u30b9\u30bf\u30f3\u30c0\u30fc\u30c9<\/a>\u306b\u306a\u3063\u3066\u3044\u308b\u3088\u3046\u3067\u3059\u3002\u3068\u3044\u3063\u3066\u3082\u3001\u3088\u304f\u3042\u308b\u554f\u3044\u5408\u308f\u305b\u5fdc\u7b54\u3084\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u691c\u7d22\u306a\u3069\u306e\u7528\u9014\u3067\u306f\u3001\u6587\u5168\u4f53\u306e\u610f\u5473\u306e\u985e\u4f3c\u6027\u3067\u691c\u7d22\u3057\u305f\u3044\u5834\u5408\u304c\u591a\u3044\u306e\u3067\u3001\u3053\u308c\u306f\u59a5\u5f53\u306a\u7d50\u679c\u306e\u3088\u3046\u306b\u601d\u3048\u307e\u3059\u3002<\/p>\n

\u4e00\u65b9\u3067\u8ab2\u984c\u3082\u3042\u308a\u307e\u3059\u3002Sentence Embedding\u306f\u6587\u5168\u4f53\u306e\u610f\u5473\u3092\u8868\u73fe\u3067\u304d\u308b\u4e00\u65b9\u3001\u7279\u5b9a\u306e\u30c8\u30fc\u30af<\/a>\u30f3\u306b\u5bfe\u3059\u308b\u985e\u4f3c\u5ea6\u306f\u8003\u616e\u3055\u308c\u306b\u304f\u3044\u50be\u5411\u304c\u3042\u308a\u307e\u3059\u3002\u3053\u308c\u306b\u3088\u308a\u4f55\u304c\u8d77\u304d\u308b\u304b\u3068\u3044\u3046\u3068\u3001\u4f8b\u3048\u3070\u6587\u306e\u610f\u5473\u3092\u6c7a\u5b9a\u3065\u3051\u308b\u91cd\u8981\u306a\u30c8\u30fc\u30af<\/a>\u30f3\u304c\u542b\u307e\u308c\u3066\u3044\u308b\u5834\u5408\u306b\u3001\u305d\u306e\u30c8\u30fc\u30af<\/a>\u30f3\u306e\u985e\u4f3c\u5ea6\u304c\u305d\u308c\u4ee5\u5916\u306e\u30c8\u30fc\u30af<\/a>\u30f3\u3068\u540c\u7387\u3067\u8a08\u7b97\u3055\u308c\u3066\u3057\u307e\u3046\u305f\u3081\u3001\u7279\u5b9a\u306e\u30c8\u30fc\u30af<\/a>\u30f3\u306e\u91cd\u8981\u5ea6\u3092\u9069\u5207\u306b\u53cd\u6620\u3067\u304d\u306a\u304f\u306a\u3063\u3066\u3057\u307e\u3044\u307e\u3059\u3002<\/p>\n

\u3053\u306e\u3088\u3046\u306a\u8ab2\u984c\u3078\u306e\u5bfe\u7b56\u3068\u3057\u3066\u306f\u3001\u591a\u304f\u306e\u5834\u5408\u3001\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u691c\u7d22\u304c\u6d3b\u7528\u3055\u308c\u308b\u3088\u3046\u3067\u3059\u3002\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u691c\u7d22\u306f\u3001\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u3060\u3051\u3067\u306a\u304f\u5f93\u6765\u306e\u30ad\u30fc\u30ef\u30fc\u30c9\u691c\u7d22\u306e\u7d50\u679c\u3092\u52a0\u5473\u3057\u3001\u305d\u308c\u305e\u308c\u306e\u691c\u7d22\u7d50\u679c\u3092\u91cd\u307f\u4ed8\u3051\u3059\u308b\u3053\u3068\u3067\u6700\u7d42\u7684\u306a\u691c\u7d22\u7d50\u679c\u3092\u8abf\u6574\u3059\u308b\u624b\u6cd5\u3067\u3059\u3002\u30cf\u30a4\u30d6\u30ea\u30c3\u30c9\u691c\u7d22\u306b\u3064\u3044\u3066\u306f\u3001\u5225\u306e\u6a5f\u4f1a\u306b\u8a73\u7d30\u3092\u66f8\u304d\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n

Token Embedding<\/h3>\n

Token Embedding\u30e2\u30c7\u30eb\u306e\u5834\u5408\u3001\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067Embedding\u304c\u751f\u6210\u3055\u308c\u307e\u3059\u3002\u3064\u307e\u308a\u3001\u5165\u529b\u30c7\u30fc\u30bf\u3092\u69cb\u6210\u3059\u308b\u30c8\u30fc\u30af<\/a>\u30f3\u306e\u6570\u3060\u3051Embedding\u304c\u751f\u6210\u3055\u308c\u308b\u3068\u3044\u3046\u3053\u3068\u3067\u3059\u3002Token Embedding\u306f\u3001\u5165\u529b\u30c7\u30fc\u30bf\u3092\u8907\u6570\u306eEmbedding\u3067\u8868\u73fe\u3059\u308b\u305f\u3081\u3001Multi-Vector<\/a> Embedding\u306b\u5206\u985e\u3055\u308c\u308b\u3053\u3068\u3082\u3042\u308b\u3088\u3046\u3067\u3059\u3002<\/p>\n

Token Embedding\u3092\u4f7f\u3063\u3066\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u3092\u884c\u3046\u5834\u5408\u3001\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067\u306e\u610f\u5473\u691c\u7d22\u304c\u53ef\u80fd\u306b\u306a\u308b\u305f\u3081\u3001\u6587\u5358\u4f4d\u3088\u308a\u3082\u9ad8\u7cbe\u5ea6\u306e\u691c\u7d22\u304c\u53ef\u80fd\u306b\u306a\u308b\u3088\u3046\u3067\u3059\u3002Qdrant\uff08\u30aa\u30fc\u30d7\u30f3\u30bd\u30fc\u30b9\u306e\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u30a8\u30f3\u30b8\u30f3\uff09\u306e\u516c\u5f0f\u30b5\u30a4\u30c8<\/a>\u306b\u306f\u3001Single Vector<\/a> Embedding\uff08Sentence Embedding\u306a\u3069\uff09\u3068\u6bd4\u8f03\u3057\u3066\u3001\u5165\u529b\u30c7\u30fc\u30bf\u306e\u3088\u308a\u8a73\u7d30\u306a\u610f\u5473\u3092\u6349\u3048\u305f\u691c\u7d22\u304c\u53ef\u80fd\u3067\u3042\u308b\u3068\u8a18\u8f09\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n

\n

ColBERT is an embedding model that produces a matrix (multivector) representation of input text, generating one vector<\/a> per token (a token being a meaningful text unit for a machine learning model). This approach allows ColBERT to capture more nuanced input semantics than many dense embedding models, which represent an entire input with a single vector<\/a>.<\/p>\n<\/blockquote>\n

\u305f\u3060\u3057\u3001\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067Embedding\u3092\u751f\u6210\u3059\u308b\u5206\u3001\u5fc5\u8981\u306a\u4fdd\u5b58\u5bb9\u91cf\u3084\u985e\u4f3c\u5ea6\u7b97\u51fa\u306e\u8a08\u7b97\u91cf\u306f\u5897\u3048\u304c\u3061\u3067\u3059\u3002\u3064\u307e\u308a\u3001\u6027\u80fd\u3068\u30b3\u30b9\u30c8\u306e\u30c8\u30ec\u30fc\u30c9\u30aa\u30d5<\/a>\u3092\u8003\u616e\u3059\u308b\u5fc5\u8981\u304c\u3042\u308b\u3068\u3044\u3046\u308f\u3051\u3067\u3059\u3002<\/p>\n

\n

By producing more granular input representations, ColBERT becomes a strong retriever. However, this advantage comes at the cost of increased resource consumption compared to traditional dense embedding models, both in terms of speed and memory.<\/p>\n<\/blockquote>\n

\u3068\u3044\u3063\u3066\u3082\u3001ColBERTv2\u306e\u3088\u3046\u306a\u6539\u826f\u30e2\u30c7\u30eb<\/a>\u3082\u767b\u5834\u3057\u3066\u304a\u308a\u3001\u3053\u306e\u3088\u3046\u306a\u8ab2\u984c\u306f\u6539\u5584\u3055\u308c\u3064\u3064\u3042\u308b\u306e\u3067\u3001\u4e00\u6982\u306b\u30c8\u30ec\u30fc\u30c9\u30aa\u30d5<\/a>\u3060\u3068\u65ad\u5b9a\u3067\u304d\u306a\u3044\u70b9\u306b\u306f\u6ce8\u610f\u304c\u5fc5\u8981\u3067\u3059\u3002<\/p>\n

\u3061\u306a\u307f\u306b\u3001\u5f15\u7528\u6587\u306e\u4e2d\u306bDense Embedding\u3068\u3044\u3046\u306e\u304c\u767b\u5834\u3057\u307e\u3059\u304c\u3001\u3053\u308c\u306f\u5404\u6b21\u5143\u304c\u8907\u96d1\u306a\u610f\u5473\u3092\u6301\u3064\u5bc6\u30d9\u30af\u30c8\u30eb\u3092\u6307\u3057\u307e\u3059\u3002\u5bfe\u7167\u7684\u306a\u6982\u5ff5\u3068\u3057\u3066Sparse Embedding\u304c\u3042\u308a\u3001\u3053\u3061\u3089\u306f\u53cd\u5bfe\u306b\u591a\u304f\u306e\u6210\u5206\u304c0\u306e\u758e\u30d9\u30af\u30c8\u30eb\u3092\u6307\u3057\u307e\u3059\u3002\u3053\u3053\u3067\u306f\u7c21\u5358\u306a\u8aac\u660e\u306b\u3068\u3069\u3081\u307e\u3059\u304c\u3001Dense\u306f\u610f\u5473\u691c\u7d22\u306e\u6587\u8108\u3001Sparse\u306f\u30ad\u30fc\u30ef\u30fc\u30c9\u691c\u7d22\u306e\u6587\u8108\u306b\u95a2\u9023\u3059\u308b\u8868\u73fe\u3067\u3042\u308b\u3068\u62bc\u3055\u3048\u3066\u304a\u3051\u3070\u826f\u3044\u3067\u3057\u3087\u3046\u3002<\/p>\n

\u3067\u306f\u3001Token Embedding\u30e2\u30c7\u30eb\u306fRAG\u306e\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306b\u4f7f\u308f\u308c\u308b\u304b\u3068\u3044\u3046\u3068\u3001\u305d\u308c\u304c\u305d\u3046\u3067\u3082\u306a\u3044\u3088\u3046\u3067\u3059\u3002
\nQdrant\u306e\u30b5\u30a4\u30c8<\/a>\u306b\u3088\u308b\u3068\u3001Token Embedding\u30e2\u30c7\u30eb\u306e\u4ee3\u8868\u4f8b\u3067\u3042\u308bColBERT\u306f\u3001\u5927\u898f\u6a21\u30b3\u30fc\u30d1\u30b9<\/a>\uff08\u691c\u7d22\u5bfe\u8c61\uff09\u306e\u5834\u5408\u306b\u306f\u3001\u691c\u7d22\u7528\u9014\u306b\u306f\u5411\u304b\u306a\u3044\u3068\u8aac\u660e\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u305d\u306e\u7406\u7531\u306f\u3001\u691c\u7d22\u901f\u5ea6\u304c\u9045\u3044\u3053\u3068\u306b\u3042\u308b\u3088\u3046\u3067\u3059\u3002<\/p>\n

\n

Despite ColBERT being a powerful retriever, its speed limitation might make it less suitable for large-scale retrieval. Therefore, we generally recommend using ColBERT for reranking a small set of already retrieved examples, rather than for first-stage retrieval.<\/p>\n<\/blockquote>\n

\u3053\u3053\u3067\u306fToken Embedding\u306e\u7528\u9014\u306b\u3082\u89e6\u308c\u3089\u308c\u3066\u304a\u308a\u3001\u305d\u308c\u306f\u691c\u7d22\u5f8c\u306e\u30ea\u30e9\u30f3\u30af\u51e6\u7406\u3067\u3059\u3002\u30ea\u30e9\u30f3\u30af\u51e6\u7406\u3068\u306f\u3001\u691c\u7d22\u7d50\u679c\u3092\u7cbe\u67fb\u3057\u3001\u3088\u308a\u6b63\u78ba\u306a\u691c\u7d22\u7d50\u679c\u306b\u30e9\u30f3\u30af\u4ed8\u3051\u3057\u76f4\u3059\u51e6\u7406\u306e\u3053\u3068\u3092\u6307\u3057\u307e\u3059\u3002\u30ea\u30e9\u30f3\u30af\u51e6\u7406\u306e\u5834\u5408\u306f\u3001\u691c\u7d22\u5bfe\u8c61\u304c\u7d5e\u3089\u308c\u3066\u3044\u308b\u306e\u3067\u3001\u8a08\u7b97\u901f\u5ea6\u3088\u308a\u3082\u7cbe\u5ea6\u3092\u512a\u5148\u3055\u305b\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u3001\u3068\u3044\u3046\u308f\u3051\u3067\u3059\u306d\u3002<\/p>\n

\u3053\u306e\u3088\u3046\u306b\u3001Token Embedding\u306f\u30ea\u30e9\u30f3\u30af\u51e6\u7406\u3067\u3088\u304f\u4f7f\u308f\u308c\u3066\u3044\u308b\u306e\u3060\u306a\u3001\u3068\u601d\u3063\u305f\u306e\u3067\u3059\u304c\u3001\u8abf\u67fb\u3092\u9032\u3081\u308b\u3068\u3001\u3069\u3046\u3084\u3089\u30ea\u30e9\u30f3\u30af\u51e6\u7406\u3067\u306fCross-Encoder\u306e\u65b9\u304c\u9069\u3057\u3066\u3044\u308b\u3088\u3046\u3067\u3059\u3002SBERT\u306e\u516c\u5f0f\u30b5\u30a4\u30c8<\/a>\u306b\u306f\u3001\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u8a18\u8f09\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n

\n

The retriever has to be efficient for large document collections with millions of entries. However, it might return irrelevant candidates. A re-ranker based on a Cross-Encoder can substantially improve<\/a> the final results for the user.<\/p>\n<\/blockquote>\n

\u3053\u3053\u3067\u65b0\u305f\u306b\u767b\u5834\u3057\u305fCross-Encoder\u3068\u306f\u4e00\u4f53\u306a\u3093\u3067\u3057\u3087\u3046\u304b\u3002SBERT\u306e\u516c\u5f0f\u30b5\u30a4\u30c8<\/a>\u306b\u3088\u308b\u3068\u3001Cross-Encoder\u306f\u6587\u7ae0\u306e\u30da\u30a2\u3092\u5165\u529b\u3068\u3057\u3066\u53d7\u3051\u53d6\u308a\u3001\u305d\u306e\u985e\u4f3c\u5ea6\u3092\u51fa\u529b\u3059\u308b\u3088\u3046\u3067\u3059\u3002\u3053\u308c\u307e\u3067\u898b\u3066\u304d\u305fSentence Embedding\u3068Token Embedding\u306fBi-Encoder\u306e\u4e00\u7a2e\u3067\u30011\u3064\u306e\u5165\u529b\u306b\u5bfe\u3057\u3066\u3001Embedding\u3092\u51fa\u529b\u3057\u307e\u3059\u3002\u3064\u307e\u308a\u3001Cross-Encoder\u306f\u500b\u5225\u306eEmbedding\u3092\u51fa\u529b\u305b\u305a\u6587\u7ae0\u306e\u30da\u30a2\u306e\u985e\u4f3c\u5ea6\u3092\u51fa\u529b\u3057\u3001Bi-Encoder\u306f\u500b\u5225\u306eEmbedding\u3092\u51fa\u529b\u3059\u308b\u3001\u3068\u3044\u3046\u308f\u3051\u3067\u3059\u306d\u3002<\/p>\n

\u307e\u305f\u3001Cross-Encoder\u306fBi-Encoder\u3088\u308a\u3082\u9ad8\u7cbe\u5ea6\u306e\u985e\u4f3c\u5ea6\u8a08\u7b97\u304c\u53ef\u80fd\u3067\u3059\u304c\u3001\u5b9f\u52d9\u306e\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u7528\u9014\u3067\u306e\u5229\u7528\u306f\u96e3\u3057\u3044\u3068\u8aac\u660e\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n

\n

As detailed in our paper, Cross-Encoder achieve better performances than Bi-Encoders. However, for many application they are not practical as they do not produce embeddings we could e.g. index or efficiently compare using cosine<\/a> similarity.<\/p>\n<\/blockquote>\n

\u305d\u306e\u7406\u7531\u306f\u3001\u8a08\u7b97\u52b9\u7387\u306b\u3042\u308a\u307e\u3059\u3002Bi-Encoder\u306e\u5834\u5408\u306b\u306f\u3001\u4e8b\u524d\u306bEmbedding\u751f\u6210\u3092\u884c\u3044\u3001\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u3092\u69cb\u7bc9\u3057\u3066\u304a\u304f\u3053\u3068\u304c\u3067\u304d\u3001\u691c\u7d22\u6642\u306b\u306fEmbedding\u3092\u4f7f\u3063\u305f\u985e\u4f3c\u5ea6\u8a08\u7b97\uff08\u30b3\u30b5\u30a4\u30f3\u985e\u4f3c\u5ea6\u306a\u3069\uff09\u3092\u3059\u308b\u3060\u3051\u3067\u6e08\u307f\u307e\u3059\u3002\u4e00\u65b9\u3001Cross-Encoder\u306f\u985e\u4f3c\u5ea6\u8a08\u7b97\u306e\u30bf\u30a4\u30df\u30f3\u30b0\u3067\u5185\u90e8\u7684\u306bEmbedding\u751f\u6210\u306b\u3088\u308b\u985e\u4f3c\u5ea6\u7b97\u51fa\u3092\u884c\u3046\u5fc5\u8981\u304c\u3042\u308a\u3001\u3088\u308a\u591a\u304f\u306e\u6642\u9593\u304c\u304b\u304b\u3063\u3066\u3057\u307e\u3046\u3068\u3044\u3046\u308f\u3051\u3067\u3059\u3002<\/p>\n

\u6700\u5f8c\u306b\u3001Token Embedding\u304cRAG\u306e\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u7528\u9014\u3067\u4f7f\u7528\u3055\u308c\u308b\u4f8b\u306f\u3042\u308b\u306e\u304b\uff1f\u304c\u6c17\u306b\u306a\u3063\u305f\u306e\u3067\u8abf\u67fb\u3057\u305f\u3068\u3053\u308d\u3001\u3042\u307e\u308a\u6570\u306f\u306a\u3055\u305d\u3046\u3067\u3059\u304c\u3001\u4f8b\u3048\u3070\u30b9\u30da\u30eb\u30c1\u30a7\u30c3\u30af\u306e\u4f8b<\/a>\u304c\u898b\u3064\u304b\u308a\u307e\u3057\u305f\u3002\u78ba\u304b\u306b\u3001\u30b9\u30da\u30eb\u30c1\u30a7\u30c3\u30af\u306f\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u306e\u8868\u8a18\u63fa\u308c\u3092\u691c\u51fa\u3059\u308b\u5fc5\u8981\u304c\u3042\u308b\u305f\u3081\u3001Token Embedding\u306e\u597d\u4f8b\u3068\u8a00\u3048\u305d\u3046\u3067\u3059\u3002<\/p>\n

\u5192\u982d\u306e\u7591\u554f\u3078\u306e\u56de\u7b54<\/h4>\n

\u3053\u3053\u3067\u3001\u5192\u982d\u3067\u89e6\u308c\u305f\u7591\u554f\u306b\u7acb\u3061\u8fd4\u3063\u3066\u307f\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002\u7591\u554f\u306f\u300c\u306a\u305c\u30d9\u30af\u30c8\u30eb\u691c\u7d22\u306eEmbedding\u30e2\u30c7\u30eb\u3092\u6587\u5358\u4f4d\u304b\u3089\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u306b\u5909\u3048\u308b\u3068\u3001\u7279\u5b9a\u306e\u30ad\u30fc\u30ef\u30fc\u30c9\u3067\u306e\u691c\u7d22\u7cbe\u5ea6\u304c\u4e0a\u304c\u3063\u305f\u306e\u304b\uff1f\u300d\u3067\u3057\u305f\u3002<\/p>\n

\u3053\u306e\u7406\u7531\u306f\u3001\u3053\u3053\u307e\u3067\u8aac\u660e\u3057\u3066\u304d\u305f\u5185\u5bb9\u304b\u3089\u660e\u3089\u304b\u306b\u306a\u308a\u307e\u3059\u3002Sentence Embedding\u306f\u6587\u5168\u4f53\u306e\u610f\u5473\u3092\u5358\u4e00\u306eEmbedding\u3067\u8868\u73fe\u3059\u308b\u305f\u3081\u3001\u5404\u30c8\u30fc\u30af<\/a>\u30f3\u306e\u610f\u5473\u304c\u5e73\u5747\u5316\u3055\u308c\u3066\u3057\u307e\u3044\u307e\u3059\u3002\u305d\u306e\u7d50\u679c\u3001\u691c\u7d22\u6642\u306b\u91cd\u8981\u306a\u30ad\u30fc\u30ef\u30fc\u30c9\u304c\u542b\u307e\u308c\u3066\u3044\u3066\u3082\u3001\u305d\u306e\u91cd\u8981\u5ea6\u304c\u4ed6\u306e\u30c8\u30fc\u30af<\/a>\u30f3\u3068\u540c\u7b49\u306b\u6271\u308f\u308c\u3001\u57cb\u3082\u308c\u3066\u3057\u307e\u3046\u53ef\u80fd\u6027\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n

\u4e00\u65b9\u3001Token Embedding\u306f\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067Embedding\u3092\u751f\u6210\u3059\u308b\u305f\u3081\u3001\u91cd\u8981\u306a\u30ad\u30fc\u30ef\u30fc\u30c9\u306e\u610f\u5473\u3092\u500b\u5225\u306b\u4fdd\u6301\u3067\u304d\u307e\u3059\u3002\u691c\u7d22\u6642\u306b\u306f\u3001\u30af\u30a8\u30ea\u306b\u542b\u307e\u308c\u308b\u30ad\u30fc\u30ef\u30fc\u30c9\u3068\u3001\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u5185\u306e\u5bfe\u5fdc\u3059\u308b\u30c8\u30fc\u30af<\/a>\u30f3\u3068\u306e\u985e\u4f3c\u5ea6\u3092\u76f4\u63a5\u8a08\u7b97\u3067\u304d\u308b\u305f\u3081\u3001\u30ad\u30fc\u30ef\u30fc\u30c9\u304c\u30de\u30c3\u30c1\u3059\u308b\u7cbe\u5ea6\u304c\u9ad8\u304f\u306a\u308a\u307e\u3059\u3002<\/p>\n

\u3064\u307e\u308a\u3001\u5192\u982d\u306e\u30b1\u30fc\u30b9\u3067\u306f\u3001\u7279\u5b9a\u306e\u30ad\u30fc\u30ef\u30fc\u30c9\uff08\u5c02\u9580\u7528\u8a9e\u3084\u56fa\u6709\u540d\u8a5e\u306a\u3069\uff09\u304c\u691c\u7d22\u306e\u6210\u5426\u3092\u6c7a\u3081\u308b\u30dd\u30a4\u30f3\u30c8\u3060\u3063\u305f\u305f\u3081\u3001\u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067\u610f\u5473\u3092\u4fdd\u6301\u3067\u304d\u308bToken Embedding\u30e2\u30c7\u30eb\u306b\u5909\u66f4\u3059\u308b\u3053\u3068\u3067\u7cbe\u5ea6\u304c\u6539\u5584\u3057\u305f\u3001\u3068\u3044\u3046\u308f\u3051\u3067\u3059\u3002<\/p>\n

\u65e9\u898b\u8868<\/h4>\n

\u3053\u3053\u307e\u3067\u66f8\u3044\u305f\u5185\u5bb9\u3092\u4e00\u76ee\u3067\u601d\u3044\u51fa\u305b\u308b\u3088\u3046\u3001\u8868\u306b\u307e\u3068\u3081\u3066\u304a\u304d\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002\uff08\u3042\u304f\u307e\u3067\u79c1\u306e\u7406\u89e3\u3067\u3042\u308a\u3001\u7c21\u7565\u5316\u3057\u3066\u3044\u308b\u90e8\u5206\u304c\u3042\u308b\u70b9\u306b\u306f\u3054\u6ce8\u610f\u304f\u3060\u3055\u3044\uff09<\/p>\n

\n\n\n\n\n\n\n\n\n\n\n\n\n
\u6bd4\u8f03\u9805\u76ee <\/th>\n Sentence Embedding <\/th>\n Token Embedding <\/th>\n<\/tr>\n<\/thead>\n
\u751f\u6210\u5358\u4f4d<\/strong> <\/td>\n \u6587\u5358\u4f4d\u306b\u5bfe\u3057\u3066\u5358\u4e00\u306eEmbedding <\/td>\n \u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067Embedding\uff08\u5165\u529b\u30c7\u30fc\u30bf\u3092\u69cb\u6210\u3059\u308b\u30c8\u30fc\u30af<\/a>\u30f3\u306e\u6570\u3060\u3051\u751f\u6210\uff09 <\/td>\n<\/tr>\n
\u751f\u6210\u30d7\u30ed\u30bb\u30b9<\/strong> <\/td>\n \u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067Embedding\u3092\u751f\u6210\u3057\u3001\u30d7\u30fc\u30ea\u30f3\u30b0\uff08\u96c6\u7d04\uff09 <\/td>\n \u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067Embedding\u3092\u751f\u6210 <\/td>\n<\/tr>\n
\u691c\u7d22\u7cbe\u5ea6<\/strong> <\/td>\n \u6587\u5168\u4f53\u306e\u610f\u5473\u691c\u7d22\u306b\u9069\u3057\u3066\u3044\u308b <\/td>\n \u30c8\u30fc\u30af<\/a>\u30f3\u5358\u4f4d\u3067\u306e\u610f\u5473\u691c\u7d22\u304c\u53ef\u80fd\u3067\u3001\u6587\u5358\u4f4d\u3088\u308a\u9ad8\u7cbe\u5ea6 <\/td>\n<\/tr>\n
\u7279\u5b9a\u30ad\u30fc\u30ef\u30fc\u30c9<\/strong> <\/td>\n \u5404\u30c8\u30fc\u30af<\/a>\u30f3\u306e\u610f\u5473\u304c\u5e73\u5747\u5316\u3055\u308c\u3001\u91cd\u8981\u306a\u30ad\u30fc\u30ef\u30fc\u30c9\u304c\u57cb\u3082\u308c\u308b\u53ef\u80fd\u6027\u304c\u3042\u308b <\/td>\n \u91cd\u8981\u306a\u30ad\u30fc\u30ef\u30fc\u30c9\u306e\u610f\u5473\u3092\u500b\u5225\u306b\u4fdd\u6301\u3067\u304d\u3001\u30ad\u30fc\u30ef\u30fc\u30c9\u30de\u30c3\u30c1\u306e\u7cbe\u5ea6\u304c\u9ad8\u3044 <\/td>\n<\/tr>\n
\u4fdd\u5b58\u5bb9\u91cf<\/strong> <\/td>\n \u5c11\u306a\u3044\uff08\u6587\u3054\u3068\u306b1\u3064\u306eEmbedding\uff09 <\/td>\n \u591a\u3044\uff08\u30c8\u30fc\u30af<\/a>\u30f3\u306e\u6570\u3060\u3051Embedding\uff09 <\/td>\n<\/tr>\n
\u8a08\u7b97\u30b3\u30b9\u30c8<\/strong> <\/td>\n \u4f4e\u3044 <\/td>\n \u9ad8\u3044 <\/td>\n<\/tr>\n
\u691c\u7d22\u901f\u5ea6<\/strong> <\/td>\n \u901f\u3044 <\/td>\n \u9045\u3044 <\/td>\n<\/tr>\n
\u4e3b\u306a\u7528\u9014<\/strong> <\/td>\n RAG\u306e\u30d9\u30af\u30c8\u30eb\u691c\u7d22\uff08\u30c7\u30d5\u30a1\u30af\u30c8\u30b9\u30bf\u30f3\u30c0\u30fc\u30c9<\/a>\uff09 <\/td>\n \u30ea\u30e9\u30f3\u30af\u51e6\u7406\u3001\u30b9\u30da\u30eb\u30c1\u30a7\u30c3\u30af\u306a\u3069 <\/td>\n<\/tr>\n
\u4ee3\u8868\u30e2\u30c7\u30eb<\/strong> <\/td>\n SBERT\uff08Sentence BERT\uff09 <\/td>\n ColBERT <\/td>\n<\/tr>\n<\/tbody>\n<\/table><\/div>\n

DataCloud\u306e\u5834\u5408<\/h3>\n

\u3067\u306f\u3001Salesforce<\/a>\u306eDataCloud\u306e\u5834\u5408\u3001\u3069\u306e\u3088\u3046\u306aEmbedding\u30e2\u30c7\u30eb\u3092\u4f7f\u7528\u3067\u304d\u308b\u306e\u304b\u78ba\u8a8d\u3057\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n

\u5229\u7528\u53ef\u80fd\u306aText Embedding\u30e2\u30c7\u30eb\u306f\u4ee5\u4e0b\u306e2\u3064\u3067\u3059\u3002<\/p>\n