{"id":874,"date":"2025-07-28T01:19:00","date_gmt":"2025-07-28T01:19:00","guid":{"rendered":"https:\/\/ouyangminwei.com\/?p=874"},"modified":"2025-07-18T09:19:19","modified_gmt":"2025-07-18T09:19:19","slug":"encoder-decoder-encoder-decoder-only-1","status":"publish","type":"post","link":"https:\/\/ouyangminwei.com\/index.php\/2025\/07\/28\/encoder-decoder-encoder-decoder-only-1\/","title":{"rendered":"encoder\/decoder\/encoder-decoder only? 1"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u5c0e\u8ad6\uff1a\u65b0\u7d00\u5143\u7684\u9ece\u660e\u8207\u5049\u5927\u7684\u5206\u6b67<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">Transformer\u4e4b\u524d\u7684\u4e16\u754c\uff1a\u8a18\u61b6\u8207\u4e26\u884c\u5316\u7684\u5371\u6a5f<\/h3>\n\n\n\n<p>\u57282017\u5e74\u4e4b\u524d\uff0c\u81ea\u7136\u8a9e\u8a00\u8655\u7406\uff08NLP\uff09\u9818\u57df\u7531\u905e\u6b78\u795e\u7d93\u7db2\u7d61\uff08Recurrent Neural Networks, RNNs\uff09\u53ca\u5176\u66f4\u8907\u96dc\u7684\u8b8a\u9ad4\uff0c\u5982\u9577\u77ed\u671f\u8a18\u61b6\uff08Long Short-Term Memory, LSTM\uff09\u7db2\u7d61\u6240\u4e3b\u5c0e <sup><\/sup>\u3002\u9019\u4e9b\u67b6\u69cb\u5728\u8a2d\u8a08\u4e0a\u662f\u70ba\u4e86\u8655\u7406\u5e8f\u5217\u6578\u64da\uff0c\u5b83\u5011\u4e00\u6b21\u8655\u7406\u4e00\u500b\u5143\u7d20\uff0c\u4e26\u901a\u904e\u4e00\u500b\u5167\u90e8\u72c0\u614b\uff08\u6216\u300c\u8a18\u61b6\u300d\uff09\u5c07\u4fe1\u606f\u5f9e\u5e8f\u5217\u7684\u4e00\u500b\u6642\u9593\u6b65\u50b3\u905e\u5230\u4e0b\u4e00\u500b\u6642\u9593\u6b65 <sup><\/sup>\u3002\u9019\u7a2e\u5e8f\u5217\u5316\u7684\u8655\u7406\u65b9\u5f0f\u5728\u7406\u8ad6\u4e0a\u4f7f\u5176\u975e\u5e38\u9069\u5408\u8a9e\u8a00\uff0c\u56e0\u70ba\u8a9e\u8a00\u672c\u8eab\u5c31\u662f\u4e00\u500b\u6709\u5e8f\u7684\u5e8f\u5217\u3002\u7136\u800c\uff0c\u9019\u7a2e\u8a2d\u8a08\u4e5f\u5e36\u4f86\u4e86\u5169\u500b\u6839\u672c\u6027\u7684\u3001\u96e3\u4ee5\u514b\u670d\u7684\u9650\u5236\uff0c\u70ba\u4e00\u5834\u5373\u5c07\u5230\u4f86\u7684\u9769\u547d\u57cb\u4e0b\u4e86\u4f0f\u7b46\u3002 &nbsp;<\/p>\n\n\n\n<p>\u7b2c\u4e00\u500b\u9650\u5236\u662f\u5176\u56fa\u6709\u7684\u5e8f\u5217\u5316\u7279\u6027\u3002RNNs\u548cLSTMs\u5fc5\u9808\u6309\u9806\u5e8f\u8655\u7406\u8f38\u5165\uff0c\u4e00\u500b\u6642\u9593\u6b65\u7684\u8a08\u7b97\u4f9d\u8cf4\u65bc\u524d\u4e00\u500b\u6642\u9593\u6b65\u7684\u7d50\u679c\u3002\u9019\u7a2e\u4f9d\u8cf4\u6027\u4f7f\u5f97\u5b83\u5011\u7121\u6cd5\u6709\u6548\u5229\u7528\u73fe\u4ee3GPU\u548cTPU\u7b49\u5927\u898f\u6a21\u4e26\u884c\u8a08\u7b97\u786c\u4ef6\u7684\u512a\u52e2 <sup><\/sup>\u3002\u7576\u9762\u5c0d\u52d5\u8f12\u6578\u5341\u5104\u8a5e\u5f59\u7684\u9f90\u5927\u6578\u64da\u96c6\u6642\uff0c\u9019\u7a2e\u7121\u6cd5\u4e26\u884c\u5316\u7684\u74f6\u9838\u6975\u5927\u5730\u9650\u5236\u4e86\u6a21\u578b\u7684\u8a13\u7df4\u901f\u5ea6\u548c\u53ef\u64f4\u5c55\u6027\uff0c\u6210\u70ba\u4e86\u8a72\u9818\u57df\u767c\u5c55\u7684\u4e00\u9053\u7121\u5f62\u58c1\u58d8\u3002 &nbsp;<\/p>\n\n\n\n<p>\u7b2c\u4e8c\u500b\uff0c\u4e5f\u662f\u66f4\u70ba\u6839\u672c\u7684\u9650\u5236\uff0c\u662f\u5b83\u5011\u96e3\u4ee5\u6355\u6349\u9577\u8ddd\u96e2\u4f9d\u8cf4\u95dc\u4fc2\uff08long-range dependencies\uff09\u3002\u7406\u8ad6\u4e0a\uff0cLSTM\u901a\u904e\u5176\u9580\u63a7\u6a5f\u5236\uff08gating mechanism\uff09\u65e8\u5728\u7de9\u89e3RNN\u4e2d\u7684\u68af\u5ea6\u6d88\u5931\u554f\u984c\uff08vanishing gradient problem\uff09\uff0c\u4f46\u5be6\u969b\u4e0a\uff0c\u7576\u5e8f\u5217\u975e\u5e38\u9577\u6642\uff0c\u4f86\u81ea\u5e8f\u5217\u65e9\u671f\u90e8\u5206\u7684\u4fe1\u606f\u5728\u50b3\u905e\u904e\u7a0b\u4e2d\u4ecd\u7136\u6703\u9010\u6f38\u8870\u6e1b\u6216\u5931\u771f <sup><\/sup>\u3002\u9019\u610f\u5473\u8457\u6a21\u578b\u5f88\u96e3\u5c07\u53e5\u5b50\u672b\u5c3e\u7684\u4e00\u500b\u8a5e\u8207\u53e5\u5b50\u958b\u982d\u7684\u4e00\u500b\u8a5e\u5efa\u7acb\u6709\u610f\u7fa9\u7684\u806f\u7e6b\uff0c\u9019\u5c0d\u65bc\u7406\u89e3\u8907\u96dc\u7684\u8a9e\u6cd5\u7d50\u69cb\u548c\u8a9e\u7fa9\u95dc\u4fc2\u81f3\u95dc\u91cd\u8981\u3002 &nbsp;<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u57282017\u5e74\uff0cNLP\u9818\u57df\u6b63\u8655\u65bc\u4e00\u500b\u5341\u5b57\u8def\u53e3\uff1a\u4e00\u65b9\u9762\uff0c\u6578\u64da\u548c\u8a08\u7b97\u8cc7\u6e90\u5448\u6307\u6578\u7d1a\u589e\u9577\uff1b\u53e6\u4e00\u65b9\u9762\uff0c\u4e3b\u6d41\u7684\u5e8f\u5217\u8655\u7406\u67b6\u69cb\u5728\u6839\u672c\u4e0a\u7121\u6cd5\u5145\u5206\u5229\u7528\u9019\u4e9b\u8cc7\u6e90\u4f86\u89e3\u6c7a\u8a9e\u8a00\u7406\u89e3\u7684\u6838\u5fc3\u6311\u6230\u3002\u6574\u500b\u9818\u57df\u8feb\u5207\u9700\u8981\u4e00\u7a2e\u65b0\u7684\u7bc4\u5f0f\uff0c\u4e00\u7a2e\u80fd\u5920\u64fa\u812b\u905e\u6b78\u675f\u7e1b\u3001\u64c1\u62b1\u4e26\u884c\u8a08\u7b97\uff0c\u4e26\u80fd\u5e73\u7b49\u5730\u770b\u5f85\u5e8f\u5217\u4e2d\u6240\u6709\u5143\u7d20\u4e4b\u9593\u95dc\u4fc2\u7684\u67b6\u69cb\u3002\u9019\u5834\u5371\u6a5f\u70baTransformer\u7684\u8a95\u751f\u5275\u9020\u4e86\u5b8c\u7f8e\u7684\u6b77\u53f2\u689d\u4ef6\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u300cAttention Is All You Need\u300d\u7684\u5947\u9ede\u6642\u523b<\/h3>\n\n\n\n<p>2017\u5e74\uff0c\u5728\u795e\u7d93\u4fe1\u606f\u8655\u7406\u7cfb\u7d71\u5927\u6703\uff08NeurIPS\uff09\u4e0a\uff0c\u4f86\u81eaGoogle\u7684\u7814\u7a76\u4eba\u54e1\u767c\u8868\u4e86\u4e00\u7bc7\u540d\u70ba\u300aAttention Is All You Need\u300b\u7684\u958b\u5275\u6027\u8ad6\u6587\uff0c\u5fb9\u5e95\u6539\u8b8a\u4e86NLP\u4e43\u81f3\u6574\u500b\u4eba\u5de5\u667a\u80fd\u9818\u57df\u7684\u683c\u5c40 <sup><\/sup>\u3002\u9019\u7bc7\u8ad6\u6587\u4e0d\u50c5\u50c5\u662f\u4e00\u6b21\u6f38\u9032\u5f0f\u7684\u6539\u9032\uff0c\u800c\u662f\u4e00\u5834\u5fb9\u5e95\u7684\u9769\u547d\u3002\u5b83\u63d0\u51fa\u4e86\u4e00\u7a2e\u5168\u65b0\u7684\u3001\u5b8c\u5168\u62cb\u68c4\u4e86\u905e\u6b78\u548c\u5377\u7a4d\u7684\u7db2\u7d61\u67b6\u69cb\u2014\u2014Transformer <sup><\/sup>\u3002\u9019\u7bc7\u8ad6\u6587\u7684\u6a19\u984c\u672c\u8eab\u5c31\u662f\u4e00\u500b\u5927\u81bd\u7684\u5ba3\u8a00\uff0c\u5b83\u5ba3\u7a31\uff0c\u50c5\u6191\u300c\u6ce8\u610f\u529b\u6a5f\u5236\u300d\uff08attention mechanism\uff09\u5c31\u8db3\u4ee5\u69cb\u5efa\u51fa\u5f37\u5927\u7684\u5e8f\u5217\u8655\u7406\u6a21\u578b\u3002 &nbsp;<\/p>\n\n\n\n<p>Transformer\u67b6\u69cb\u88ab\u8996\u70ba\u73fe\u4ee3\u6240\u6709\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u5171\u540c\u7956\u5148\uff0c\u662f\u4e09\u5927\u67b6\u69cb\u738b\u671d\u7684\u8d77\u6e90\u3002\u5b83\u7684\u8a2d\u8a08\u521d\u8877\u662f\u70ba\u4e86\u89e3\u6c7a\u6a5f\u5668\u7ffb\u8b6f\u4efb\u52d9\uff0c\u4f46\u5176\u6838\u5fc3\u7d44\u4ef6\u7684\u901a\u7528\u6027\u548c\u5f37\u5927\u80fd\u529b\u4f7f\u5176\u8fc5\u901f\u64f4\u5c55\u5230\u5e7e\u4e4e\u6240\u6709NLP\u4efb\u52d9\u4e2d\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u539f\u59cbTransformer\u7684\u6838\u5fc3\u7d44\u4ef6<\/h4>\n\n\n\n<p>\u539f\u59cb\u7684Transformer\u6a21\u578b\u662f\u4e00\u500b\u5b8c\u6574\u7684\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\uff08Encoder-Decoder\uff09\u67b6\u69cb\uff0c\u5176\u8a2d\u8a08\u7cbe\u5de7\uff0c\u70ba\u5f8c\u4f86\u7684\u67b6\u69cb\u5206\u6b67\u63d0\u4f9b\u4e86\u6240\u6709\u5fc5\u8981\u7684\u57fa\u56e0 <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7de8\u78bc\u5668\u8207\u89e3\u78bc\u5668\u5806\u758a\uff08The Encoder and Decoder Stacks\uff09<\/strong>\uff1a\u539f\u59cb\u6a21\u578b\u7531N=6\u500b\u76f8\u540c\u7684\u7de8\u78bc\u5668\u5c64\u5806\u758a\u800c\u6210\uff0c\u4ee5\u53caN=6\u500b\u76f8\u540c\u7684\u89e3\u78bc\u5668\u5c64\u5806\u758a\u800c\u6210\u3002\u6bcf\u500b\u7de8\u78bc\u5668\u5c64\u5305\u542b\u5169\u500b\u5b50\u5c64\uff1a\u4e00\u500b\u591a\u982d\u81ea\u8a3b\u610f\u529b\u6a5f\u5236\u548c\u4e00\u500b\u7c21\u55ae\u7684\u4f4d\u7f6e\u5168\u9023\u63a5\u524d\u994b\u7db2\u7d61\u3002\u6bcf\u500b\u89e3\u78bc\u5668\u5c64\u5247\u5728\u7de8\u78bc\u5668\u5c64\u7684\u57fa\u790e\u4e0a\uff0c\u63d2\u5165\u4e86\u7b2c\u4e09\u500b\u5b50\u5c64\uff0c\u7528\u65bc\u5c0d\u7de8\u78bc\u5668\u7684\u8f38\u51fa\u57f7\u884c\u591a\u982d\u6ce8\u610f\u529b \u3002\u9019\u7a2e\u5806\u758a\u7d50\u69cb\u5141\u8a31\u6a21\u578b\u9010\u6b65\u69cb\u5efa\u66f4\u8907\u96dc\u3001\u66f4\u62bd\u8c61\u7684\u6578\u64da\u8868\u793a\u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u81ea\u8a3b\u610f\u529b\u6a5f\u5236\uff08Self-Attention \/ Intra-Attention\uff09<\/strong>\uff1a\u9019\u662fTransformer\u67b6\u69cb\u7684\u57fa\u77f3\uff0c\u4e5f\u662f\u5176\u64fa\u812b\u905e\u6b78\u675f\u7e1b\u7684\u95dc\u9375 \u3002\u81ea\u8a3b\u610f\u529b\u6a5f\u5236\u5141\u8a31\u6a21\u578b\u5728\u8655\u7406\u4e00\u500b\u5e8f\u5217\u6642\uff0c\u540c\u6642\u6b0a\u8861\u5e8f\u5217\u4e2d\u6240\u6709\u5176\u4ed6\u8a5e\u7684\u91cd\u8981\u6027\u3002\u5b83\u901a\u904e\u5c07\u6bcf\u500b\u8f38\u5165\u8a5e\u8f49\u63db\u70ba\u4e09\u500b\u4e0d\u540c\u7684\u5411\u91cf\u4f86\u5be6\u73fe\u9019\u4e00\u9ede\uff1a\u67e5\u8a62\uff08Query, Q\uff09\u3001\u9375\uff08Key, K\uff09\u548c\u503c\uff08Value, V\uff09\u3002\u6a21\u578b\u901a\u904e\u8a08\u7b97\u4e00\u500b\u7279\u5b9a\u8a5e\u7684Q\u5411\u91cf\u8207\u6240\u6709\u5176\u4ed6\u8a5e\u7684K\u5411\u91cf\u4e4b\u9593\u7684\u9ede\u7a4d\u4f86\u78ba\u5b9a\u6ce8\u610f\u529b\u5206\u6578\uff0c\u9019\u500b\u5206\u6578\u6c7a\u5b9a\u4e86\u5728\u751f\u6210\u8a72\u8a5e\u7684\u65b0\u8868\u793a\u6642\uff0c\u61c9\u8a72\u7d66\u4e88\u5176\u4ed6\u8a5e\u7684V\u5411\u91cf\u591a\u5c11\u6b0a\u91cd \u3002\u5176\u6838\u5fc3\u8a08\u7b97\u516c\u5f0f\u70ba\uff1a \u00a0 Attention(Q,K,V)=softmax(dk\u200b<img decoding=\"async\" src=\"\">\u200bQKT\u200b)V\u5176\u4e2ddk\u200b\u662f\u9375\u5411\u91cf\u7684\u7dad\u5ea6\uff0c\u7528\u65bc\u7e2e\u653e\u9ede\u7a4d\uff0c\u9632\u6b62\u68af\u5ea6\u904e\u5c0f \u3002\u7531\u65bc\u9019\u7a2e\u6a5f\u5236\u76f4\u63a5\u8a08\u7b97\u5e8f\u5217\u4e2d\u4efb\u610f\u5169\u500b\u4f4d\u7f6e\u4e4b\u9593\u7684\u95dc\u4fc2\uff0c\u56e0\u6b64\u5b83\u5c0d\u8ddd\u96e2\u4e0d\u654f\u611f\uff0c\u5b8c\u7f8e\u5730\u89e3\u6c7a\u4e86RNN\u7684\u9577\u8ddd\u96e2\u4f9d\u8cf4\u554f\u984c\u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u591a\u982d\u6ce8\u610f\u529b\u6a5f\u5236\uff08Multi-Head Attention\uff09<\/strong>\uff1a\u70ba\u4e86\u589e\u5f37\u81ea\u8a3b\u610f\u529b\u7684\u80fd\u529b\uff0cTransformer\u4e26\u884c\u5730\u904b\u884c\u591a\u500b\u6ce8\u610f\u529b\u8a08\u7b97\uff0c\u5373\u300c\u591a\u982d\u300d\u3002\u5b83\u4e0d\u662f\u53ea\u9032\u884c\u4e00\u6b21\u55ae\u4e00\u7684\u6ce8\u610f\u529b\u8a08\u7b97\uff0c\u800c\u662f\u5c07Q\u3001K\u3001V\u5411\u91cf\u7dda\u6027\u6295\u5f71\u5230\u4e0d\u540c\u7684\u5b50\u7a7a\u9593\u4e2d\uff0c\u4e26\u5728\u6bcf\u500b\u5b50\u7a7a\u9593\u4e2d\u4e26\u884c\u57f7\u884c\u6ce8\u610f\u529b\u51fd\u6578\u3002\u7136\u5f8c\uff0c\u5c07\u6240\u6709\u982d\u7684\u8f38\u51fa\u62fc\u63a5\u8d77\u4f86\u4e26\u518d\u6b21\u9032\u884c\u7dda\u6027\u6295\u5f71\u3002\u9019\u4f7f\u5f97\u6a21\u578b\u80fd\u5920\u540c\u6642\u95dc\u6ce8\u4f86\u81ea\u4e0d\u540c\u8868\u793a\u5b50\u7a7a\u9593\u7684\u4fe1\u606f\uff0c\u5f9e\u800c\u6355\u6349\u5230\u66f4\u8c50\u5bcc\u3001\u66f4\u591a\u6a23\u7684\u8a9e\u7fa9\u95dc\u4fc2 \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u4f4d\u7f6e\u7de8\u78bc\uff08Positional Encoding\uff09<\/strong>\uff1a\u7531\u65bc\u81ea\u8a3b\u610f\u529b\u6a5f\u5236\u672c\u8eab\u662f\u4f4d\u7f6e\u7121\u95dc\u7684\uff08position-agnostic\uff09\uff0c\u5b83\u5c07\u8f38\u5165\u8996\u70ba\u4e00\u500b\u7121\u5e8f\u7684\u96c6\u5408\u3002\u70ba\u4e86\u8b93\u6a21\u578b\u80fd\u5920\u5229\u7528\u5e8f\u5217\u7684\u9806\u5e8f\u4fe1\u606f\uff0c\u7814\u7a76\u4eba\u54e1\u5728\u8f38\u5165\u5d4c\u5165\u4e2d\u52a0\u5165\u4e86\u300c\u4f4d\u7f6e\u7de8\u78bc\u300d\u3002\u9019\u4e9b\u7de8\u78bc\u662f\u6839\u64da\u8a5e\u5728\u5e8f\u5217\u4e2d\u7684\u7d55\u5c0d\u6216\u76f8\u5c0d\u4f4d\u7f6e\u8a08\u7b97\u51fa\u7684\u5411\u91cf\uff0c\u5b83\u5011\u88ab\u52a0\u5230\u8a5e\u5d4c\u5165\u4e0a\uff0c\u70ba\u6a21\u578b\u63d0\u4f9b\u4e86\u95dc\u65bc\u8a5e\u5e8f\u7684\u95dc\u9375\u4fe1\u606f \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u4e26\u884c\u5316\uff08Parallelization\uff09<\/strong>\uff1a\u901a\u904e\u5b8c\u5168\u62cb\u68c4\u905e\u6b78\uff0cTransformer\u7684\u8a08\u7b97\u53ef\u4ee5\u5728\u5e8f\u5217\u7dad\u5ea6\u4e0a\u5927\u898f\u6a21\u4e26\u884c\u5316\u3002\u5728\u8655\u7406\u4e00\u500b\u5e8f\u5217\u6642\uff0c\u6240\u6709\u4f4d\u7f6e\u7684Q\u3001K\u3001V\u5411\u91cf\u4ee5\u53ca\u5b83\u5011\u4e4b\u9593\u7684\u6ce8\u610f\u529b\u5206\u6578\u90fd\u53ef\u4ee5\u540c\u6642\u8a08\u7b97\u3002\u9019\u8207RNN\u7684\u5e8f\u5217\u5316\u8655\u7406\u5f62\u6210\u4e86\u9bae\u660e\u5c0d\u6bd4\uff0c\u6975\u5927\u5730\u63d0\u9ad8\u4e86\u8a13\u7df4\u6548\u7387\uff0c\u4f7f\u5f97\u5728\u7576\u6642\u524d\u6240\u672a\u6709\u7684\u5927\u898f\u6a21\u6578\u64da\u96c6\u4e0a\u8a13\u7df4\u5de8\u578b\u6a21\u578b\u6210\u70ba\u53ef\u80fd \u3002 \u00a0<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u5049\u5927\u7684\u5206\u6b67\uff1a\u67b6\u69cb\u9053\u8def\u4e0a\u7684\u5c94\u53e3<\/h3>\n\n\n\n<p>Transformer\u8ad6\u6587\u7684\u767c\u8868\uff0c\u5982\u540c\u5728\u8a9e\u8a00\u6a21\u578b\u9818\u57df\u6295\u4e0b\u4e86\u4e00\u9846\u5947\u9ede\u70b8\u5f48\uff0c\u5176\u80fd\u91cf\u91cb\u653e\u5f8c\u4e26\u672a\u5f62\u6210\u55ae\u4e00\u7684\u767c\u5c55\u8def\u5f91\uff0c\u800c\u662f\u8fc5\u901f\u5f15\u767c\u4e86\u4e00\u5834\u300c\u5049\u5927\u7684\u5206\u6b67\u300d\uff08The Great Divergence\uff09\u3002\u539f\u59cb\u7684Transformer\u662f\u4e00\u500b\u7528\u65bc\u6a5f\u5668\u7ffb\u8b6f\u7684\u5b8c\u6574\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u7cfb\u7d71\uff0c\u4f46\u7814\u7a76\u754c\u5f88\u5feb\u610f\u8b58\u5230\uff0c\u5176\u7de8\u78bc\u5668\u548c\u89e3\u78bc\u5668\u90e8\u5206\u53ef\u4ee5\u88ab\u62c6\u5206\u958b\u4f86\uff0c\u7528\u65bc\u89e3\u6c7a\u4e0d\u540c\u985e\u578b\u7684\u554f\u984c\u3002\u9019\u5834\u5206\u6b67\u4e26\u975e\u5076\u7136\u7684\u5de5\u7a0b\u6c7a\u7b56\uff0c\u800c\u662f\u4e00\u5834\u6df1\u523b\u7684\u54f2\u5b78\u5206\u91ce\uff0c\u6e90\u65bc\u4e0d\u540c\u7814\u7a76\u5718\u968a\u5c0d\u4eba\u5de5\u667a\u80fd\u672a\u4f86\u767c\u5c55\u8def\u5f91\u7684\u4e0d\u540c\u9858\u666f\u3002<\/p>\n\n\n\n<p>\u4e00\u65b9\u9762\uff0c\u4ee5Google\u7684\u7814\u7a76\u4eba\u54e1\u70ba\u4ee3\u8868\u7684\u9663\u71df\uff0c\u5176\u6838\u5fc3\u76ee\u6a19\u662f\u63d0\u5347\u5c0d\u8a9e\u8a00\u7684<strong>\u6df1\u5ea6\u7406\u89e3<\/strong>\uff0c\u4ee5\u670d\u52d9\u65bc\u641c\u7d22\u3001\u4fe1\u606f\u6aa2\u7d22\u7b49\u4efb\u52d9 <sup><\/sup>\u3002\u5c0d\u65bc\u9019\u4e9b\u61c9\u7528\u800c\u8a00\uff0c\u6700\u95dc\u9375\u7684\u80fd\u529b\u662f\u80fd\u5920\u5c0d\u8f38\u5165\u6587\u672c\uff08\u4f8b\u5982\u4e00\u500b\u641c\u7d22\u67e5\u8a62\uff09\u9032\u884c\u5168\u9762\u3001\u96d9\u5411\u7684\u8a9e\u5883\u7406\u89e3\u3002\u9019\u9a45\u4f7f\u4ed6\u5011\u4fdd\u7559\u4e26\u5f37\u5316\u4e86Transformer\u7684\u7de8\u78bc\u5668\u90e8\u5206\uff0c\u56e0\u70ba\u7de8\u78bc\u5668\u7684\u96d9\u5411\u81ea\u8a3b\u610f\u529b\u6a5f\u5236\u6b63\u662f\u70ba\u6b64\u800c\u751f\u3002\u4ed6\u5011\u62cb\u68c4\u4e86\u89e3\u78bc\u5668\uff0c\u5c08\u6ce8\u65bc\u6253\u9020\u80fd\u5920\u751f\u6210\u6700\u8c50\u5bcc\u3001\u6700\u7cbe\u78ba\u6587\u672c\u8868\u793a\u7684\u300c\u7406\u89e3\u5f15\u64ce\u300d\u3002\u9019\u689d\u8def\u5f91\u50ac\u751f\u4e86\u7b2c\u4e00\u500b\u5049\u5927\u7684\u738b\u671d\u2014\u2014 &nbsp;<\/p>\n\n\n\n<p><strong>\u7de8\u78bc\u5668-Only\u738b\u671d<\/strong>\u3002<\/p>\n\n\n\n<p>\u53e6\u4e00\u65b9\u9762\uff0c\u4ee5OpenAI\u70ba\u4ee3\u8868\u7684\u9663\u71df\uff0c\u5176\u76ee\u6a19\u5247\u66f4\u70ba\u5b8f\u5927\uff0c\u4ed6\u5011\u81f4\u529b\u65bc\u63a2\u7d22<strong>\u901a\u7528\u4eba\u5de5\u667a\u80fd\uff08AGI\uff09\u7684\u8def\u5f91\uff0c\u800c\u5f37\u5927\u7684\u751f\u6210\u80fd\u529b<\/strong>\u88ab\u8996\u70ba\u5176\u4e2d\u7684\u95dc\u9375\u4e00\u6b65 <sup><\/sup>\u3002\u4ed6\u5011\u8a8d\u70ba\uff0c\u4e00\u500b\u80fd\u5920\u6839\u64da\u4efb\u610f\u63d0\u793a\u9023\u8cab\u5730\u3001\u5275\u9020\u6027\u5730\u751f\u6210\u6587\u672c\u7684\u6a21\u578b\uff0c\u662f\u901a\u5f80\u66f4\u9ad8\u5c64\u6b21\u667a\u80fd\u7684\u57fa\u790e\u3002\u70ba\u6b64\uff0c\u4ed6\u5011\u9078\u64c7\u4e86Transformer\u7684\u89e3\u78bc\u5668\u90e8\u5206\uff0c\u56e0\u70ba\u5176\u81ea\u56de\u6b78\uff08auto-regressive\uff09\u548c\u56e0\u679c\u6ce8\u610f\u529b\uff08causal attention\uff09\u7684\u7279\u6027\u5929\u7136\u9069\u5408\u65bc\u751f\u6210\u4efb\u52d9\u3002\u4ed6\u5011\u62cb\u68c4\u4e86\u7de8\u78bc\u5668\uff0c\u5168\u529b\u6253\u9020\u4e00\u500b\u7d14\u7cb9\u7684\u300c\u751f\u6210\u5f15\u64ce\u300d\u3002\u9019\u689d\u8def\u5f91\u5247\u958b\u5275\u4e86\u7b2c\u4e8c\u500b\u5049\u5927\u7684\u738b\u671d\u2014\u2014 &nbsp;<\/p>\n\n\n\n<p><strong>\u89e3\u78bc\u5668-Only\u738b\u671d<\/strong>\u3002<\/p>\n\n\n\n<p>\u6700\u521d\u7684\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u67b6\u69cb\u4e26\u672a\u5c31\u6b64\u6d88\u4ea1\u3002\u5728\u5169\u5927\u5c08\u7cbe\u738b\u671d\u5404\u81ea\u767c\u5c55\u7684\u540c\u6642\uff0c\u4e00\u4e9b\u7814\u7a76\u8005\u8a66\u5716\u91cd\u65b0\u6574\u5408\u5169\u8005\u7684\u512a\u52e2\uff0c\u6253\u9020\u51fa\u65e2\u80fd\u6df1\u5ea6\u7406\u89e3\u8f38\u5165\u53c8\u80fd\u9748\u6d3b\u751f\u6210\u8f38\u51fa\u7684\u901a\u7528\u6846\u67b6\u3002\u9019\u4ee3\u8868\u4e86\u4e00\u7a2e\u300c\u518d\u878d\u5408\u300d\u7684\u8da8\u52e2\uff0c\u5f62\u6210\u4e86\u7b2c\u4e09\u80a1\u529b\u91cf\uff0c\u5373<strong>\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u6587\u85dd\u5fa9\u8208<\/strong>\u3002<\/p>\n\n\n\n<p>\u9019\u4e09\u5927\u738b\u671d\u7684\u8208\u8870\u3001\u6f14\u8b8a\u8207\u7af6\u722d\uff0c\u5171\u540c\u8b5c\u5beb\u4e86\u73fe\u4ee3\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u7684\u5b8f\u5049\u53f2\u8a69\u3002\u4e0b\u8868\u70ba\u9019\u6bb5\u6b77\u53f2\u4e2d\u7684\u4e00\u4e9b\u91cc\u7a0b\u7891\u5f0f\u6a21\u578b\u63d0\u4f9b\u4e86\u4e00\u500b\u7c21\u660e\u7684\u6642\u9593\u7dda\u3002<\/p>\n\n\n\n<p><strong>\u88681\uff1a\u91cc\u7a0b\u7891\u5f0fTransformer\u6a21\u578b\u5e74\u8868<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><td>\u6a21\u578b\u540d\u7a31<\/td><td>\u5e74\u4efd<\/td><td>\u4e3b\u8981\u958b\u767c\u8005<\/td><td>\u67b6\u69cb\u985e\u578b<\/td><td>\u95dc\u9375\u5275\u65b0\/\u8ca2\u737b<\/td><\/tr><\/thead><tbody><tr><td><strong>Transformer<\/strong><\/td><td>2017<\/td><td>Google<\/td><td>Encoder-Decoder<\/td><td>\u9996\u6b21\u63d0\u51fa\u5b8c\u5168\u57fa\u65bc\u6ce8\u610f\u529b\u7684\u67b6\u69cb\uff0c\u62cb\u68c4\u905e\u6b78\uff0c\u5be6\u73fe\u5927\u898f\u6a21\u4e26\u884c\u5316 <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>BERT<\/strong><\/td><td>2018<\/td><td>Google<\/td><td>Encoder-Only<\/td><td>\u5f15\u5165\u96d9\u5411\u4e0a\u4e0b\u6587\u7406\u89e3\u548cMLM\u9810\u8a13\u7df4\u4efb\u52d9\uff0c\u5fb9\u5e95\u6539\u8b8aNLU\u9818\u57df <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>GPT-1<\/strong><\/td><td>2018<\/td><td>OpenAI<\/td><td>Decoder-Only<\/td><td>\u9a57\u8b49\u4e86\u751f\u6210\u5f0f\u9810\u8a13\u7df4\u5728Decoder-Only\u67b6\u69cb\u4e0a\u7684\u6709\u6548\u6027 <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>GPT-2<\/strong><\/td><td>2019<\/td><td>OpenAI<\/td><td>Decoder-Only<\/td><td>\u5c55\u793a\u4e86\u5927\u898f\u6a21\u6a21\u578b\u5728\u96f6\u6a23\u672c\uff08Zero-shot\uff09\u8a2d\u7f6e\u4e0b\u7684\u5f37\u5927\u751f\u6210\u80fd\u529b <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>RoBERTa<\/strong><\/td><td>2019<\/td><td>Facebook AI<\/td><td>Encoder-Only<\/td><td>\u512a\u5316\u4e86BERT\u7684\u8a13\u7df4\u7b56\u7565\uff0c\u8b49\u660eBERT\u88ab\u56b4\u91cd\u8a13\u7df4\u4e0d\u8db3 <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>ALBERT<\/strong><\/td><td>2019<\/td><td>Google<\/td><td>Encoder-Only<\/td><td>\u901a\u904e\u53c3\u6578\u5171\u4eab\u7b49\u6280\u8853\uff0c\u6975\u5927\u5730\u63d0\u9ad8\u4e86BERT\u7684\u53c3\u6578\u6548\u7387 <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>T5<\/strong><\/td><td>2019<\/td><td>Google<\/td><td>Encoder-Decoder<\/td><td>\u63d0\u51fa\u300cText-to-Text\u300d\u7d71\u4e00\u6846\u67b6\uff0c\u5c07\u6240\u6709NLP\u4efb\u52d9\u8996\u70ba\u5e8f\u5217\u751f\u6210\u554f\u984c <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>BART<\/strong><\/td><td>2019<\/td><td>Facebook AI<\/td><td>Encoder-Decoder<\/td><td>\u7d50\u5408BERT\u7684\u96d9\u5411\u7de8\u78bc\u5668\u548cGPT\u7684\u81ea\u56de\u6b78\u89e3\u78bc\u5668\uff0c\u63a1\u7528\u53bb\u566a\u81ea\u7de8\u78bc\u5668\u76ee\u6a19 <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>GPT-3<\/strong><\/td><td>2020<\/td><td>OpenAI<\/td><td>Decoder-Only<\/td><td>\u6191\u85c91750\u5104\u53c3\u6578\u7684\u5de8\u5927\u898f\u6a21\uff0c\u5c55\u73fe\u51fa\u9a5a\u4eba\u7684\u4e0a\u4e0b\u6587\u5b78\u7fd2\uff08In-context Learning\uff09\u80fd\u529b <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>PaLM<\/strong><\/td><td>2022<\/td><td>Google<\/td><td>Decoder-Only<\/td><td>\u5c07\u6a21\u578b\u898f\u6a21\u63a8\u54115400\u5104\u53c3\u6578\uff0c\u4e26\u7cfb\u7d71\u6027\u5730\u7814\u7a76\u4e86\u300c\u6e67\u73fe\u80fd\u529b\u300d <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>Llama 2<\/strong><\/td><td>2023<\/td><td>Meta AI<\/td><td>Decoder-Only<\/td><td>\u767c\u5e03\u4e86\u5f37\u5927\u7684\u958b\u6e90\u6a21\u578b\u7cfb\u5217\uff0c\u63a8\u52d5\u4e86\u793e\u5340\u7684\u767c\u5c55\u548c\u7814\u7a76 <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>Mixtral 8x7B<\/strong><\/td><td>2023<\/td><td>Mistral AI<\/td><td>Decoder-Only (MoE)<\/td><td>\u63a1\u7528\u7a00\u758f\u5c08\u5bb6\u6df7\u5408\uff08MoE\uff09\u67b6\u69cb\uff0c\u5728\u4fdd\u6301\u9ad8\u6027\u80fd\u7684\u540c\u6642\u5927\u5e45\u63d0\u5347\u63a8\u7406\u6548\u7387 <sup><\/sup> &nbsp;<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u7b2c\u4e00\u7ae0\uff1a\u7406\u89e3\u7684\u6642\u4ee3 &#8211; \u7de8\u78bc\u5668-Only\u738b\u671d<\/h2>\n\n\n\n<p>\u5728Transformer\u5f15\u767c\u7684\u5049\u5927\u5206\u6b67\u4e4b\u5f8c\uff0c\u7b2c\u4e00\u500b\u8fc5\u901f\u5d1b\u8d77\u4e26\u78ba\u7acb\u7d71\u6cbb\u5730\u4f4d\u7684\u738b\u671d\uff0c\u662f\u5c08\u6ce8\u65bc\u6df1\u5ea6\u8a9e\u8a00\u7406\u89e3\u7684\u7de8\u78bc\u5668-Only\u6a21\u578b\u3002\u9019\u500b\u738b\u671d\u7684\u958b\u570b\u541b\u4e3b\u662fBERT\uff0c\u5b83\u7684\u51fa\u73fe\u4e0d\u50c5\u50c5\u662f\u6280\u8853\u4e0a\u7684\u52dd\u5229\uff0c\u66f4\u662f\u4e00\u7a2e\u601d\u60f3\u4e0a\u7684\u9769\u65b0\uff0c\u5b83\u91cd\u65b0\u5b9a\u7fa9\u4e86\u6a5f\u5668\u5982\u4f55\u300c\u95b1\u8b80\u300d\u548c\u300c\u7406\u89e3\u300d\u6587\u672c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">BERT\u7684\u7d71\u6cbb\uff1a\u4e0a\u4e0b\u6587\u4e4b\u738b\u7684\u65b0\u751f<\/h3>\n\n\n\n<p>2018\u5e74\uff0cGoogle\u7684\u7814\u7a76\u4eba\u54e1\u63a8\u51fa\u4e86BERT\uff08Bidirectional Encoder Representations from Transformers\uff09\uff0c\u9019\u500b\u6a21\u578b\u8fc5\u901f\u6210\u70baNLP\u9818\u57df\u7121\u8655\u4e0d\u5728\u7684\u57fa\u790e\u8a2d\u65bd <sup><\/sup>\u3002BERT\u7684\u67b6\u69cb\u5728\u6982\u5ff5\u4e0a\u975e\u5e38\u7c21\u6f54\uff1a\u5b83\u76f4\u63a5\u63a1\u7528\u4e86\u539f\u59cbTransformer\u7684\u7de8\u78bc\u5668\u5806\u758a\u90e8\u5206 <sup><\/sup>\u3002\u7136\u800c\uff0c\u5176\u771f\u6b63\u7684\u9769\u547d\u6027\u5728\u65bc\u5b83\u5982\u4f55\u88ab\u8a13\u7df4\u4ee5\u53ca\u5b83\u56e0\u6b64\u7372\u5f97\u7684\u7368\u7279\u80fd\u529b\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u96d9\u5411\u6027\u7684\u529b\u91cf<\/h4>\n\n\n\n<p>BERT\u6700\u6838\u5fc3\u7684\u8ca2\u737b\u5728\u65bc\u5176<strong>\u96d9\u5411\u6027<\/strong>\u3002\u5728BERT\u4e4b\u524d\uff0c\u8a9e\u8a00\u6a21\u578b\uff08\u5982GPT-1\uff09\u901a\u5e38\u662f\u55ae\u5411\u7684\uff0c\u5373\u5f9e\u5de6\u5230\u53f3\u8655\u7406\u6587\u672c\uff0c\u9810\u6e2c\u4e0b\u4e00\u500b\u8a5e\u3002\u9019\u7a2e\u65b9\u6cd5\u5c0d\u65bc\u751f\u6210\u4efb\u52d9\u662f\u81ea\u7136\u7684\uff0c\u4f46\u5c0d\u65bc\u9700\u8981\u6df1\u523b\u7406\u89e3\u6574\u500b\u53e5\u5b50\u8a9e\u5883\u7684\u4efb\u52d9\u5247\u5b58\u5728\u5c40\u9650\u3002\u4f8b\u5982\uff0c\u5728\u53e5\u5b50\u300cThe man went to the bank to withdraw money\u300d\u4e2d\uff0c\u8981\u7406\u89e3\u300cbank\u300d\u7684\u542b\u7fa9\uff08\u662f\u9280\u884c\u800c\u975e\u6cb3\u5cb8\uff09\uff0c\u6a21\u578b\u9700\u8981\u540c\u6642\u8003\u616e\u5176\u5de6\u908a\u7684\u4e0a\u4e0b\u6587\uff08\u300cwent to the\u300d\uff09\u548c\u53f3\u908a\u7684\u4e0a\u4e0b\u6587\uff08\u300cto withdraw money\u300d\uff09\u3002\u55ae\u5411\u6a21\u578b\u5728\u8655\u7406\u300cbank\u300d\u9019\u500b\u8a5e\u6642\uff0c\u7121\u6cd5\u770b\u5230\u5f8c\u9762\u7684\u4fe1\u606f\u3002<\/p>\n\n\n\n<p>BERT\u901a\u904e\u5176\u9810\u8a13\u7df4\u76ee\u6a19\uff0c\u5b8c\u7f8e\u5730\u89e3\u6c7a\u4e86\u9019\u500b\u554f\u984c\u3002\u5b83\u80fd\u5920\u5728\u6240\u6709\u5c64\u4e2d\u540c\u6642\u5229\u7528\u5de6\u5074\u548c\u53f3\u5074\u7684\u4e0a\u4e0b\u6587\u4f86\u9810\u8a13\u7df4\u6df1\u5ea6\u7684\u96d9\u5411\u8868\u793a <sup><\/sup>\u3002\u9019\u610f\u5473\u8457\u6a21\u578b\u5728\u70ba\u6bcf\u500b\u8a5e\u751f\u6210\u8868\u793a\u6642\uff0c\u90fd\u5df2\u7d93\u300c\u770b\u904e\u300d\u4e86\u6574\u500b\u53e5\u5b50\uff0c\u5f9e\u800c\u80fd\u5920\u6355\u6349\u5230\u6975\u5176\u8c50\u5bcc\u548c\u7cbe\u78ba\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u5275\u65b0\u7684\u9810\u8a13\u7df4\u76ee\u6a19<\/h4>\n\n\n\n<p>\u70ba\u4e86\u5be6\u73fe\u9019\u7a2e\u96d9\u5411\u5b78\u7fd2\uff0cBERT\u7684\u8a2d\u8a08\u8005\u5011\u5fc5\u9808\u653e\u68c4\u50b3\u7d71\u7684\u8a9e\u8a00\u5efa\u6a21\u76ee\u6a19\uff0c\u56e0\u70ba\u5728\u96d9\u5411\u7684\u8a2d\u5b9a\u4e0b\uff0c\u300c\u9810\u6e2c\u4e0b\u4e00\u500b\u8a5e\u300d\u8b8a\u5f97\u5fae\u4e0d\u8db3\u9053\u3002\u53d6\u800c\u4ee3\u4e4b\uff0c\u4ed6\u5011\u8a2d\u8a08\u4e86\u5169\u500b\u5de7\u5999\u7684\u81ea\u76e3\u7763\u9810\u8a13\u7df4\u4efb\u52d9\uff1a<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li><strong>\u63a9\u78bc\u8a9e\u8a00\u6a21\u578b\uff08Masked Language Model, MLM\uff09<\/strong>\uff1a\u9019\u662fBERT\u7684\u9748\u9b42\u6240\u5728\u3002\u5728\u8a13\u7df4\u904e\u7a0b\u4e2d\uff0c\u6a21\u578b\u6703\u96a8\u6a5f\u5730\u5c07\u8f38\u5165\u5e8f\u5217\u4e2d15%\u7684\u8a5e\u5143\uff08token\uff09\u66ff\u63db\u70ba\u4e00\u500b\u7279\u6b8a\u7684&#8220;\u6a19\u8a18 \u3002\u6a21\u578b\u7684\u4efb\u52d9\u5c31\u662f\u6839\u64da\u5468\u570d\u672a\u88ab\u63a9\u84cb\u7684\u4e0a\u4e0b\u6587\uff0c\u4f86\u9810\u6e2c\u9019\u4e9b\u88ab\u63a9\u84cb\u7684\u539f\u59cb\u8a5e\u5143\u3002\u9019\u500b\u904e\u7a0b\u5c31\u50cf\u662f\u8b93\u6a21\u578b\u505a\u300c\u5b8c\u5f62\u586b\u7a7a\u300d\u3002\u56e0\u70ba\u9810\u6e2c\u76ee\u6a19\u662f\u53e5\u5b50\u4e2d\u9593\u7684\u8a5e\uff0c\u6a21\u578b\u88ab\u8feb\u5b78\u7fd2\u878d\u5408\u4f86\u81ea\u5de6\u53f3\u5169\u908a\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u5f9e\u800c\u5b78\u6703\u4e86\u771f\u6b63\u7684\u96d9\u5411\u8a9e\u5883\u8868\u793a\u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u4e0b\u4e00\u53e5\u9810\u6e2c\uff08Next Sentence Prediction, NSP\uff09<\/strong>\uff1a\u70ba\u4e86\u8b93\u6a21\u578b\u7406\u89e3\u53e5\u5b50\u4e4b\u9593\u7684\u95dc\u4fc2\uff08\u5982\u860a\u542b\u3001\u56e0\u679c\uff09\uff0cBERT\u9084\u5f15\u5165\u4e86NSP\u4efb\u52d9\u3002\u5728\u8a13\u7df4\u6642\uff0c\u6a21\u578b\u6703\u63a5\u6536\u4e00\u5c0d\u53e5\u5b50A\u548cB\uff0c\u4e26\u9700\u8981\u5224\u65b7\u53e5\u5b50B\u662f\u5426\u662f\u53e5\u5b50A\u5728\u539f\u6587\u4e2d\u7684\u4e0b\u4e00\u53e5 \u3002\u9019\u500b\u4e8c\u5143\u5206\u985e\u4efb\u52d9\u65e8\u5728\u6355\u6349\u9577\u8ddd\u96e2\u7684\u53e5\u5b50\u9023\u8cab\u6027\uff0c\u9019\u5c0d\u65bc\u554f\u7b54\uff08QA\uff09\u548c\u81ea\u7136\u8a9e\u8a00\u63a8\u65b7\uff08NLI\uff09\u7b49\u4efb\u52d9\u81f3\u95dc\u91cd\u8981\u3002 \u00a0<\/li>\n<\/ol>\n\n\n\n<p>\u9019\u7a2e\u300c\u9810\u8a13\u7df4-\u5fae\u8abf\u300d\uff08pre-train then fine-tune\uff09\u7684\u7bc4\u5f0f\u6210\u70ba\u4e86BERT\u738b\u671d\u7684\u6a19\u8a8c\u3002\u4e00\u500b\u5728\u6d77\u91cf\u7121\u6a19\u7c64\u6587\u672c\u4e0a\u901a\u904eMLM\u548cNSP\u9810\u8a13\u7df4\u597d\u7684BERT\u6a21\u578b\uff0c\u53ef\u4ee5\u901a\u904e\u5728\u5176\u9802\u90e8\u6dfb\u52a0\u4e00\u500b\u7c21\u55ae\u7684\u8f38\u51fa\u5c64\uff0c\u7136\u5f8c\u5728\u7279\u5b9a\u4efb\u52d9\u7684\u5c0f\u898f\u6a21\u6709\u6a19\u7c64\u6578\u64da\u4e0a\u9032\u884c\u5fae\u8abf\uff0c\u5c31\u80fd\u5728\u5404\u7a2e\u4e0b\u6e38\u4efb\u52d9\u4e0a\u53d6\u5f97\u7576\u6642\u6700\u5148\u9032\u7684\u6210\u679c\uff0c\u800c\u7121\u9700\u5c0d\u6a21\u578b\u67b6\u69cb\u9032\u884c\u5927\u7684\u4fee\u6539 <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7e7c\u627f\u8005\u8207\u512a\u5316\u8005\uff1aBERT\u5b78\u7684\u6f14\u9032<\/h3>\n\n\n\n<p>BERT\u7684\u6210\u529f\u5f15\u767c\u4e86\u4e00\u80a1\u7814\u7a76\u71b1\u6f6e\uff0c\u7814\u7a76\u8005\u5011\u958b\u59cb\u63a2\u7d22\u5982\u4f55\u6539\u9032\u548c\u512a\u5316\u5176\u914d\u65b9\u3002\u5f88\u5feb\uff0c\u4e00\u7cfb\u5217\u91cd\u8981\u7684\u7e7c\u627f\u8005\u51fa\u73fe\u4e86\uff0c\u5b83\u5011\u96d6\u7136\u6c92\u6709\u6539\u8b8aBERT\u7684\u6838\u5fc3\u67b6\u69cb\uff0c\u4f46\u901a\u904e\u5c0d\u8a13\u7df4\u65b9\u6cd5\u7684\u7cbe\u7149\uff0c\u9032\u4e00\u6b65\u978f\u56fa\u4e86\u7de8\u78bc\u5668-Only\u738b\u671d\u7684\u7d71\u6cbb\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">RoBERTa\uff1a\u4e00\u500b\u88ab\u5145\u5206\u512a\u5316\u7684BERT<\/h4>\n\n\n\n<p>2019\u5e74\uff0c\u4f86\u81eaFacebook AI\u7684\u7814\u7a76\u4eba\u54e1\u767c\u5e03\u4e86RoBERTa\uff08A Robustly Optimized BERT Pretraining Approach\uff09\u3002RoBERTa\u4e26\u975e\u4e00\u500b\u5168\u65b0\u7684\u67b6\u69cb\uff0c\u800c\u662f\u4e00\u9805\u56b4\u8b39\u7684\u8907\u88fd\u7814\u7a76\uff08replication study\uff09\uff0c\u5176\u7d50\u8ad6\u4ee4\u4eba\u9707\u9a5a\uff1aBERT\u672c\u8eab\u88ab\u56b4\u91cd\u5730<strong>\u8a13\u7df4\u4e0d\u8db3<\/strong>\uff08undertrained\uff09<sup><\/sup>\u3002\u901a\u904e\u4e00\u7cfb\u5217\u7cbe\u5fc3\u8a2d\u8a08\u7684\u8a13\u7df4\u7b56\u7565\u8abf\u6574\uff0cRoBERTa\u5728\u6027\u80fd\u4e0a\u5927\u5e45\u8d85\u8d8a\u4e86BERT\u3002\u5176\u95dc\u9375\u6539\u9032\u5305\u62ec\uff1a &nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u66f4\u5927\u898f\u6a21\u7684\u8a13\u7df4<\/strong>\uff1aRoBERTa\u8a13\u7df4\u7684\u6642\u9593\u66f4\u9577\uff0c\u4f7f\u7528\u4e86\u66f4\u5927\u7684\u6279\u91cf\u5927\u5c0f\uff08batch size\uff09\uff0c\u4e26\u5728\u66f4\u591a\u7684\u6578\u64da\u4e0a\u9032\u884c\u8a13\u7df4 \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u52d5\u614b\u63a9\u78bc\uff08Dynamic Masking\uff09<\/strong>\uff1a\u539f\u59cbBERT\u5728\u6578\u64da\u9810\u8655\u7406\u968e\u6bb5\u53ea\u9032\u884c\u4e00\u6b21\u975c\u614b\u63a9\u78bc\u3002\u800cRoBERTa\u5247\u5728\u6bcf\u6b21\u5411\u6a21\u578b\u8f38\u5165\u5e8f\u5217\u6642\u52d5\u614b\u751f\u6210\u65b0\u7684\u63a9\u78bc\u6a21\u5f0f\uff0c\u589e\u52a0\u4e86\u6578\u64da\u7684\u591a\u6a23\u6027\uff0c\u4f7f\u6a21\u578b\u5b78\u7fd2\u66f4\u70ba\u9b6f\u68d2 \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u62cb\u68c4NSP\u4efb\u52d9<\/strong>\uff1aRoBERTa\u7684\u7814\u7a76\u767c\u73fe\uff0c\u539f\u59cb\u7684NSP\u4efb\u52d9\u4e0d\u50c5\u6548\u679c\u4e0d\u4f73\uff0c\u751a\u81f3\u53ef\u80fd\u5c0d\u67d0\u4e9b\u4e0b\u6e38\u4efb\u52d9\u7684\u6027\u80fd\u7522\u751f\u8ca0\u9762\u5f71\u97ff\u3002\u56e0\u6b64\uff0c\u4ed6\u5011\u679c\u65b7\u5730\u79fb\u9664\u4e86NSP\u76ee\u6a19\uff0c\u50c5\u4f7f\u7528MLM\u9032\u884c\u9810\u8a13\u7df4\uff0c\u7d50\u679c\u767c\u73fe\u6027\u80fd\u53cd\u800c\u5f97\u5230\u4e86\u63d0\u5347 \u3002 \u00a0<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">ALBERT\uff1a\u8f15\u91cf\u7d1a\u7684BERT<\/h4>\n\n\n\n<p>\u5118\u7ba1BERT\u53ca\u5176\u8b8a\u9ad4\u6027\u80fd\u5f37\u5927\uff0c\u4f46\u5176\u5de8\u5927\u7684\u6a21\u578b\u5c3a\u5bf8\uff08BERT-Large\u64c1\u67093.4\u5104\u53c3\u6578\uff09\u5e36\u4f86\u4e86\u9ad8\u6602\u7684\u8a08\u7b97\u548c\u5167\u5b58\u6210\u672c\u3002\u70ba\u4e86\u89e3\u6c7a\u9019\u500b\u554f\u984c\uff0cGoogle\u57282019\u5e74\u63a8\u51fa\u4e86ALBERT\uff08A Lite BERT\uff09\uff0c\u65e8\u5728\u5927\u5e45\u63d0\u5347\u6a21\u578b\u7684<strong>\u53c3\u6578\u6548\u7387<\/strong> <sup><\/sup>\u3002ALBERT\u901a\u904e\u5169\u7a2e\u5de7\u5999\u7684\u6280\u8853\u5be6\u73fe\u4e86\u9019\u4e00\u9ede\uff1a &nbsp;<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li><strong>\u56e0\u5f0f\u5206\u89e3\u7684\u5d4c\u5165\u53c3\u6578\u5316\uff08Factorized Embedding Parameterization\uff09<\/strong>\uff1a\u5728BERT\u4e2d\uff0c\u8a5e\u5d4c\u5165\u5c64\u7684\u7dad\u5ea6E\u8207\u96b1\u85cf\u5c64\u7684\u7dad\u5ea6H\u662f\u7d81\u5b9a\u7684\uff08E=H\uff09\u3002\u8003\u616e\u5230\u8a5e\u5d4c\u5165\u5b78\u7fd2\u7684\u662f\u8207\u4e0a\u4e0b\u6587\u7121\u95dc\u7684\u8868\u793a\uff0c\u800c\u96b1\u85cf\u5c64\u5b78\u7fd2\u7684\u662f\u8207\u4e0a\u4e0b\u6587\u76f8\u95dc\u7684\u8868\u793a\uff0c\u9019\u7a2e\u7d81\u5b9a\u662f\u4e0d\u5fc5\u8981\u7684\uff0c\u4e26\u4e14\u6703\u56e0\u5de8\u5927\u7684\u8a5e\u5f59\u8868V\u800c\u5c0e\u81f4\u5d4c\u5165\u77e9\u9663\uff08V\u00d7H\uff09\u53c3\u6578\u904e\u591a\u3002ALBERT\u5c07\u9019\u500b\u5927\u77e9\u9663\u5206\u89e3\u70ba\u5169\u500b\u8f03\u5c0f\u7684\u77e9\u9663\uff0c\u5c07\u53c3\u6578\u5f9e$O(V \\times H)$\u6e1b\u5c11\u5230$O(V \\times E + E \\times H)$\uff0c\u5176\u4e2dE\u226aH\u3002\u9019\u6975\u5927\u5730\u6e1b\u5c11\u4e86\u53c3\u6578\u6578\u91cf \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u8de8\u5c64\u53c3\u6578\u5171\u4eab\uff08Cross-Layer Parameter Sharing\uff09<\/strong>\uff1aALBERT\u5728\u6240\u6709\u7de8\u78bc\u5668\u5c64\u4e4b\u9593\u5171\u4eab\u53c3\u6578\u3002\u9019\u610f\u5473\u8457\u6a21\u578b\u53ea\u9700\u8981\u5b78\u7fd2\u4e00\u7d44\u7de8\u78bc\u5668\u5c64\u7684\u53c3\u6578\uff0c\u7136\u5f8c\u5728\u7db2\u7d61\u7684\u6240\u6709\u6df1\u5ea6\u91cd\u8907\u4f7f\u7528\u3002\u9019\u7a2e\u65b9\u6cd5\u4e0d\u50c5\u986f\u8457\u6e1b\u5c11\u4e86\u7e3d\u53c3\u6578\uff0c\u9084\u8d77\u5230\u4e86\u4e00\u7a2e\u6b63\u5247\u5316\u7684\u4f5c\u7528\uff0c\u4f7f\u6a21\u578b\u8a13\u7df4\u66f4\u7a69\u5b9a \u3002\u4e00\u500b\u8207BERT-Large\u914d\u7f6e\u76f8\u4f3c\u7684ALBERT\u6a21\u578b\uff0c\u53c3\u6578\u5c11\u4e8618\u500d\uff0c\u8a13\u7df4\u901f\u5ea6\u537b\u5feb\u4e861.7\u500d\u3002 \u00a0<\/li>\n<\/ol>\n\n\n\n<p>\u6b64\u5916\uff0cALBERT\u4e5f\u7528\u4e00\u500b\u66f4\u6709\u6548\u7684**\u53e5\u5b50\u9806\u5e8f\u9810\u6e2c\uff08Sentence-Order Prediction, SOP\uff09**\u4efb\u52d9\u53d6\u4ee3\u4e86NSP\u3002SOP\u4efb\u52d9\u8981\u6c42\u6a21\u578b\u5340\u5206\u5169\u500b\u9023\u7e8c\u53e5\u5b50\u7684\u539f\u59cb\u9806\u5e8f\u548c\u4ea4\u63db\u5f8c\u7684\u9806\u5e8f\uff0c\u9019\u6bd4NSP\u66f4\u80fd\u96c6\u4e2d\u65bc\u5b78\u7fd2\u53e5\u5b50\u9593\u7684\u9023\u8cab\u6027\uff0c\u800c\u4e0d\u662f\u4e3b\u984c\u9810\u6e2c <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7d71\u6cbb\u7684\u8b49\u64da\uff1a\u5f81\u670dNLU\u57fa\u6e96\u6e2c\u8a66<\/h3>\n\n\n\n<p>\u7de8\u78bc\u5668-Only\u738b\u671d\u7684\u7d71\u6cbb\u5730\u4f4d\u4e26\u975e\u7a7a\u8ac7\uff0c\u800c\u662f\u5efa\u7acb\u5728\u5c0d\u7576\u6642\u5b78\u8853\u754c\u516c\u8a8d\u7684\u81ea\u7136\u8a9e\u8a00\u7406\u89e3\uff08NLU\uff09\u57fa\u6e96\u6e2c\u8a66\u7684\u5168\u9762\u5f81\u670d\u4e4b\u4e0a\u3002\u9019\u4e9b\u57fa\u6e96\u6e2c\u8a66\uff0c\u5982\u540c\u53e4\u4ee3\u7684\u7af6\u6280\u5834\uff0c\u662f\u6240\u6709\u6a21\u578b\u8b49\u660e\u81ea\u8eab\u5be6\u529b\u7684\u821e\u53f0\u3002<\/p>\n\n\n\n<p>\u5176\u4e2d\u6700\u8457\u540d\u7684\u5c31\u662f<strong>GLUE<\/strong>\uff08General Language Understanding Evaluation\uff09\u57fa\u6e96\u6e2c\u8a66 <sup><\/sup>\u3002GLUE\u5305\u542b\u4e86\u4e00\u7cfb\u5217\u591a\u6a23\u5316\u7684NLU\u4efb\u52d9\uff0c\u5982\u81ea\u7136\u8a9e\u8a00\u63a8\u65b7\uff08MNLI, RTE\uff09\u3001\u8a9e\u7fa9\u76f8\u4f3c\u5ea6\uff08MRPC, QQP\uff09\u3001\u60c5\u611f\u5206\u6790\uff08SST-2\uff09\u548c\u8a9e\u6cd5\u53ef\u63a5\u53d7\u6027\uff08CoLA\uff09\u7b49\u3002BERT\u4e00\u7d93\u554f\u4e16\uff0c\u5c31\u572811\u9805NLP\u4efb\u52d9\u4e0a\u53d6\u5f97\u4e86\u7576\u6642\u7684\u6700\u5148\u9032\uff08SOTA\uff09\u6210\u679c\uff0c\u5305\u62ec\u5c07GLUE\u7684\u5e73\u5747\u5206\u63a8\u9ad8\u81f380.5%\uff0c\u5be6\u73fe\u4e867.7\u500b\u767e\u5206\u9ede\u7684\u7d55\u5c0d\u63d0\u5347\uff0c\u9019\u5728\u7576\u6642\u662f\u524d\u6240\u672a\u6709\u7684\u6210\u5c31 <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<p>\u96a8\u8457BERT\u7684\u6210\u529f\uff0cGLUE\u7684\u6311\u6230\u6027\u9010\u6f38\u964d\u4f4e\uff0c\u6a21\u578b\u6027\u80fd\u5f88\u5feb\u8d85\u8d8a\u4e86\u4eba\u985e\u57fa\u6e96 <sup><\/sup>\u3002\u70ba\u6b64\uff0c\u7814\u7a76\u754c\u63a8\u51fa\u4e86\u66f4\u5177\u6311\u6230\u6027\u7684 &nbsp;<\/p>\n\n\n\n<p><strong>SuperGLUE<\/strong>\u57fa\u6e96\u6e2c\u8a66 <sup><\/sup>\u3002SuperGLUE\u4fdd\u7559\u4e86GLUE\u4e2d\u6700\u96e3\u7684\u4efb\u52d9\uff0c\u4e26\u52a0\u5165\u4e86\u66f4\u8907\u96dc\u7684\u4efb\u52d9\u683c\u5f0f\uff0c\u5982\u554f\u7b54\u548c\u5171\u6307\u6d88\u89e3\u3002\u5373\u4fbf\u5982\u6b64\uff0c\u7de8\u78bc\u5668-Only\u738b\u671d\u7684\u7e7c\u627f\u8005\u5011\uff0c\u5982RoBERTa\u548c\u5f8c\u4f86\u7684\u8b8a\u9ad4\uff0c\u4f9d\u7136\u5728\u9019\u500b\u66f4\u56f0\u96e3\u7684\u7af6\u6280\u5834\u4e0a\u7e7c\u7e8c\u5237\u65b0\u7d00\u9304 <sup><\/sup>\u3002\u4e0b\u8868\u5c55\u793a\u4e86BERT\u548cRoBERTa\u5728GLUE\u958b\u767c\u96c6\u4e0a\u7684\u4e00\u4e9b\u6027\u80fd\u6578\u64da\uff0c\u76f4\u89c0\u5730\u8b49\u660e\u4e86\u5b83\u5011\u5728NLU\u4efb\u52d9\u4e0a\u7684\u7d71\u6cbb\u529b\u3002 &nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><td>\u4efb\u52d9 (\u6307\u6a19)<\/td><td>BERT-Base<\/td><td>RoBERTa-Base<\/td><\/tr><\/thead><tbody><tr><td>CoLA (Matthew&#8217;s Corr)<\/td><td>59.39<\/td><td>64.83<\/td><\/tr><tr><td>MNLI-m (Accuracy)<\/td><td>83.92<\/td><td>87.31<\/td><\/tr><tr><td>MRPC (F1)<\/td><td>91.28<\/td><td>91.68<\/td><\/tr><tr><td>QNLI (Accuracy)<\/td><td>88.52<\/td><td>90.82<\/td><\/tr><tr><td>QQP (F1)<\/td><td>87.17<\/td><td>88.77<\/td><\/tr><tr><td>RTE (Accuracy)<\/td><td>69.31<\/td><td>73.65<\/td><\/tr><tr><td>SST-2 (Accuracy)<\/td><td>92.55<\/td><td>94.50<\/td><\/tr><tr><td>STS-B (Spearman)<\/td><td>87.88<\/td><td>89.09<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u532f\u51fa\u5230\u8a66\u7b97\u8868<\/p>\n\n\n\n<p>\u6578\u64da\u4f86\u6e90\uff1a<sup><\/sup> &nbsp;<\/p>\n\n\n\n<p>\u9019\u4e00\u7cfb\u5217\u58d3\u5012\u6027\u7684\u5be6\u8b49\u7d50\u679c\uff0c\u7262\u56fa\u5730\u78ba\u7acb\u4e86\u7de8\u78bc\u5668-Only\u67b6\u69cb\u5728\u8a9e\u8a00\u7406\u89e3\u4efb\u52d9\u4e0a\u7684\u9738\u4e3b\u5730\u4f4d\u3002\u9019\u4e5f\u6df1\u523b\u5730\u63ed\u793a\u4e86\u4e00\u500b\u6838\u5fc3\u539f\u7406\uff1a\u67b6\u69cb\u7684\u8a2d\u8a08\u8207\u5176\u8a13\u7df4\u76ee\u6a19\u4e4b\u9593\u5b58\u5728\u8457\u4e00\u7a2e\u5171\u751f\u6f14\u5316\u7684\u95dc\u4fc2\u3002\u4e00\u500b\u96d9\u5411\u7684\u7de8\u78bc\u5668\u67b6\u69cb\uff0c\u7531\u65bc\u5176\u80fd\u5920\u300c\u770b\u5230\u300d\u6574\u500b\u8f38\u5165\u5e8f\u5217\uff0c\u56e0\u6b64\u7121\u6cd5\u4f7f\u7528\u50b3\u7d71\u7684\u5f9e\u5de6\u5230\u53f3\u7684\u751f\u6210\u5f0f\u76ee\u6a19\u9032\u884c\u8a13\u7df4 <sup><\/sup>\u3002\u9019\u7a2e\u67b6\u69cb\u7279\u6027\u53cd\u904e\u4f86 &nbsp;<\/p>\n\n\n\n<p><strong>\u50ac\u751f<\/strong>\u4e86\u50cfMLM\u9019\u6a23\u7684\u81ea\u76e3\u7763\u4efb\u52d9\u7684\u767c\u660e\u3002\u800cMLM\u7684\u6210\u529f\uff0c\u53c8\u6070\u6070\u662f\u8ce6\u4e88\u6a21\u578b\u6df1\u5ea6\u4e0a\u4e0b\u6587\u7406\u89e3\u80fd\u529b\u7684\u95dc\u9375\u3002\u5f8c\u7e8cRoBERTa\u5c0dNSP\u7684\u63da\u68c4\u4ee5\u53caALBERT\u7528SOP\u5c0d\u5176\u7684\u66ff\u4ee3\uff0c\u90fd\u5c55\u793a\u4e86\u9019\u4e00\u5171\u751f\u6f14\u5316\u904e\u7a0b\u7684\u7cbe\u7149\u2014\u2014\u7814\u7a76\u8005\u5011\u4e0d\u65b7\u8abf\u6574\u8a13\u7df4\u76ee\u6a19\uff0c\u4ee5\u66f4\u597d\u5730\u767c\u63ee\u67b6\u69cb\u5167\u5728\u7684\u3001\u70ba\u300c\u7406\u89e3\u300d\u800c\u751f\u7684\u6f5b\u529b\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u7b2c\u4e8c\u7ae0\uff1a\u751f\u6210\u7684\u6642\u4ee3 &#8211; \u89e3\u78bc\u5668-Only\u738b\u671d<\/h2>\n\n\n\n<p>\u5728\u7de8\u78bc\u5668-Only\u6a21\u578b\u5c08\u6ce8\u65bc\u6df1\u5ea6\u7406\u89e3\u7684\u540c\u6642\uff0c\u53e6\u4e00\u689d\u5e73\u884c\u7684\u6f14\u5316\u8def\u5f91\u6b63\u5728\u6084\u7136\u7a4d\u84c4\u529b\u91cf\uff0c\u4e26\u6700\u7d42\u958b\u5275\u4e86\u4e00\u500b\u4ee5<strong>\u751f\u6210<\/strong>\u70ba\u6838\u5fc3\u7684\u8f1d\u714c\u6642\u4ee3\u3002\u9019\u500b\u738b\u671d\u7684\u7de0\u9020\u8005\u662fOpenAI\uff0c\u5176\u65d7\u4e0b\u7684GPT\uff08Generative Pre-trained Transformer\uff09\u7cfb\u5217\u6a21\u578b\uff0c\u901a\u904e\u5c0dTransformer\u89e3\u78bc\u5668\u90e8\u5206\u7684\u6975\u81f4\u6316\u6398\u548c\u5c0d\u6a21\u578b\u898f\u6a21\u7684\u7121\u754f\u63a2\u7d22\uff0c\u5fb9\u5e95\u6539\u8b8a\u4e86\u4eba\u6a5f\u4ea4\u4e92\u7684\u7bc4\u5f0f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">GPT\u7684\u5d1b\u8d77\uff1a\u81ea\u56de\u6b78\u9810\u6e2c\u7684\u529b\u91cf<\/h3>\n\n\n\n<p>\u89e3\u78bc\u5668-Only\u738b\u671d\u7684\u57fa\u77f3\u662fGPT\u6a21\u578b\u3002\u5176\u67b6\u69cb\u672c\u8cea\u4e0a\u662f\u539f\u59cbTransformer\u7684\u89e3\u78bc\u5668\u5806\u758a\uff0c\u4f46\u79fb\u9664\u4e86\u7528\u65bc\u8207\u7de8\u78bc\u5668\u4ea4\u4e92\u7684\u4ea4\u53c9\u6ce8\u610f\u529b\uff08cross-attention\uff09\u5b50\u5c64 <sup><\/sup>\u3002\u9019\u4f7f\u5f97\u5b83\u6210\u70ba\u4e00\u500b\u7d14\u7cb9\u7684\u3001\u81ea\u7d66\u81ea\u8db3\u7684\u751f\u6210\u6a21\u578b\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u56e0\u679c\u8a9e\u8a00\u5efa\u6a21\uff08Causal Language Modeling, CLM\uff09<\/h4>\n\n\n\n<p>\u8207BERT\u7684MLM\u76ee\u6a19\u4e0d\u540c\uff0cGPT\u7cfb\u5217\u7684\u6838\u5fc3\u8a13\u7df4\u76ee\u6a19\u662f<strong>\u56e0\u679c\u8a9e\u8a00\u5efa\u6a21<\/strong>\uff0c\u4e5f\u5c31\u662f\u50b3\u7d71\u7684\u300c\u9810\u6e2c\u4e0b\u4e00\u500b\u8a5e\u300d\u4efb\u52d9 <sup><\/sup>\u3002\u7d66\u5b9a\u4e00\u500b\u6587\u672c\u5e8f\u5217\uff0c\u6a21\u578b\u7684\u4efb\u52d9\u662f\u9810\u6e2c\u5e8f\u5217\u4e2d\u7684\u4e0b\u4e00\u500b\u8a5e\u5143\u3002\u9019\u500b\u904e\u7a0b\u662f &nbsp;<\/p>\n\n\n\n<p><strong>\u81ea\u56de\u6b78<\/strong>\u7684\uff08auto-regressive\uff09\uff0c\u5373\u6a21\u578b\u5728\u751f\u6210\u7b2ci\u500b\u8a5e\u5143\u6642\uff0c\u6703\u5c07\u524di\u22121\u500b\u8a5e\u5143\u4f5c\u70ba\u8f38\u5165\u3002<\/p>\n\n\n\n<p>\u70ba\u4e86\u56b4\u683c\u57f7\u884c\u9019\u7a2e\u5f9e\u5de6\u5230\u53f3\u7684\u751f\u6210\u6d41\u7a0b\uff0c\u89e3\u78bc\u5668-Only\u67b6\u69cb\u63a1\u7528\u4e86<strong>\u63a9\u78bc\u81ea\u8a3b\u610f\u529b<\/strong>\uff08masked self-attention\uff09\u6a5f\u5236 <sup><\/sup>\u3002\u5728\u6ce8\u610f\u529b\u8a08\u7b97\u904e\u7a0b\u4e2d\uff0c\u4e00\u500b\u4f4d\u65bc\u4f4d\u7f6e &nbsp;<\/p>\n\n\n\n<p>i\u7684\u8a5e\u5143\u53ea\u80fd\u95dc\u6ce8\uff08attend to\uff09\u4f4d\u7f6ej\u2264i\u7684\u8a5e\u5143\uff0c\u800c\u4f86\u81ea\u672a\u4f86\u7684\u8a5e\u5143\uff08j&gt;i\uff09\u5247\u88ab\u63a9\u84cb\u6389\u3002\u9019\u7a2e\u6a5f\u5236\u78ba\u4fdd\u4e86\u6a21\u578b\u5728\u8a13\u7df4\u548c\u751f\u6210\u6642\u4e0d\u6703\u300c\u5077\u770b\u300d\u7b54\u6848\uff0c\u5f9e\u800c\u5b78\u6703\u4e86\u5982\u4f55\u9023\u8cab\u5730\u3001\u4e00\u6b65\u4e00\u6b65\u5730\u69cb\u5efa\u6587\u672c\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u65e9\u671f\u7684\u541b\u4e3b\uff1aGPT-1\u8207GPT-2<\/h4>\n\n\n\n<p>GPT\u738b\u671d\u7684\u958b\u7aef\u53ef\u4ee5\u8ffd\u6eaf\u52302018\u5e74\u7684<strong>GPT-1<\/strong> <sup><\/sup>\u3002\u5b83\u6210\u529f\u5730\u8b49\u660e\u4e86\u5728\u89e3\u78bc\u5668-Only\u67b6\u69cb\u4e0a\u9032\u884c\u5927\u898f\u6a21\u751f\u6210\u5f0f\u9810\u8a13\u7df4\uff0c\u7136\u5f8c\u91dd\u5c0d\u4e0b\u6e38\u4efb\u52d9\u9032\u884c\u5fae\u8abf\u7684\u7bc4\u5f0f\u662f\u53ef\u884c\u7684\u3002 &nbsp;<\/p>\n\n\n\n<p>\u7136\u800c\uff0c\u771f\u6b63\u8b93\u4e16\u754c\u77da\u76ee\u7684\u662f2019\u5e74\u767c\u5e03\u7684<strong>GPT-2<\/strong> <sup><\/sup>\u3002GPT-2\u7684\u53c3\u6578\u898f\u6a21\u9054\u5230\u4e8615\u5104\uff0c\u9060\u8d85\u7576\u6642\u7684\u540c\u985e\u6a21\u578b\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5b83\u5c55\u793a\u4e86\u9a5a\u4eba\u7684**\u96f6\u6a23\u672c\uff08zero-shot\uff09**\u751f\u6210\u80fd\u529b <sup><\/sup>\u3002\u5728\u6c92\u6709\u4efb\u4f55\u7279\u5b9a\u4efb\u52d9\u5fae\u8abf\u7684\u60c5\u6cc1\u4e0b\uff0c\u50c5\u50c5\u901a\u904e\u7d66\u4e88\u4e00\u500b\u63d0\u793a\uff08prompt\uff09\uff0cGPT-2\u5c31\u80fd\u5920\u751f\u6210\u9023\u8cab\u3001\u6709\u908f\u8f2f\u751a\u81f3\u9817\u5177\u6587\u91c7\u7684\u6bb5\u843d\uff0c\u6db5\u84cb\u65b0\u805e\u3001\u6545\u4e8b\u3001\u8a69\u6b4c\u7b49\u591a\u7a2e\u98a8\u683c\u3002GPT-2\u7684\u5f37\u5927\u80fd\u529b\u751a\u81f3\u8b93OpenAI\u6700\u521d\u51fa\u65bc\u5c0d\u5176\u6f5b\u5728\u6feb\u7528\u7684\u64d4\u6182\uff0c\u800c\u9078\u64c7\u4e0d\u5b8c\u5168\u516c\u958b\u6a21\u578b <sup><\/sup>\u3002GPT-2\u7684\u6210\u529f\uff0c\u9996\u6b21\u5411\u4e16\u4eba\u63ed\u793a\u4e86\u300c\u898f\u6a21\u300d\u672c\u8eab\u53ef\u80fd\u5c31\u662f\u901a\u5f80\u66f4\u5f37\u5927\u901a\u7528\u667a\u80fd\u7684\u95dc\u9375\u3002 &nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">GPT-3\u7684\u7bc4\u5f0f\u8f49\u79fb\uff1a\u7576\u898f\u6a21\u6210\u70ba\u4e00\u7a2e\u672c\u8cea<\/h3>\n\n\n\n<p>\u5982\u679c\u8aaaGPT-2\u662f\u9ece\u660e\u524d\u7684\u66d9\u5149\uff0c\u90a3\u9ebc2020\u5e74\u767c\u5e03\u7684<strong>GPT-3<\/strong>\u5247\u6a19\u8a8c\u8457\u4e00\u500b\u5168\u65b0\u6642\u4ee3\u7684\u5230\u4f86 <sup><\/sup>\u3002GPT-3\u4e0d\u50c5\u50c5\u662fGPT-2\u7684\u653e\u5927\u7248\uff0c\u5b83\u5728\u8cea\u4e0a\u6539\u8b8a\u4e86\u4eba\u5011\u8207AI\u4ea4\u4e92\u7684\u65b9\u5f0f\uff0c\u5f15\u767c\u4e86\u4e00\u5834\u6df1\u523b\u7684\u7bc4\u5f0f\u8f49\u79fb\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u524d\u6240\u672a\u6709\u7684\u898f\u6a21<\/h4>\n\n\n\n<p>GPT-3\u64c1\u67091750\u5104\u500b\u53c3\u6578\uff0c\u6bd4\u4e4b\u524d\u4efb\u4f55\u975e\u7a00\u758f\u8a9e\u8a00\u6a21\u578b\u7684\u898f\u6a21\u90fd\u5927\u51fa\u4e00\u500b\u6578\u91cf\u7d1a <sup><\/sup>\u3002\u9019\u7a2e\u5de8\u5927\u7684\u898f\u6a21\u4e26\u975e\u50c5\u50c5\u5e36\u4f86\u4e86\u91cf\u7684\u7a4d\u7d2f\uff0c\u800c\u662f\u5f15\u767c\u4e86\u8cea\u7684\u98db\u8e8d\uff0c\u50ac\u751f\u4e86\u5168\u65b0\u7684\u80fd\u529b\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e0a\u4e0b\u6587\u5b78\u7fd2\uff08In-Context Learning\uff09<\/h4>\n\n\n\n<p>GPT-3\u6700\u91cd\u5927\u7684\u8ca2\u737b\u662f\u8b49\u660e\u4e86<strong>\u4e0a\u4e0b\u6587\u5b78\u7fd2<\/strong>\u7684\u9a5a\u4eba\u6f5b\u529b <sup><\/sup>\u3002\u8207BERT\u6240\u5efa\u7acb\u7684\u300c\u9810\u8a13\u7df4-\u5fae\u8abf\u300d\u7bc4\u5f0f\u622a\u7136\u4e0d\u540c\uff0cGPT-3\u53ef\u4ee5\u5728\u4e0d\u9032\u884c\u4efb\u4f55\u68af\u5ea6\u66f4\u65b0\u6216\u5fae\u8abf\u7684\u60c5\u6cc1\u4e0b\u57f7\u884c\u65b0\u4efb\u52d9\u3002\u5b83\u901a\u904e\u5728\u63d0\u793a\u4e2d\u7d66\u51fa\u4efb\u52d9\u7684\u63cf\u8ff0\u548c\u5e7e\u500b\u793a\u4f8b\uff08demonstrations\uff09\u4f86\u300c\u5b78\u7fd2\u300d\u4efb\u52d9\u3002\u6839\u64da\u63d0\u4f9b\u793a\u4f8b\u7684\u6578\u91cf\uff0c\u9019\u7a2e\u5b78\u7fd2\u65b9\u5f0f\u5206\u70ba\uff1a &nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5c11\u6a23\u672c\u5b78\u7fd2\uff08Few-shot\uff09<\/strong>\uff1a\u5728\u63d0\u793a\u4e2d\u63d0\u4f9b\u5e7e\u500b\u793a\u4f8b\u3002<\/li>\n\n\n\n<li><strong>\u55ae\u6a23\u672c\u5b78\u7fd2\uff08One-shot\uff09<\/strong>\uff1a\u5728\u63d0\u793a\u4e2d\u63d0\u4f9b\u4e00\u500b\u793a\u4f8b\u3002<\/li>\n\n\n\n<li><strong>\u96f6\u6a23\u672c\u5b78\u7fd2\uff08Zero-shot\uff09<\/strong>\uff1a\u5728\u63d0\u793a\u4e2d\u53ea\u63d0\u4f9b\u4efb\u52d9\u63cf\u8ff0\uff0c\u4e0d\u63d0\u4f9b\u4efb\u4f55\u793a\u4f8b\u3002<\/li>\n<\/ul>\n\n\n\n<p>GPT-3\u5728\u8a31\u591aNLP\u6578\u64da\u96c6\u4e0a\uff0c\u50c5\u901a\u904e\u5c11\u6a23\u672c\u5b78\u7fd2\u5c31\u9054\u5230\u4e86\u8207\u7d93\u904e\u5c08\u9580\u5fae\u8abf\u7684SOTA\u6a21\u578b\u76f8\u5ab2\u7f8e\u751a\u81f3\u8d85\u8d8a\u7684\u6027\u80fd <sup><\/sup>\u3002\u9019\u4e00\u767c\u73fe\u5177\u6709\u9769\u547d\u6027\u610f\u7fa9\u3002\u5b83\u610f\u5473\u8457\u8207\u6a21\u578b\u7684\u4ea4\u4e92\u65b9\u5f0f\u5f9e\u9700\u8981\u5c08\u696d\u77e5\u8b58\u7684\u300c\u6a21\u578b\u8a13\u7df4\u300d\u8f49\u8b8a\u70ba\u66f4\u76f4\u89c0\u3001\u66f4\u81ea\u7136\u7684\u300c\u5c0d\u8a71\u5f0f\u6307\u4ee4\u300d\u3002\u9019\u7a2e\u7bc4\u5f0f\u7684\u8f49\u8b8a\u6975\u5927\u5730\u964d\u4f4e\u4e86\u4f7f\u7528\u5f37\u5927AI\u7684\u9580\u6abb\uff0c\u70ba\u5f8c\u4f86ChatGPT\u7b49\u61c9\u7528\u7684\u7206\u767c\u92ea\u5e73\u4e86\u9053\u8def\u3002\u5b83\u5c07\u300c\u7de8\u7a0b\u300d\u6a21\u578b\u7684\u91cd\u5fc3\u5f9e\u8abf\u6574\u6a21\u578b\u6b0a\u91cd\u8f49\u79fb\u5230\u4e86\u8a2d\u8a08\u6709\u6548\u7684\u63d0\u793a\u4e0a\uff0c\u50ac\u751f\u4e86\u300c\u63d0\u793a\u5de5\u7a0b\u300d\uff08prompt engineering\uff09\u9019\u4e00\u5168\u65b0\u9818\u57df\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u7e2e\u653e\u5b9a\u5f8b\u8207\u6e67\u73fe\u80fd\u529b<\/h4>\n\n\n\n<p>GPT-3\u7684\u6210\u529f\u4e5f\u8b93\u7814\u7a76\u754c\u958b\u59cb\u7cfb\u7d71\u6027\u5730\u95dc\u6ce8<strong>\u7e2e\u653e\u5b9a\u5f8b\uff08Scaling Laws\uff09<\/strong>\u3002\u7814\u7a76\u767c\u73fe\uff0c\u96a8\u8457\u6a21\u578b\u898f\u6a21\u3001\u6578\u64da\u96c6\u5927\u5c0f\u548c\u8a08\u7b97\u91cf\u7684\u589e\u52a0\uff0c\u6a21\u578b\u7684\u6027\u80fd\u6703\u4ee5\u4e00\u7a2e\u53ef\u9810\u6e2c\u7684\u65b9\u5f0f\uff08\u901a\u5e38\u662f\u51aa\u5f8b\uff09\u5e73\u6ed1\u63d0\u5347 <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<p>\u7136\u800c\uff0c\u66f4\u6709\u8da3\u7684\u767c\u73fe\u5728\u65bc\uff0c\u67d0\u4e9b\u80fd\u529b\u4e26\u975e\u96a8\u8457\u898f\u6a21\u7684\u589e\u52a0\u800c\u5e73\u6ed1\u51fa\u73fe\u3002\u76f8\u53cd\uff0c\u5b83\u5011\u4f3c\u4e4e\u5728\u6a21\u578b\u9054\u5230\u67d0\u500b\u81e8\u754c\u898f\u6a21\u5f8c\u7a81\u7136<strong>\u6e67\u73fe\uff08emerge\uff09\u51fa\u4f86\u3002\u9019\u4e9b\u6e67\u73fe\u80fd\u529b<\/strong>\uff08emergent abilities\uff09\u5728\u5c0f\u6a21\u578b\u4e0a\u5b8c\u5168\u4e0d\u5b58\u5728\uff08\u8868\u73fe\u70ba\u96a8\u6a5f\u731c\u6e2c\uff09\uff0c\u4f46\u5728\u5927\u6a21\u578b\u4e0a\u537b\u8868\u73fe\u51fa\u8272 <sup><\/sup>\u3002\u4f8b\u5982\uff0c\u591a\u6b65\u7b97\u8853\u3001\u7406\u89e3\u6bd4\u55bb\u7b49\u8907\u96dc\u63a8\u7406\u80fd\u529b\uff0c\u90fd\u662f\u5728\u6a21\u578b\u898f\u6a21\u9054\u5230\u6578\u767e\u5104\u751a\u81f3\u5343\u5104\u53c3\u6578\u5f8c\u624d\u986f\u73fe\u51fa\u4f86\u7684\u3002 &nbsp;<\/p>\n\n\n\n<p>\u95dc\u65bc\u6e67\u73fe\u80fd\u529b\u7684\u672c\u8cea\uff0c\u5b78\u8853\u754c\u5b58\u5728\u4e00\u4e9b\u722d\u8b70\u3002\u4e00\u4e9b\u7814\u7a76\u8a8d\u70ba\uff0c\u9019\u4e9b\u80fd\u529b\u662f\u6a21\u578b\u9054\u5230\u8db3\u5920\u8907\u96dc\u5ea6\u5f8c\u771f\u6b63\u7684\u8cea\u8b8a <sup><\/sup>\u3002\u800c\u53e6\u4e00\u4e9b\u7814\u7a76\u5247\u6307\u51fa\uff0c\u6240\u8b02\u7684\u300c\u6e67\u73fe\u300d\u53ef\u80fd\u53ea\u662f\u7814\u7a76\u8005\u9078\u64c7\u7684\u8a55\u4f30\u6307\u6a19\uff08\u5982\u8981\u6c42\u5b8c\u5168\u5339\u914d\u7684\u6e96\u78ba\u7387\uff09\u6240\u5c0e\u81f4\u7684\u5047\u8c61\uff1b\u5982\u679c\u63db\u7528\u66f4\u9023\u7e8c\u7684\u6307\u6a19\uff08\u5982\u4ea4\u53c9\u71b5\u640d\u5931\uff09\uff0c\u6027\u80fd\u7684\u63d0\u5347\u5176\u5be6\u662f\u5e73\u6ed1\u548c\u53ef\u9810\u6e2c\u7684 <sup><\/sup>\u3002\u5118\u7ba1\u5b58\u5728\u722d\u8b70\uff0c\u4f46\u6e67\u73fe\u80fd\u529b\u7684\u6982\u5ff5\u6975\u5927\u5730\u6fc0\u767c\u4e86\u5c0d\u8d85\u5927\u898f\u6a21\u6a21\u578b\u6f5b\u529b\u7684\u60f3\u8c61\u3002 &nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u958b\u6e90\u529b\u91cf\u7684\u5d1b\u8d77\uff1aLlama\u3001PaLM\u8207\u6b0a\u529b\u7684\u6c11\u4e3b\u5316<\/h3>\n\n\n\n<p>\u5728GPT-3\u78ba\u7acb\u4e86\u5927\u898f\u6a21\u89e3\u78bc\u5668-Only\u6a21\u578b\u7684\u7d71\u6cbb\u5730\u4f4d\u5f8c\uff0c\u9019\u80a1\u529b\u91cf\u958b\u59cb\u5f9e\u5c11\u6578\u9802\u7d1a\u5be6\u9a57\u5ba4\u5411\u66f4\u5ee3\u6cdb\u7684\u793e\u5340\u64f4\u6563\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Google\u7684PaLM<\/strong>\uff1aGoogle\u57282022\u5e74\u767c\u5e03\u4e86PaLM\uff08Pathways Language Model\uff09\uff0c\u5c07\u6a21\u578b\u898f\u6a21\u9032\u4e00\u6b65\u63a8\u5411\u4e865400\u5104\u53c3\u6578\u7684\u9a5a\u4eba\u6c34\u5e73 \u3002PaLM\u5728\u773e\u591a\u57fa\u6e96\u6e2c\u8a66\u4e0a\u53d6\u5f97\u4e86SOTA\u6210\u7e3e\uff0c\u4e26\u70ba\u6e67\u73fe\u80fd\u529b\u7684\u5b58\u5728\u63d0\u4f9b\u4e86\u66f4\u591a\u6709\u529b\u7684\u8b49\u64da\uff0c\u7279\u5225\u662f\u5728\u591a\u6b65\u63a8\u7406\u4efb\u52d9\u4e0a\u5c55\u73fe\u4e86\u7a81\u7834\u6027\u7684\u6027\u80fd \u3002 \u00a0<\/li>\n\n\n\n<li><strong>Meta\u7684Llama\u7cfb\u5217<\/strong>\uff1aMeta AI\u57282023\u5e74\u767c\u5e03\u7684Llama\u7cfb\u5217\uff0c\u7279\u5225\u662f<strong>Llama 2<\/strong>\uff0c\u6210\u70ba\u4e86\u958b\u6e90\u793e\u5340\u7684\u4e00\u500b\u91cc\u7a0b\u7891\u4e8b\u4ef6 \u3002Llama 2\u4e0d\u50c5\u6027\u80fd\u5f37\u5927\uff0c\u53ef\u8207\u9589\u6e90\u6a21\u578b\u76f8\u5ab2\u7f8e\uff0c\u800c\u4e14\u5176\u6b0a\u91cd\u5c0d\u7814\u7a76\u548c\u5546\u696d\u7528\u9014\u958b\u653e\uff08\u6709\u4e00\u5b9a\u9650\u5236\uff09\u3002\u9019\u6975\u5927\u5730\u63a8\u52d5\u4e86LLM\u7814\u7a76\u7684\u6c11\u4e3b\u5316\uff0c\u4f7f\u5f97\u66f4\u5ee3\u6cdb\u7684\u7814\u7a76\u8005\u548c\u958b\u767c\u8005\u80fd\u5920\u5728\u6b64\u57fa\u790e\u4e0a\u9032\u884c\u5275\u65b0\u3002Llama 2\u7684\u67b6\u69cb\u63a1\u7528\u4e86\u7576\u6642\u6700\u5148\u9032\u7684\u7d44\u4ef6\uff0c\u5982RMSNorm\u9810\u6b78\u4e00\u5316\u3001SwiGLU\u6fc0\u6d3b\u51fd\u6578\u548c\u65cb\u8f49\u4f4d\u7f6e\u5d4c\u5165\uff08RoPE\uff09\uff0c\u5176\u5c0d\u8a71\u7248\u672cLlama 2-Chat\u66f4\u662f\u901a\u904e\u76e3\u7763\u5fae\u8abf\uff08SFT\uff09\u548c\u5e36\u6709\u4eba\u985e\u53cd\u994b\u7684\u5f37\u5316\u5b78\u7fd2\uff08RLHF\uff09\u9032\u884c\u4e86\u7cbe\u5fc3\u7684\u5c0d\u9f4a \u3002 \u00a0<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u7d71\u6cbb\u7684\u8b49\u64da\uff1a\u77e5\u8b58\u8207\u751f\u6210\u7684\u96d9\u91cd\u5236\u9738<\/h3>\n\n\n\n<p>\u89e3\u78bc\u5668-Only\u738b\u671d\u7684\u5f37\u5927\u5be6\u529b\u9ad4\u73fe\u5728\u5176\u5c0d\u77e5\u8b58\u5bc6\u96c6\u578b\u4efb\u52d9\u548c\u958b\u653e\u5f0f\u751f\u6210\u4efb\u52d9\u7684\u96d9\u91cd\u638c\u63a7\u4e0a\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>MMLU\u57fa\u6e96\u6e2c\u8a66<\/strong>\uff1a\u5982\u679c\u8aaaGLUE\/SuperGLUE\u662f\u8861\u91cfNLU\u80fd\u529b\u7684\u8a66\u91d1\u77f3\uff0c\u90a3\u9ebc<strong>MMLU<\/strong>\uff08Massive Multitask Language Understanding\uff09\u5c31\u662f\u6aa2\u9a57\u6a21\u578b\u901a\u7528\u77e5\u8b58\u5ee3\u5ea6\u548c\u6df1\u5ea6\u7684\u6b0a\u5a01\u6a19\u6e96 \u3002MMLU\u6db5\u84cb\u4e86\u5f9e\u521d\u7b49\u6578\u5b78\u5230\u7f8e\u570b\u6b77\u53f2\u3001\u5f9e\u8a08\u7b97\u6a5f\u79d1\u5b78\u5230\u5c08\u696d\u91ab\u5b78\u7b4957\u500b\u4e0d\u540c\u5b78\u79d1\u7684\u5c08\u696d\u7d1a\u554f\u984c\u3002\u5927\u898f\u6a21\u89e3\u78bc\u5668-Only\u6a21\u578b\uff0c\u5982GPT-4\u3001PaLM\u548cLlama\uff0c\u5728MMLU\u4e0a\u7684\u5353\u8d8a\u8868\u73fe\uff08\u901a\u5e38\u5728\u5c11\u6a23\u672c\u8a2d\u7f6e\u4e0b\u9032\u884c\u8a55\u4f30\uff09\u8b49\u660e\u4e86\u5b83\u5011\u5f9e\u6d77\u91cf\u8a13\u7df4\u6578\u64da\u4e2d\u5b78\u5230\u4e86\u8c50\u5bcc\u7684\u4e16\u754c\u77e5\u8b58 \u3002\u4f8b\u5982\uff0c\u6839\u64daHELM v0.3.0\u7684\u6578\u64da\uff0cGPT-3 (davinci)\u76845-shot MMLU\u6e96\u78ba\u7387\u70ba42.2%\uff0c\u800c\u66f4\u5f37\u5927\u7684\u6a21\u578b\u5982Llama 2 (70B)\u5247\u9054\u5230\u4e8658.2% \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u751f\u6210\u80fd\u529b\u7684\u5c55\u73fe<\/strong>\uff1a\u96d6\u7136\u751f\u6210\u8cea\u91cf\u96e3\u4ee5\u7528\u55ae\u4e00\u6307\u6a19\u91cf\u5316\uff0c\u4f46\u89e3\u78bc\u5668-Only\u6a21\u578b\u7684\u751f\u6210\u80fd\u529b\u662f\u6709\u76ee\u5171\u7779\u7684\u3002\u5f9e<strong>ChatGPT<\/strong>\uff08\u57fa\u65bcGPT-3.5\u548cGPT-4\uff09\u5f15\u767c\u7684\u5168\u7403\u71b1\u6f6e\uff0c\u5230\u5728\u4ee3\u78bc\u751f\u6210\u9818\u57df\u7684\u5ee3\u6cdb\u61c9\u7528\uff08\u5982GitHub Copilot\uff09\uff0c\u518d\u5230\u5404\u7a2e\u5275\u610f\u5beb\u4f5c\u548c\u5167\u5bb9\u5275\u4f5c\u5de5\u5177\uff0c\u90fd\u8b49\u660e\u4e86\u9019\u4e00\u67b6\u69cb\u5728\u751f\u6210\u4efb\u52d9\u4e0a\u7684\u7d55\u5c0d\u7d71\u6cbb\u529b \u3002 \u00a0<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u7b2c\u4e09\u7ae0\uff1a\u7d9c\u5408\u7684\u6642\u4ee3 &#8211; \u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u6587\u85dd\u5fa9\u8208<\/h2>\n\n\n\n<p>\u5728\u7de8\u78bc\u5668-Only\u548c\u89e3\u78bc\u5668-Only\u5169\u5927\u738b\u671d\u5404\u81ea\u8d70\u5411\u5c08\u7cbe\u5316\u7684\u540c\u6642\uff0c\u4e00\u80a1\u8a66\u5716\u300c\u518d\u878d\u5408\u300d\u7684\u529b\u91cf\u4e5f\u5728\u6084\u7136\u8208\u8d77\u3002\u9019\u80a1\u529b\u91cf\u56de\u6eaf\u5230\u539f\u59cbTransformer\u7684\u8a2d\u8a08\uff0c\u65e8\u5728\u6253\u9020\u51fa\u65e2\u5177\u5099\u7de8\u78bc\u5668\u6df1\u5ea6\u7406\u89e3\u80fd\u529b\uff0c\u53c8\u64c1\u6709\u89e3\u78bc\u5668\u9748\u6d3b\u751f\u6210\u80fd\u529b\u7684\u901a\u7528\u6a21\u578b\u3002\u9019\u5834\u300c\u6587\u85dd\u5fa9\u8208\u300d\u7684\u4ee3\u8868\u4f5c\u662fT5\u548cBART\uff0c\u5b83\u5011\u901a\u904e\u5275\u65b0\u7684\u6846\u67b6\u548c\u9810\u8a13\u7df4\u76ee\u6a19\uff0c\u70ba\u5e8f\u5217\u5230\u5e8f\u5217\uff08Sequence-to-Sequence\uff09\u4efb\u52d9\u6a39\u7acb\u4e86\u65b0\u7684\u6a19\u6746\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">T5\u7684\u7d71\u4e00\u6230\u7dda\uff1a\u300cText-to-Text\u300d\u6846\u67b6\u6db5\u84cb\u4e00\u5207<\/h3>\n\n\n\n<p>2019\u5e74\uff0cGoogle\u7684\u7814\u7a76\u4eba\u54e1\u63a8\u51fa\u4e86T5\uff08Text-to-Text Transfer Transformer\uff09\uff0c\u5176\u6838\u5fc3\u601d\u60f3\u6975\u5177\u985b\u8986\u6027\uff1a\u5c07\u6240\u6709NLP\u4efb\u52d9\u90fd\u7d71\u4e00\u5230\u4e00\u500b\u55ae\u4e00\u7684\u300c\u6587\u672c\u5230\u6587\u672c\u300d\u6846\u67b6\u4e2d <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u6838\u5fc3\u54f2\u5b78<\/h4>\n\n\n\n<p>T5\u7684\u54f2\u5b78\u662f\uff0c\u7121\u8ad6\u662f\u7ffb\u8b6f\u3001\u5206\u985e\u3001\u554f\u7b54\u9084\u662f\u6458\u8981\uff0c\u6240\u6709\u4efb\u52d9\u90fd\u53ef\u4ee5\u88ab\u91cd\u65b0\u8868\u8ff0\u70ba\u4e00\u500b\u6a21\u578b\u63a5\u6536\u4e00\u6bb5\u6587\u672c\u8f38\u5165\uff0c\u4e26\u751f\u6210\u4e00\u6bb5\u65b0\u7684\u6587\u672c\u8f38\u51fa\u7684\u904e\u7a0b <sup><\/sup>\u3002\u4f8b\u5982\uff1a &nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7ffb\u8b6f<\/strong>\uff1a\u8f38\u5165<code>translate English to German: That is good.<\/code>\uff0c\u6a21\u578b\u61c9\u8f38\u51fa<code>Das ist gut.<\/code>\u3002<\/li>\n\n\n\n<li><strong>\u5206\u985e\uff08\u60c5\u611f\u5206\u6790\uff09<\/strong>\uff1a\u8f38\u5165<code>sentiment: This movie is fantastic!<\/code>\uff0c\u6a21\u578b\u61c9\u8f38\u51fa<code>positive<\/code>\u3002<\/li>\n\n\n\n<li><strong>\u6458\u8981<\/strong>\uff1a\u8f38\u5165<code>summarize: [long article text]...<\/code>\uff0c\u6a21\u578b\u61c9\u8f38\u51fa<code>[short summary text]<\/code>\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u9019\u7a2e\u7d71\u4e00\u7684\u6846\u67b6\u5e36\u4f86\u4e86\u5de8\u5927\u7684\u597d\u8655\uff1a\u7814\u7a76\u4eba\u54e1\u53ef\u4ee5\u4f7f\u7528\u5b8c\u5168\u76f8\u540c\u7684\u6a21\u578b\u67b6\u69cb\u3001\u640d\u5931\u51fd\u6578\u3001\u8d85\u53c3\u6578\u548c\u89e3\u78bc\u904e\u7a0b\u4f86\u8655\u7406\u6975\u5176\u591a\u6a23\u5316\u7684\u4efb\u52d9\uff0c\u6975\u5927\u5730\u7c21\u5316\u4e86\u5be6\u9a57\u548c\u90e8\u7f72\u6d41\u7a0b <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u67b6\u69cb\u8207\u9810\u8a13\u7df4\u76ee\u6a19<\/h4>\n\n\n\n<p>T5\u63a1\u7528\u4e86\u6a19\u6e96\u7684Transformer\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u67b6\u69cb <sup><\/sup>\u3002\u5176\u4e3b\u8981\u7684\u9810\u8a13\u7df4\u76ee\u6a19\u88ab\u7a31\u70ba &nbsp;<\/p>\n\n\n\n<p><strong>\u8de8\u5ea6\u640d\u58de\uff08Span Corruption\uff09<\/strong>\uff0c\u9019\u662f\u4e00\u7a2e\u66f4\u63a5\u8fd1\u771f\u5be6\u4e16\u754c\u566a\u8072\u7684MLM\u8b8a\u9ad4 <sup><\/sup>\u3002T5\u4e0d\u662f\u63a9\u84cb\u55ae\u500b\u8a5e\u5143\uff0c\u800c\u662f\u96a8\u6a5f\u5730\u5f9e\u8f38\u5165\u6587\u672c\u4e2d\u79fb\u9664\u9023\u7e8c\u7684\u6587\u672c\u8de8\u5ea6\uff08span\uff09\uff0c\u4e26\u7528\u4e00\u500b\u55ae\u4e00\u7684\u7279\u6b8a\u54e8\u5175\u8a5e\u5143\uff08sentinel token\uff09\u66ff\u63db\u5b83\u5011\u3002\u6a21\u578b\u7684\u4efb\u52d9\u5247\u662f\u9810\u6e2c\u51fa\u6240\u6709\u88ab\u79fb\u9664\u7684\u539f\u59cb\u6587\u672c\u8de8\u5ea6\uff0c\u4e26\u7528\u76f8\u61c9\u7684\u54e8\u5175\u8a5e\u5143\u9032\u884c\u5206\u9694\u3002\u9019\u7a2e\u76ee\u6a19\u8feb\u4f7f\u6a21\u578b\u4e0d\u50c5\u8981\u586b\u88dc\u7f3a\u5931\u7684\u5167\u5bb9\uff0c\u9084\u8981\u5b78\u6703\u8655\u7406\u9577\u5ea6\u548c\u7d50\u69cb\u90fd\u6709\u8b8a\u5316\u7684\u8f38\u5165\u548c\u8f38\u51fa\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">C4\u6578\u64da\u96c6<\/h4>\n\n\n\n<p>\u70ba\u4e86\u652f\u6301\u5927\u898f\u6a21\u7684\u9810\u8a13\u7df4\uff0cT5\u5718\u968a\u9084\u5275\u5efa\u4e26\u958b\u6e90\u4e86**C4\uff08Colossal Clean Crawled Corpus\uff09**\u6578\u64da\u96c6 <sup><\/sup>\u3002C4\u662f\u5c0dCommon Crawl\u7db2\u7d61\u722c\u53d6\u6578\u64da\u9032\u884c\u5927\u91cf\u6e05\u6d17\u548c\u904e\u6ffe\u5f8c\u5f97\u5230\u7684\uff0c\u65e8\u5728\u63d0\u4f9b\u4e00\u500b\u9ad8\u8cea\u91cf\u7684\u3001\u901a\u7528\u7684\u5927\u898f\u6a21\u6587\u672c\u8a9e\u6599\u5eab\uff0c\u9019\u4e5f\u6210\u70ba\u4e86T5\u9805\u76ee\u5c0d\u793e\u5340\u7684\u91cd\u5927\u8ca2\u737b\u4e4b\u4e00\u3002 &nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">BART\u7684\u54f2\u5b78\uff1a\u70ba\u6cdb\u5316\u800c\u751f\u7684\u53bb\u566a<\/h3>\n\n\n\n<p>\u5e7e\u4e4e\u5728\u540c\u4e00\u6642\u9593\uff0cFacebook AI\u7684\u7814\u7a76\u4eba\u54e1\u63d0\u51fa\u4e86BART\uff08Bidirectional and Auto-Regressive Transformer\uff09\uff0c\u9019\u662f\u53e6\u4e00\u500b\u5f37\u5927\u7684\u7d9c\u5408\u6a21\u578b <sup><\/sup>\u3002BART\u7684\u8a2d\u8a08\u7406\u5ff5\u662f\u660e\u78ba\u5730\u5c07BERT\u548cGPT\u7684\u512a\u9ede\u7d50\u5408\u8d77\u4f86\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u7d50\u5408\u5169\u5927\u738b\u671d\u7684\u512a\u52e2<\/h4>\n\n\n\n<p>BART\u7684\u67b6\u69cb\u53ef\u4ee5\u88ab\u8996\u70ba\u5c0dBERT\u548cGPT\u7684\u76f4\u63a5\u6982\u62ec <sup><\/sup>\u3002\u5b83\u4f7f\u7528\u4e86\u4e00\u500b &nbsp;<\/p>\n\n\n\n<p><strong>\u96d9\u5411\u7684\u7de8\u78bc\u5668<\/strong>\uff08\u985e\u4f3cBERT\uff09\u4f86\u8655\u7406\u88ab\u640d\u58de\u7684\u8f38\u5165\u6587\u672c\uff0c\u4f7f\u5176\u80fd\u5920\u5145\u5206\u7406\u89e3\u8f38\u5165\u7684\u5168\u5c40\u4e0a\u4e0b\u6587\u3002\u7136\u5f8c\uff0c\u5b83\u4f7f\u7528\u4e00\u500b<strong>\u5f9e\u5de6\u5230\u53f3\u7684\u81ea\u56de\u6b78\u89e3\u78bc\u5668<\/strong>\uff08\u985e\u4f3cGPT\uff09\u4f86\u9010\u6b65\u91cd\u5efa\u539f\u59cb\u7684\u3001\u672a\u640d\u58de\u7684\u6587\u672c <sup><\/sup>\u3002\u9019\u7a2e\u8a2d\u8a08\u4f7f\u5f97BART\u5728\u9700\u8981\u5c0d\u6e90\u6587\u672c\u6709\u6df1\u523b\u7406\u89e3\u7684\u751f\u6210\u4efb\u52d9\u4e0a\uff08\u5982\u6458\u8981\uff09\u5177\u6709\u5929\u7136\u7684\u512a\u52e2\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u53bb\u566a\u81ea\u7de8\u78bc\u5668\u9810\u8a13\u7df4\u76ee\u6a19<\/h4>\n\n\n\n<p>BART\u7684\u9810\u8a13\u7df4\u904e\u7a0b\u88ab\u5b9a\u7fa9\u70ba\u4e00\u500b<strong>\u53bb\u566a\u81ea\u7de8\u78bc\u5668\uff08denoising autoencoder\uff09<\/strong>\u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f\uff0c\u9996\u5148\u7528\u4efb\u610f\u7684\u566a\u8072\u51fd\u6578\u300c\u640d\u58de\u300d\u539f\u59cb\u6587\u672c\uff0c\u7136\u5f8c\u8a13\u7df4\u6a21\u578b\u4f86\u6062\u5fa9\u539f\u59cb\u6587\u672c <sup><\/sup>\u3002\u9019\u7a2e\u9748\u6d3b\u7684\u6846\u67b6\u5141\u8a31\u7814\u7a76\u4eba\u54e1\u5617\u8a66\u5404\u7a2e\u5404\u6a23\u7684\u6587\u672c\u640d\u58de\u7b56\u7565\uff0c\u9060\u6bd4BERT\u7684MLM\u66f4\u70ba\u8c50\u5bcc\uff1a &nbsp;<\/p>\n\n\n\n<ol start=\"1\" class=\"wp-block-list\">\n<li><strong>\u8a5e\u5143\u63a9\u78bc\uff08Token Masking\uff09<\/strong>\uff1a\u8207BERT\u76f8\u540c\u3002<\/li>\n\n\n\n<li><strong>\u8a5e\u5143\u522a\u9664\uff08Token Deletion\uff09<\/strong>\uff1a\u96a8\u6a5f\u522a\u9664\u8a5e\u5143\uff0c\u6a21\u578b\u9700\u8981\u5b78\u6703\u5728\u54ea\u88e1\u63d2\u5165\u7f3a\u5931\u7684\u5167\u5bb9\u3002<\/li>\n\n\n\n<li><strong>\u6587\u672c\u586b\u5145\uff08Text Infilling\uff09<\/strong>\uff1a\u5c07\u4e00\u500b\u6216\u591a\u500b\u6587\u672c\u8de8\u5ea6\u66ff\u63db\u70ba\u55ae\u4e00\u7684&#8220;\u8a5e\u5143\u3002\u6a21\u578b\u9700\u8981\u9810\u6e2c\u51fa\u88ab\u66ff\u63db\u7684\u5167\u5bb9\u4ee5\u53ca\u5176\u9577\u5ea6\u3002<\/li>\n\n\n\n<li><strong>\u53e5\u5b50\u6392\u5217\uff08Sentence Permutation\uff09<\/strong>\uff1a\u96a8\u6a5f\u6253\u4e82\u6587\u6a94\u4e2d\u53e5\u5b50\u7684\u9806\u5e8f\uff0c\u6a21\u578b\u9700\u8981\u6062\u5fa9\u539f\u59cb\u9806\u5e8f\u3002<\/li>\n\n\n\n<li><strong>\u6587\u6a94\u65cb\u8f49\uff08Document Rotation\uff09<\/strong>\uff1a\u96a8\u6a5f\u9078\u64c7\u4e00\u500b\u8a5e\u5143\u4f5c\u70ba\u6587\u6a94\u7684\u958b\u982d\uff0c\u6a21\u578b\u9700\u8981\u8b58\u5225\u51fa\u539f\u59cb\u7684\u6587\u6a94\u8d77\u9ede\u3002<\/li>\n<\/ol>\n\n\n\n<p>BART\u7684\u8ad6\u6587\u767c\u73fe\uff0c\u5c07<strong>\u6587\u672c\u586b\u5145<\/strong>\u548c<strong>\u53e5\u5b50\u6392\u5217<\/strong>\u9019\u5169\u7a2e\u640d\u58de\u7b56\u7565\u7d50\u5408\u4f7f\u7528\u6642\uff0c\u6a21\u578b\u5728\u4e0b\u6e38\u4efb\u52d9\u4e0a\u8868\u73fe\u6700\u597d <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7d71\u6cbb\u7684\u8b49\u64da\uff1a\u5e8f\u5217\u5230\u5e8f\u5217\u4efb\u52d9\u7684SOTA\u8868\u73fe<\/h3>\n\n\n\n<p>\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u6587\u85dd\u5fa9\u8208\u7684\u6210\u679c\uff0c\u6700\u76f4\u63a5\u5730\u9ad4\u73fe\u5728\u90a3\u4e9b\u9700\u8981\u540c\u6642\u9032\u884c\u6df1\u5ea6\u7406\u89e3\u548c\u8907\u96dc\u751f\u6210\u7684\u5e8f\u5217\u5230\u5e8f\u5217\u4efb\u52d9\u4e0a\u3002\u5728\u9019\u4e9b\u9818\u57df\uff0cT5\u548cBART\u90fd\u53d6\u5f97\u4e86\u7576\u6642\u7684SOTA\u6210\u7e3e\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6458\u8981\uff08CNN\/DailyMail &amp; XSum\uff09<\/strong>\uff1a\u62bd\u8c61\u5f0f\u6458\u8981\u662f\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u67b6\u69cb\u7684\u7d93\u5178\u61c9\u7528\u5834\u666f\u3002\u6a21\u578b\u9700\u8981\u300c\u95b1\u8b80\u300d\u4e00\u7bc7\u9577\u6587\uff08\u7531\u7de8\u78bc\u5668\u8655\u7406\uff09\uff0c\u7136\u5f8c\u300c\u5beb\u300d\u51fa\u4e00\u6bb5\u7c21\u6f54\u7684\u6458\u8981\uff08\u7531\u89e3\u78bc\u5668\u751f\u6210\uff09\u3002T5\u548cBART\u5728\u9019\u500b\u4efb\u52d9\u4e0a\u8868\u73fe\u51fa\u8272\uff0c\u5176\u6027\u80fd\u901a\u5e38\u4f7f\u7528<strong>ROUGE<\/strong>\uff08Recall-Oriented Understudy for Gisting Evaluation\uff09\u5206\u6578\u4f86\u8861\u91cf\uff0c\u8a72\u6307\u6a19\u8a08\u7b97\u751f\u6210\u6458\u8981\u8207\u53c3\u8003\u6458\u8981\u4e4b\u9593\u7684n-gram\u91cd\u758a\u5ea6 \u3002BART\u5728\u5176\u8ad6\u6587\u4e2d\u5831\u544a\uff0c\u5728\u6458\u8981\u3001\u5c0d\u8a71\u548c\u554f\u7b54\u7b49\u4e00\u7cfb\u5217\u751f\u6210\u4efb\u52d9\u4e0a\u53d6\u5f97\u4e86\u9ad8\u90543.5 ROUGE\u7684\u63d0\u5347 \u3002T5-11B\u5728CNN\/DailyMail\u4e0a\u7684ROUGE-L\u5206\u6578\u9054\u5230\u4e8640.69\uff0c\u662f\u7576\u6642\u7684\u9802\u7d1a\u6c34\u5e73 \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u6a5f\u5668\u7ffb\u8b6f\uff08WMT\uff09<\/strong>\uff1a\u6a5f\u5668\u7ffb\u8b6f\u662f\u539f\u59cbTransformer\u7684\u8a95\u751f\u4e4b\u5730\uff0c\u4e5f\u662f\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u67b6\u69cb\u7684\u50b3\u7d71\u5f37\u9805\u3002T5\u548cBART\u8b49\u660e\uff0c\u901a\u904e\u5927\u898f\u6a21\u9810\u8a13\u7df4\u548c\u5fae\u8abf\uff0c\u9019\u4e00\u67b6\u69cb\u5728\u7ffb\u8b6f\u4efb\u52d9\u4e0a\u4f9d\u7136\u6975\u5177\u7af6\u722d\u529b\u3002\u5176\u6027\u80fd\u901a\u5e38\u4f7f\u7528<strong>BLEU<\/strong>\uff08Bilingual Evaluation Understudy\uff09\u5206\u6578\u4f86\u8a55\u4f30\uff0c\u8a72\u6307\u6a19\u8861\u91cf\u6a5f\u5668\u7ffb\u8b6f\u8207\u5c08\u696d\u4eba\u5de5\u7ffb\u8b6f\u4e4b\u9593\u7684\u76f8\u4f3c\u5ea6 \u3002BART\u5728\u5176\u767c\u5e03\u6642\uff0c\u50c5\u901a\u904e\u76ee\u6a19\u8a9e\u8a00\u7684\u9810\u8a13\u7df4\uff0c\u5c31\u5728\u6a5f\u5668\u7ffb\u8b6f\u4efb\u52d9\u4e0a\u6bd4\u4e00\u500b\u5f37\u5927\u7684\u53cd\u5411\u7ffb\u8b6f\uff08back-translation\uff09\u7cfb\u7d71\u9ad8\u51fa1.1 BLEU \u3002\u539f\u59cbTransformer\u5728WMT 2014\u82f1\u5fb7\u7ffb\u8b6f\u4efb\u52d9\u4e0a\u7684BLEU\u5206\u6578\u70ba28.4\uff0c\u800cT5\u548cBART\u7684\u8b8a\u9ad4\u5247\u5728\u6b64\u57fa\u790e\u4e0a\u9032\u4e00\u6b65\u63d0\u5347\u4e86\u6027\u80fd \u3002 \u00a0<\/li>\n<\/ul>\n\n\n\n<p>\u9019\u4e9b\u5be6\u8b49\u7d50\u679c\u6e05\u6670\u5730\u8868\u660e\uff0c\u5118\u7ba1\u5c08\u7cbe\u5316\u7684\u7de8\u78bc\u5668-Only\u548c\u89e3\u78bc\u5668-Only\u6a21\u578b\u5728\u5404\u81ea\u7684\u9818\u57df\u7a31\u9738\uff0c\u4f46\u7d9c\u5408\u6027\u7684\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u67b6\u69cb\u5728\u90a3\u4e9b\u9700\u8981\u5169\u8005\u80fd\u529b\u7d50\u5408\u7684\u8907\u96dc\u8f49\u63db\u4efb\u52d9\u4e2d\uff0c\u627e\u5230\u4e86\u81ea\u5df1\u4e0d\u53ef\u66ff\u4ee3\u7684\u751f\u614b\u4f4d\u3002<\/p>\n\n\n\n<p>\u9019\u5834\u6587\u85dd\u5fa9\u8208\u4e5f\u63ed\u793a\u4e86AI\u67b6\u69cb\u8a2d\u8a08\u4e2d\u7684\u300c\u6c92\u6709\u514d\u8cbb\u7684\u5348\u9910\u300d\uff08No Free Lunch\uff09\u539f\u5247\u3002T5\u548cBART\u662f\u51fa\u8272\u7684\u591a\u9762\u624b\u548c\u5e8f\u5217\u5230\u5e8f\u5217\u4efb\u52d9\u7684\u5927\u5e2b\uff0c\u4f46\u5b83\u5011\u7684\u901a\u7528\u6027\u4e5f\u610f\u5473\u8457\u5728\u67d0\u4e9b\u6975\u7aef\u5c08\u7cbe\u7684\u4efb\u52d9\u4e0a\uff0c\u53ef\u80fd\u6703\u88ab\u7279\u5316\u6a21\u578b\u8d85\u8d8a\u3002\u4f8b\u5982\uff0c\u5728SuperGLUE\u9019\u6a23\u7684\u7d14NLU\u57fa\u6e96\u6e2c\u8a66\u4e2d\uff0c\u4e00\u500b\u7cbe\u5fc3\u512a\u5316\u7684\u7de8\u78bc\u5668-Only\u6a21\u578b\uff08\u5982RoBERTa\uff09\u901a\u5e38\u8868\u73fe\u66f4\u4f73 <sup><\/sup>\uff1b\u800c\u5728\u9700\u8981\u6975\u81f4\u5275\u9020\u529b\u548c\u96f6\u6a23\u672c\u6cdb\u5316\u80fd\u529b\u7684\u958b\u653e\u5f0f\u751f\u6210\u4efb\u52d9\u4e0a\uff0c\u4e00\u500b\u898f\u6a21\u5de8\u5927\u7684\u89e3\u78bc\u5668-Only\u6a21\u578b\uff08\u5982GPT-3\uff09\u5247\u7121\u53ef\u5339\u6575 <sup><\/sup>\u3002\u9019\u51f8\u986f\u4e86\u5728LLM\u67b6\u69cb\u8a2d\u8a08\u4e2d\uff0c\u5c08\u7cbe\u5316\u8207\u6cdb\u5316\u80fd\u529b\u4e4b\u9593\u6c38\u6046\u7684\u6b0a\u8861\u3002 &nbsp;<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u7b2c\u56db\u7ae0\uff1a\u67b6\u69cb\u6df1\u5ea6\u5256\u6790\uff1a\u4e09\u65b9\u5c0d\u6bd4<\/h2>\n\n\n\n<p>\u5728\u56de\u9867\u4e86\u4e09\u5927\u738b\u671d\u5404\u81ea\u7684\u8208\u8870\u53f2\u5f8c\uff0c\u672c\u7ae0\u5c07\u5c0d\u9019\u4e09\u7a2e\u6838\u5fc3\u67b6\u69cb\u9032\u884c\u4e00\u6b21\u76f4\u63a5\u7684\u3001\u6280\u8853\u5c64\u9762\u7684\u6df1\u5ea6\u6bd4\u8f03\uff0c\u5f9e\u4fe1\u606f\u6d41\u3001\u6ce8\u610f\u529b\u6a5f\u5236\u548c\u9810\u8a13\u7df4\u76ee\u6a19\u4e09\u500b\u7dad\u5ea6\uff0c\u63ed\u793a\u5b83\u5011\u7684\u672c\u8cea\u5dee\u7570\u548c\u7531\u6b64\u7522\u751f\u7684\u80fd\u529b\u5206\u91ce\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4fe1\u606f\u6d41\u8207\u6ce8\u610f\u529b\u6a5f\u5236<\/h3>\n\n\n\n<p>\u6a21\u578b\u5982\u4f55\u8655\u7406\u548c\u50b3\u905e\u4fe1\u606f\uff0c\u662f\u5176\u67b6\u69cb\u7684\u6839\u672c\u3002\u4e09\u5927\u738b\u671d\u5728\u9019\u65b9\u9762\u6709\u8457\u622a\u7136\u4e0d\u540c\u7684\u8a2d\u8a08\u54f2\u5b78\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7de8\u78bc\u5668-Only\uff08\u96d9\u5411\u4fe1\u606f\u6d41\uff09<\/strong>\uff1a\u5728BERT\u53ca\u5176\u5f8c\u7e7c\u8005\u4e2d\uff0c\u4fe1\u606f\u6d41\u662f\u5b8c\u5168<strong>\u96d9\u5411\u7684\uff08bidirectional\uff09<\/strong>\u3002\u5728\u5176\u6838\u5fc3\u7684\u81ea\u8a3b\u610f\u529b\u5c64\u4e2d\uff0c\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e00\u500b\u8a5e\u5143\u90fd\u53ef\u4ee5\u95dc\u6ce8\u5230\u5e8f\u5217\u4e2d\u7684\u6240\u6709\u5176\u4ed6\u8a5e\u5143\uff0c\u7121\u8ad6\u662f\u5728\u5176\u4e4b\u524d\u9084\u662f\u4e4b\u5f8c\u3002\u9019\u5c31\u597d\u50cf\u5728\u95b1\u8b80\u4e00\u500b\u53e5\u5b50\u6642\uff0c\u53ef\u4ee5\u53cd\u8986\u901a\u8b80\u5168\u6587\uff0c\u76f4\u5230\u5c0d\u6bcf\u500b\u8a5e\u5728\u5b8c\u6574\u8a9e\u5883\u4e2d\u7684\u542b\u7fa9\u90fd\u6709\u4e86\u6df1\u523b\u7684\u7406\u89e3\u3002\u9019\u7a2e\u7121\u9650\u5236\u7684\u4fe1\u606f\u6d41\u52d5\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u5920\u69cb\u5efa\u51fa\u5c0d\u8f38\u5165\u6587\u672c\u6975\u5176\u8c50\u5bcc\u548c\u6df1\u5165\u7684\u4e0a\u4e0b\u6587\u8868\u793a\u3002\u4f8b\u5982\uff0c\u8981\u5340\u5206\u300criver bank\u300d\uff08\u6cb3\u5cb8\uff09\u548c\u300cmoney bank\u300d\uff08\u9280\u884c\uff09\u4e2d\u7684\u300cbank\u300d\uff0c\u6a21\u578b\u53ef\u4ee5\u540c\u6642\u5229\u7528\u300criver\u300d\u548c\u300cmoney\u300d\u9019\u5169\u500b\u95dc\u9375\u8a5e\uff0c\u7121\u8ad6\u5b83\u5011\u51fa\u73fe\u5728\u53e5\u5b50\u7684\u54ea\u500b\u4f4d\u7f6e\u3002<\/li>\n\n\n\n<li><strong>\u89e3\u78bc\u5668-Only\uff08\u56e0\u679c\/\u81ea\u56de\u6b78\u4fe1\u606f\u6d41\uff09<\/strong>\uff1a\u5728GPT\u7cfb\u5217\u6a21\u578b\u4e2d\uff0c\u4fe1\u606f\u6d41\u662f\u56b4\u683c<strong>\u55ae\u5411\u7684\uff08unidirectional\uff09\u6216\u56e0\u679c\u7684\uff08causal\uff09<\/strong>\u3002\u5176\u81ea\u8a3b\u610f\u529b\u6a5f\u5236\u88ab\u65bd\u52a0\u4e86\u63a9\u78bc\uff0c\u4f7f\u5f97\u4e00\u500b\u4f4d\u65bc\u4f4d\u7f6ei\u7684\u8a5e\u5143\u53ea\u80fd\u95dc\u6ce8\u5230\u5176\u81ea\u8eab\u4ee5\u53ca\u4e4b\u524d\u7684\u6240\u6709\u8a5e\u5143\uff08\u4f4d\u7f6e1\u5230i\uff09\u3002\u5b83\u5c0d\u672a\u4f86\u7684\u8a5e\u5143\u662f\u300c\u76f2\u300d\u7684 \u3002\u9019\u7a2e\u8a2d\u8a08\u5f37\u5236\u6a21\u578b\u9075\u5faa\u5f9e\u5de6\u5230\u53f3\u7684\u81ea\u56de\u6b78\u751f\u6210\u904e\u7a0b\uff0c\u5373\u6839\u64da\u5df2\u7d93\u751f\u6210\u7684\u5167\u5bb9\u4f86\u9810\u6e2c\u4e0b\u4e00\u500b\u5167\u5bb9\u3002\u9019\u5c31\u50cf\u4e00\u500b\u4f5c\u5bb6\u5728\u5beb\u4f5c\u6642\uff0c\u53ea\u80fd\u57fa\u65bc\u5df2\u7d93\u5beb\u4e0b\u7684\u6587\u5b57\u4f86\u69cb\u601d\u4e0b\u4e00\u53e5\u8a71\uff0c\u800c\u4e0d\u80fd\u9810\u77e5\u672a\u4f86\u7684\u60c5\u7bc0\u3002\u9019\u7a2e\u4fe1\u606f\u6d41\u5c0d\u65bc\u751f\u6210\u9023\u8cab\u3001\u6d41\u66a2\u7684\u6587\u672c\u81f3\u95dc\u91cd\u8981\u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\uff08\u6df7\u5408\u4fe1\u606f\u6d41\uff09<\/strong>\uff1a\u9019\u7a2e\u67b6\u69cb\u64c1\u6709\u6700\u8907\u96dc\u7684\u4fe1\u606f\u6d41\uff0c\u878d\u5408\u4e86\u524d\u5169\u8005\u7684\u7279\u9ede\uff0c\u4e26\u5f15\u5165\u4e86\u7b2c\u4e09\u7a2e\u4ea4\u4e92\u65b9\u5f0f\u3002\n<ol start=\"1\" class=\"wp-block-list\">\n<li><strong>\u7de8\u78bc\u5668\u4e2d\u7684\u96d9\u5411\u81ea\u8a3b\u610f\u529b<\/strong>\uff1a\u7de8\u78bc\u5668\u90e8\u5206\u8207BERT\u4e00\u6a23\uff0c\u5c0d\u8f38\u5165\u7684\u6e90\u5e8f\u5217\uff08source sequence\uff09\u9032\u884c\u96d9\u5411\u8655\u7406\uff0c\u4ee5\u5f62\u6210\u4e00\u500b\u5b8c\u6574\u7684\u4e0a\u4e0b\u6587\u8868\u793a\u3002<\/li>\n\n\n\n<li><strong>\u89e3\u78bc\u5668\u4e2d\u7684\u56e0\u679c\u81ea\u8a3b\u610f\u529b<\/strong>\uff1a\u89e3\u78bc\u5668\u90e8\u5206\u8207GPT\u4e00\u6a23\uff0c\u5c0d\u5176\u81ea\u8eab\u5df2\u7d93\u751f\u6210\u7684\u76ee\u6a19\u5e8f\u5217\uff08target sequence\uff09\u9032\u884c\u56e0\u679c\u81ea\u8a3b\u610f\u529b\u8a08\u7b97\uff0c\u4ee5\u78ba\u4fdd\u751f\u6210\u7684\u9023\u8cab\u6027\u3002<\/li>\n\n\n\n<li><strong>\u4ea4\u53c9\u6ce8\u610f\u529b\uff08Cross-Attention\uff09<\/strong>\uff1a\u9019\u662f\u9023\u63a5\u7de8\u78bc\u5668\u548c\u89e3\u78bc\u5668\u7684\u6a4b\u6a11\uff0c\u4e5f\u662f\u8a72\u67b6\u69cb\u7684\u95dc\u9375\u6240\u5728 \u3002\u5728\u89e3\u78bc\u5668\u7684\u6bcf\u4e00\u5c64\uff0c\u9664\u4e86\u56e0\u679c\u81ea\u8a3b\u610f\u529b\uff0c\u9084\u6709\u4e00\u500b\u4ea4\u53c9\u6ce8\u610f\u529b\u5b50\u5c64\u3002\u5728\u9019\u500b\u5b50\u5c64\u4e2d\uff0c\u89e3\u78bc\u5668\u751f\u6210\u7684\u67e5\u8a62\uff08Q\uff09\u5411\u91cf\u6703\u53bb\u95dc\u6ce8\u7de8\u78bc\u5668\u6700\u7d42\u8f38\u51fa\u7684\u9375\uff08K\uff09\u548c\u503c\uff08V\uff09\u5411\u91cf\u3002\u9019\u5141\u8a31\u89e3\u78bc\u5668\u5728\u751f\u6210\u6bcf\u4e00\u500b\u65b0\u8a5e\u5143\u6642\uff0c\u90fd\u80fd\u300c\u56de\u770b\u300d\u4e26\u501f\u9452\u6e90\u5e8f\u5217\u4e2d\u7684\u6240\u6709\u4fe1\u606f\u3002\u9019\u5c0d\u65bc\u7ffb\u8b6f\uff08\u5c07\u76ee\u6a19\u8a5e\u8207\u6e90\u8a5e\u5c0d\u9f4a\uff09\u548c\u6458\u8981\uff08\u78ba\u4fdd\u6458\u8981\u5167\u5bb9\u5fe0\u5be6\u65bc\u539f\u6587\uff09\u7b49\u4efb\u52d9\u662f\u4e0d\u53ef\u6216\u7f3a\u7684\u3002 \u00a0<\/li>\n<\/ol>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u9810\u8a13\u7df4\u76ee\u6a19\uff1a\u6a5f\u5668\u7684\u9748\u9b42<\/h3>\n\n\n\n<p>\u5982\u679c\u8aaa\u67b6\u69cb\u662f\u6a21\u578b\u7684\u9aa8\u67b6\uff0c\u90a3\u9ebc\u9810\u8a13\u7df4\u76ee\u6a19\u5c31\u662f\u6ce8\u5165\u5176\u9748\u9b42\u7684\u5100\u5f0f\u3002\u4e0d\u540c\u7684\u76ee\u6a19\u5851\u9020\u4e86\u6a21\u578b\u622a\u7136\u4e0d\u540c\u7684\u300c\u4e16\u754c\u89c0\u300d\u548c\u6838\u5fc3\u80fd\u529b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>MLM\uff08\u63a9\u78bc\u8a9e\u8a00\u5efa\u6a21\uff09<\/strong>\uff1a\u9019\u662f\u7de8\u78bc\u5668-Only\u6a21\u578b\u7684\u5c08\u5c6c\u76ee\u6a19\u3002\u5176\u6839\u672c\u76ee\u7684\u4e0d\u662f\u751f\u6210\uff0c\u800c\u662f<strong>\u7406\u89e3<\/strong>\u548c<strong>\u8868\u793a<\/strong>\u3002\u5b83\u901a\u904e\u300c\u5b8c\u5f62\u586b\u7a7a\u300d\u7684\u65b9\u5f0f\uff0c\u8a13\u7df4\u6a21\u578b\u53bb\u63a8\u65b7\u7f3a\u5931\u7684\u4fe1\u606f\uff0c\u5f9e\u800c\u8feb\u4f7f\u6a21\u578b\u5b78\u7fd2\u8a5e\u8207\u8a5e\u4e4b\u9593\u6df1\u5c64\u6b21\u7684\u8a9e\u7fa9\u548c\u8a9e\u6cd5\u95dc\u4fc2\u3002MLM\u8a13\u7df4\u51fa\u7684\u6a21\u578b\uff0c\u5176\u6700\u7d42\u8f38\u51fa\u662f\u4e00\u500b\u9ad8\u8cea\u91cf\u7684\u4e0a\u4e0b\u6587\u5d4c\u5165\u5411\u91cf\uff0c\u975e\u5e38\u9069\u5408\u7528\u65bc\u5206\u985e\u3001\u5be6\u9ad4\u8b58\u5225\u7b49\u5224\u5225\u6027\u4efb\u52d9\u3002<\/li>\n\n\n\n<li><strong>CLM\uff08\u56e0\u679c\u8a9e\u8a00\u5efa\u6a21\uff09<\/strong>\uff1a\u9019\u662f\u89e3\u78bc\u5668-Only\u6a21\u578b\u7684\u6a19\u6e96\u76ee\u6a19\u3002\u5176\u76ee\u7684\u975e\u5e38\u7d14\u7cb9\uff1a<strong>\u751f\u6210<\/strong>\u3002\u901a\u904e\u4e0d\u65b7\u5730\u9810\u6e2c\u4e0b\u4e00\u500b\u8a5e\uff0c\u6a21\u578b\u5b78\u6703\u4e86\u8a9e\u8a00\u7684\u6982\u7387\u5206\u5e03\uff0c\u638c\u63e1\u4e86\u8a9e\u6cd5\u3001\u98a8\u683c\u548c\u4e8b\u5be6\u77e5\u8b58\u3002CLM\u8a13\u7df4\u51fa\u7684\u6a21\u578b\u5929\u7136\u5c31\u662f\u4e00\u500b\u6587\u672c\u751f\u6210\u5668\uff0c\u64c5\u9577\u958b\u653e\u5f0f\u7684\u3001\u5275\u9020\u6027\u7684\u4efb\u52d9\u3002<\/li>\n\n\n\n<li><strong>\u53bb\u566a\/\u8de8\u5ea6\u640d\u58de\uff08Denoising \/ Span Corruption\uff09<\/strong>\uff1a\u9019\u662f\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u6a21\u578b\u7684\u6df7\u5408\u76ee\u6a19\u3002\u5b83\u7d50\u5408\u4e86\u7406\u89e3\u548c\u751f\u6210\u7684\u96d9\u91cd\u6311\u6230\u3002\u6a21\u578b\u9996\u5148\u9700\u8981\u901a\u904e\u7de8\u78bc\u5668\u7406\u89e3\u4e00\u500b\u88ab\u300c\u640d\u58de\u300d\u6216\u4e0d\u5b8c\u6574\u7684\u8f38\u5165\uff0c\u7136\u5f8c\u901a\u904e\u89e3\u78bc\u5668\u751f\u6210\u4e00\u500b\u300c\u4e7e\u6de8\u300d\u7684\u3001\u5b8c\u6574\u7684\u8f38\u51fa\u3002\u9019\u7a2e**\u8f49\u63db\uff08transformation\uff09<strong>\u6216<\/strong>\u7de8\u8f2f\uff08editing\uff09**\u7684\u8a13\u7df4\u7bc4\u5f0f\uff0c\u4f7f\u5f97\u6a21\u578b\u975e\u5e38\u64c5\u9577\u90a3\u4e9b\u8f38\u5165\u548c\u8f38\u51fa\u5e8f\u5217\u4e4b\u9593\u5b58\u5728\u7dca\u5bc6\u6620\u5c04\u95dc\u4fc2\u7684\u4efb\u52d9\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u4efb\u52d9\u5c08\u7cbe\u77e9\u9663\u8207\u5be6\u8b49\u652f\u6301<\/h3>\n\n\n\n<p>\u7d9c\u5408\u4ee5\u4e0a\u5206\u6790\uff0c\u6211\u5011\u53ef\u4ee5\u69cb\u5efa\u4e00\u500b\u6e05\u6670\u7684\u4efb\u52d9\u5c08\u7cbe\u77e9\u9663\uff0c\u4e26\u7528\u524d\u5e7e\u7ae0\u63d0\u5230\u7684\u57fa\u6e96\u6e2c\u8a66\u7d50\u679c\u4f5c\u70ba\u5be6\u8b49\u652f\u6301\u3002<\/p>\n\n\n\n<p><strong>\u88682\uff1a\u4e09\u5927\u67b6\u69cb\u8207\u4efb\u52d9\u6027\u80fd\u5c0d\u6bd4<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><td>\u7279\u6027<\/td><td>\u7de8\u78bc\u5668-Only (\u4f8b\u5982, BERT, RoBERTa)<\/td><td>\u89e3\u78bc\u5668-Only (\u4f8b\u5982, GPT-3, Llama 2)<\/td><td>\u7de8\u78bc\u5668-\u89e3\u78bc\u5668 (\u4f8b\u5982, T5, BART)<\/td><\/tr><\/thead><tbody><tr><td><strong>\u6838\u5fc3\u539f\u5247<\/strong><\/td><td>\u96d9\u5411\u4e0a\u4e0b\u6587\u7406\u89e3<\/td><td>\u81ea\u56de\u6b78\u751f\u6210<\/td><td>\u5e8f\u5217\u5230\u5e8f\u5217\u8f49\u63db<\/td><\/tr><tr><td><strong>\u4fe1\u606f\u6d41<\/strong><\/td><td>\u7121\u9650\u5236\u96d9\u5411\u81ea\u8a3b\u610f\u529b<\/td><td>\u56e0\u679c\uff08\u55ae\u5411\uff09\u81ea\u8a3b\u610f\u529b<\/td><td><strong>\u7de8\u78bc\u5668<\/strong>: \u96d9\u5411\u81ea\u8a3b\u610f\u529b <strong>\u89e3\u78bc\u5668<\/strong>: \u56e0\u679c\u81ea\u8a3b\u610f\u529b + \u4ea4\u53c9\u6ce8\u610f\u529b<\/td><\/tr><tr><td><strong>\u4e3b\u8981\u9810\u8a13\u7df4\u76ee\u6a19<\/strong><\/td><td>\u63a9\u78bc\u8a9e\u8a00\u5efa\u6a21 (MLM)<\/td><td>\u56e0\u679c\u8a9e\u8a00\u5efa\u6a21 (CLM)<\/td><td>\u53bb\u566a \/ \u8de8\u5ea6\u640d\u58de<\/td><\/tr><tr><td><strong>\u64c5\u9577\u4efb\u52d9<\/strong><\/td><td><strong>\u81ea\u7136\u8a9e\u8a00\u7406\u89e3 (NLU)<\/strong>: &#8211; \u6587\u672c\u5206\u985e &#8211; \u60c5\u611f\u5206\u6790 &#8211; \u547d\u540d\u5be6\u9ad4\u8b58\u5225 (NER)<\/td><td><strong>\u958b\u653e\u5f0f\u751f\u6210<\/strong>: &#8211; \u5c0d\u8a71\u7cfb\u7d71\/\u804a\u5929\u6a5f\u5668\u4eba &#8211; \u5275\u610f\u5beb\u4f5c &#8211; \u4ee3\u78bc\u751f\u6210 &#8211; \u4e0a\u4e0b\u6587\u5b78\u7fd2<\/td><td><strong>\u5e8f\u5217\u5230\u5e8f\u5217\u4efb\u52d9<\/strong>: &#8211; \u62bd\u8c61\u5f0f\u6458\u8981 &#8211; \u6a5f\u5668\u7ffb\u8b6f &#8211; \u554f\u7b54 &#8211; \u6587\u672c\u98a8\u683c\u9077\u79fb<\/td><\/tr><tr><td><strong>\u4e3b\u8981\u5be6\u8b49 (\u57fa\u6e96 &amp; \u6307\u6a19)<\/strong><\/td><td><strong>GLUE\/SuperGLUE (\u5e73\u5747\u5206)<\/strong>: &#8211; BERT-Large (GLUE): 82.1 <sup><\/sup> &nbsp;<\/td><td>&#8211; RoBERTa\u5728\u591a\u9805\u4efb\u52d9\u4e0a\u8d85\u8d8aBERT <sup><\/sup> &nbsp;<\/td><td><strong>MMLU (\u6e96\u78ba\u7387)<\/strong>: &#8211; GPT-3 (davinci, 5-shot): 42.2% <sup><\/sup> &nbsp;<\/td><td>&#8211; Llama 2 (70B, 5-shot): 58.2% <sup><\/sup> &nbsp;<\/td><td><strong>WMT En-De (BLEU)<\/strong>: &#8211; Transformer: 28.4 <sup><\/sup> &nbsp;<\/td><td>&#8211; BART: \u8f03\u5f37\u57fa\u7dda\u63d0\u53471.1 <sup><\/sup> &nbsp;<\/td><td><strong>CNN\/DM (ROUGE-L)<\/strong>: &#8211; T5-11B: 40.69 <sup><\/sup> &nbsp;<\/td><td>&#8211; BART: \u8f03\u5f37\u57fa\u7dda\u63d0\u5347\u986f\u8457 <sup><\/sup> &nbsp;<\/td><\/tr><tr><td><strong>\u6838\u5fc3\u5f31\u9ede<\/strong><\/td><td>\u4e0d\u5929\u7136\u9069\u7528\u65bc\u751f\u6210\u4efb\u52d9\uff1b\u9700\u8981\u70ba\u4e0b\u6e38\u4efb\u52d9\u6dfb\u52a0\u7279\u5b9a\u7684\u300c\u982d\u300d\u90e8\u7d50\u69cb\u3002<\/td><td>\u5728\u9700\u8981\u6df1\u5ea6\u96d9\u5411\u7406\u89e3\u7684\u4efb\u52d9\u4e0a\u53ef\u80fd\u8868\u73fe\u4e0d\u4f73\uff1b\u5bb9\u6613\u7522\u751f\u5e7b\u89ba\uff08hallucination\uff09\u3002<\/td><td>\u5c0d\u65bc\u975e\u5e8f\u5217\u5230\u5e8f\u5217\u4efb\u52d9\uff0c\u5176\u67b6\u69cb\u6bd4\u5c08\u7cbe\u6a21\u578b\u66f4\u8907\u96dc\uff0c\u8a08\u7b97\u6210\u672c\u66f4\u9ad8\u3002<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u9019\u5f35\u8868\u683c\u6e05\u6670\u5730\u5c55\u793a\u4e86\u4e09\u5927\u67b6\u69cb\u4e4b\u9593\u7684\u5206\u5de5\u8207\u6b0a\u8861\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8b49\u64da\u8868\u660e<\/strong>\uff0c\u7576\u4efb\u52d9\u7684\u6838\u5fc3\u662f<strong>\u7406\u89e3<\/strong>\u8f38\u5165\u6587\u672c\u4e26\u5f9e\u4e2d\u63d0\u53d6\u4fe1\u606f\u6216\u5c0d\u5176\u9032\u884c\u5206\u985e\u6642\uff0c\u7de8\u78bc\u5668-Only\u6a21\u578b\u7121\u7591\u662f\u738b\u8005\u3002\u5b83\u5011\u5728GLUE\u548cSuperGLUE\u7b49NLU\u57fa\u6e96\u4e0a\u7684\u7d71\u6cbb\u5730\u4f4d\u5c31\u662f\u6700\u597d\u7684\u8b49\u660e \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u8b49\u64da\u8868\u660e<\/strong>\uff0c\u7576\u4efb\u52d9\u7684\u6838\u5fc3\u662f<strong>\u751f\u6210<\/strong>\u9023\u8cab\u3001\u591a\u6a23\u4e14\u8207\u4e0a\u4e0b\u6587\u76f8\u95dc\u7684\u6587\u672c\u6642\uff0c\u5927\u898f\u6a21\u7684\u89e3\u78bc\u5668-Only\u6a21\u578b\u8655\u65bc\u9818\u5148\u5730\u4f4d\u3002\u5b83\u5011\u5728MMLU\u7b49\u77e5\u8b58\u5bc6\u96c6\u578b\u57fa\u6e96\u4e0a\u7684\u9ad8\u5206\uff0c\u4ee5\u53ca\u5728ChatGPT\u7b49\u5be6\u969b\u61c9\u7528\u4e2d\u7684\u6210\u529f\uff0c\u90fd\u9ad4\u73fe\u4e86\u5b83\u5011\u7684\u5f37\u5927\u80fd\u529b \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u8b49\u64da\u8868\u660e<\/strong>\uff0c\u7576\u4efb\u52d9\u9700\u8981\u5c0d\u4e00\u500b\u5b8c\u6574\u7684\u8f38\u5165\u5e8f\u5217\u9032\u884c<strong>\u7406\u89e3<\/strong>\uff0c\u4e26\u57fa\u65bc\u6b64<strong>\u751f\u6210<\/strong>\u4e00\u500b\u5168\u65b0\u7684\u3001\u76f8\u95dc\u7684\u8f38\u51fa\u5e8f\u5217\u6642\uff0c\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u6a21\u578b\u5247\u8868\u73fe\u6700\u4f73\u3002\u5b83\u5011\u5728\u6458\u8981\uff08ROUGE\u5206\u6578\uff09\u548c\u7ffb\u8b6f\uff08BLEU\u5206\u6578\uff09\u7b49\u7d93\u5178Seq2Seq\u4efb\u52d9\u4e0a\u7684SOTA\u8868\u73fe\uff0c\u9a57\u8b49\u4e86\u5176\u67b6\u69cb\u8a2d\u8a08\u7684\u6709\u6548\u6027 \u3002 \u00a0<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u7b2c\u4e94\u7ae0\uff1a\u4e0b\u4e00\u500b\u738b\u671d &#8211; LLM\u67b6\u69cb\u7684\u672a\u4f86<\/h2>\n\n\n\n<p>\u4e09\u5927\u738b\u671d\u7684\u6f14\u7fa9\u8b5c\u5beb\u4e86\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u767c\u5c55\u7684\u7b2c\u4e00\u500b\u9ec3\u91d1\u6642\u4ee3\u3002\u7136\u800c\uff0c\u6b63\u5982\u6b77\u53f2\u7684\u8eca\u8f2a\u6efe\u6efe\u5411\u524d\uff0c\u6280\u8853\u7684\u6f14\u9032\u4e5f\u6c38\u4e0d\u505c\u6b47\u3002\u63a8\u52d5\u89e3\u78bc\u5668-Only\u738b\u671d\u8d70\u5411\u5dd4\u5cf0\u7684\u7e2e\u653e\u5b9a\u5f8b\uff0c\u672c\u8eab\u4e5f\u9810\u793a\u4e86\u5176\u5c40\u9650\u6027\u3002\u4e00\u500b\u65b0\u7684\u67b6\u69cb\u7bc4\u5f0f\u6b63\u5728\u5d1b\u8d77\uff0c\u5b83\u8a66\u5716\u89e3\u6c7a\u5bc6\u96c6\u6a21\u578b\u7684\u5167\u5728\u77db\u76fe\uff0c\u9810\u793a\u8457\u4e0b\u4e00\u500b\u738b\u671d\u7684\u5230\u4f86\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7e2e\u653e\u5b9a\u5f8b\u7684\u96d9\u5203\u528d\uff1a\u8a08\u7b97\u4e4b\u7246<\/h3>\n\n\n\n<p>\u7e2e\u653e\u5b9a\u5f8b\u63ed\u793a\u4e86\u4e00\u500b\u7c21\u55ae\u800c\u5f37\u5927\u7684\u771f\u7406\uff1a\u66f4\u5927\u7684\u6a21\u578b\u3001\u66f4\u591a\u7684\u6578\u64da\u3001\u66f4\u5f37\u7684\u8a08\u7b97\u529b\uff0c\u6703\u5e36\u4f86\u66f4\u597d\u3001\u66f4\u53ef\u9810\u6e2c\u7684\u6027\u80fd <sup><\/sup>\u3002\u9019\u4e00\u898f\u5f8b\u662fGPT-3\u7b49\u5bc6\u96c6\uff08dense\uff09\u6a21\u578b\u53d6\u5f97\u6210\u529f\u7684\u6838\u5fc3\u9a45\u52d5\u529b\u3002\u5728\u5bc6\u96c6\u6a21\u578b\u4e2d\uff0c\u6bcf\u4e00\u500b\u8f38\u5165\u8a5e\u5143\u90fd\u6703\u6fc0\u6d3b\u4e26\u5229\u7528\u6a21\u578b\u7684\u6240\u6709\u53c3\u6578\u4f86\u9032\u884c\u8a08\u7b97\u3002\u7136\u800c\uff0c\u9019\u7a2e\u300c\u591a\u591a\u76ca\u5584\u300d\u7684\u7b56\u7565\u662f\u4e00\u628a\u96d9\u5203\u528d\u3002 &nbsp;<\/p>\n\n\n\n<p>\u96a8\u8457\u6a21\u578b\u53c3\u6578\u5f9e\u6578\u5341\u5104\u64f4\u5c55\u5230\u6578\u5343\u5104\uff0c\u751a\u81f3\u842c\u5104\uff0c\u5176\u8a13\u7df4\u548c\u63a8\u7406\u7684\u6210\u672c\u4e5f\u5448\u6307\u6578\u7d1a\u589e\u9577\u3002\u8a13\u7df4\u4e00\u500b\u842c\u5104\u53c3\u6578\u7684\u5bc6\u96c6\u6a21\u578b\u6240\u9700\u7684\u8a08\u7b97\u8cc7\u6e90\u3001\u80fd\u6e90\u6d88\u8017\u548c\u6642\u9593\u6210\u672c\uff0c\u5c0d\u65bc\u7d55\u5927\u591a\u6578\u7d44\u7e54\u800c\u8a00\u90fd\u662f\u96e3\u4ee5\u627f\u53d7\u7684\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5373\u4f7f\u5728\u63a8\u7406\u968e\u6bb5\uff0c\u6bcf\u6b21\u8655\u7406\u4e00\u500b\u8a5e\u5143\u90fd\u8981\u52d5\u7528\u5168\u90e8\u7684\u842c\u5104\u53c3\u6578\uff0c\u9019\u4f7f\u5f97\u6a21\u578b\u7684\u90e8\u7f72\u548c\u670d\u52d9\u6210\u672c\u6975\u9ad8\uff0c\u6548\u7387\u4f4e\u4e0b\u3002\u6574\u500b\u9818\u57df\u6b63\u9010\u6f38\u649e\u4e0a\u4e00\u5835\u7531\u7269\u7406\u548c\u7d93\u6fdf\u898f\u5f8b\u7bc9\u6210\u7684\u300c\u8a08\u7b97\u4e4b\u7246\u300d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7a00\u758f\u6027\u7684\u5d1b\u8d77\uff1a\u5c08\u5bb6\u6df7\u5408\uff08Mixture of Experts, MoE\uff09<\/h3>\n\n\n\n<p>\u70ba\u4e86\u89e3\u6c7a\u9019\u4e00\u56f0\u5883\uff0c\u7814\u7a76\u754c\u5c07\u76ee\u5149\u6295\u5411\u4e86\u4e00\u7a2e\u66f4\u70ba\u9ad8\u6548\u7684\u67b6\u69cb\u7bc4\u5f0f\u2014\u2014<strong>\u5c08\u5bb6\u6df7\u5408\uff08Mixture of Experts, MoE\uff09<\/strong> <sup><\/sup>\u3002MoE\u4e26\u975e\u4e00\u500b\u5168\u65b0\u7684\u6982\u5ff5\uff0c\u4f46\u5728LLM\u6642\u4ee3\uff0c\u5b83\u88ab\u91cd\u65b0\u767c\u6398\u4e26\u8ce6\u4e88\u4e86\u65b0\u7684\u751f\u547d\uff0c\u6210\u70ba\u4e86\u6253\u7834\u7e2e\u653e\u74f6\u9838\u7684\u95dc\u9375\u3002 &nbsp;<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u6838\u5fc3\u6982\u5ff5<\/h4>\n\n\n\n<p>MoE\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5f9e\u300c\u6240\u6709\u53c3\u6578\u5171\u540c\u5de5\u4f5c\u300d\u8f49\u8b8a\u70ba\u300c\u5c08\u696d\u5206\u5de5\uff0c\u5354\u540c\u5de5\u4f5c\u300d\u3002\u5b83\u5c0d\u6a19\u6e96Transformer\u67b6\u69cb\u4e2d\u7684\u524d\u994b\u7db2\u7d61\uff08FFN\uff09\u5c64\u9032\u884c\u4e86\u6539\u9020 <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5c08\u5bb6\u7db2\u7d61\uff08Experts\uff09<\/strong>\uff1a\u5728\u4e00\u500bMoE\u5c64\u4e2d\uff0c\u55ae\u4e00\u7684\u3001\u5de8\u5927\u7684FFN\u88ab\u66ff\u63db\u70ba\u591a\u500b\uff08\u4f8b\u59828\u500b\u300164\u500b\uff09\u898f\u6a21\u8f03\u5c0f\u3001\u529f\u80fd\u7368\u7acb\u7684\u300c\u5c08\u5bb6\u300d\u7db2\u7d61\u3002\u6bcf\u500b\u5c08\u5bb6\u672c\u8eab\u5c31\u662f\u4e00\u500b\u6a19\u6e96\u7684\u524d\u994b\u7db2\u7d61 \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u9580\u63a7\u7db2\u7d61\/\u8def\u7531\u5668\uff08Gating Network\/Router\uff09<\/strong>\uff1a\u5728MoE\u5c64\u7684\u524d\u7aef\uff0c\u6709\u4e00\u500b\u5c0f\u578b\u7684\u3001\u8f15\u91cf\u7d1a\u7684\u795e\u7d93\u7db2\u7d61\uff0c\u88ab\u7a31\u70ba\u300c\u8def\u7531\u5668\u300d\u3002\u8def\u7531\u5668\u7684\u4f5c\u7528\u662f\u6839\u64da\u7576\u524d\u7684\u8f38\u5165\u8a5e\u5143\uff0c\u52d5\u614b\u5730\u3001\u667a\u80fd\u5730\u6c7a\u5b9a\u61c9\u8a72\u5c07\u9019\u500b\u8a5e\u5143\u767c\u9001\u7d66\u54ea\u4e9b\u5c08\u5bb6\u9032\u884c\u8655\u7406 \u3002 \u00a0<\/li>\n\n\n\n<li><strong>\u7a00\u758f\u6fc0\u6d3b\uff08Sparse Activation\uff09<\/strong>\uff1a\u5c0d\u65bc\u6bcf\u4e00\u500b\u8f38\u5165\u7684\u8a5e\u5143\uff0c\u8def\u7531\u5668\u4e26\u4e0d\u6703\u6fc0\u6d3b\u6240\u6709\u7684\u5c08\u5bb6\uff0c\u800c\u662f\u53ea\u9078\u64c7\u4e00\u500b\u7a00\u758f\u7684\u5b50\u96c6\uff08\u4f8b\u5982\uff0c\u5f9e8\u500b\u5c08\u5bb6\u4e2d\u9078\u64c72\u500b\uff09\u3002\u7136\u5f8c\uff0c\u8a72\u8a5e\u5143\u7684\u6700\u7d42\u8f38\u51fa\u662f\u9019\u5e7e\u500b\u88ab\u9078\u4e2d\u5c08\u5bb6\u7684\u8f38\u51fa\u7d50\u679c\u7684\u52a0\u6b0a\u548c \u3002 \u00a0<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">MoE\u7684\u512a\u52e2<\/h4>\n\n\n\n<p>\u9019\u7a2e\u8a2d\u8a08\u5e36\u4f86\u4e86\u9769\u547d\u6027\u7684\u597d\u8655\uff1a\u5b83\u6210\u529f\u5730\u89e3\u8026\u4e86\u6a21\u578b\u7684<strong>\u7e3d\u53c3\u6578\u6578\u91cf<\/strong>\u548c<strong>\u55ae\u6b21\u63a8\u7406\u7684\u8a08\u7b97\u91cf<\/strong>\u3002\u4e00\u500bMoE\u6a21\u578b\u7684\u7e3d\u53c3\u6578\u53ef\u4ee5\u8f15\u6613\u5730\u64f4\u5c55\u5230\u842c\u5104\u7d1a\u5225\uff0c\u56e0\u70ba\u5b83\u53ea\u662f\u589e\u52a0\u4e86\u5c08\u5bb6\u7684\u6578\u91cf\u3002\u7136\u800c\uff0c\u5728\u8655\u7406\u4efb\u4f55\u4e00\u500b\u8a5e\u5143\u6642\uff0c\u5be6\u969b\u53c3\u8207\u8a08\u7b97\u7684\u53ea\u6709\u88ab\u8def\u7531\u5668\u9078\u4e2d\u7684\u90a3\u4e00\u5c0f\u90e8\u5206\u5c08\u5bb6\u3002\u9019\u610f\u5473\u8457\uff0c\u6a21\u578b\u7684\u898f\u6a21\u53ef\u4ee5\u6301\u7e8c\u64f4\u5927\uff0c\u800c\u5176\u8a13\u7df4\u548c\u63a8\u7406\u7684\u8a08\u7b97\u6210\u672c\uff08\u4ee5FLOPs\u8861\u91cf\uff09\u537b\u53ef\u4ee5\u4fdd\u6301\u5728\u4e00\u500b\u76f8\u5c0d\u8f03\u4f4e\u7684\u6c34\u5e73 <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u65b0\u738b\u8005\u7684\u96db\u5f62\uff1aMixtral 8x7B<\/h3>\n\n\n\n<p>\u5982\u679c\u8aaaMoE\u662f\u672a\u4f86\u738b\u671d\u7684\u7406\u8ad6\u57fa\u790e\uff0c\u90a3\u9ebc\u7531Mistral AI\u57282023\u5e74\u767c\u5e03\u7684<strong>Mixtral 8x7B<\/strong>\u5c31\u662f\u9019\u4e00\u7406\u8ad6\u7684\u6709\u529b\u5be6\u8e10\u8005\u548c\u65b0\u738b\u8005\u7684\u96db\u5f62 <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<p>Mixtral 8x7B\u662f\u4e00\u500b\u57fa\u65bc\u89e3\u78bc\u5668-Only\u67b6\u69cb\u7684\u7a00\u758f\u5c08\u5bb6\u6df7\u5408\u6a21\u578b\u3002\u5176\u540d\u7a31\u4e2d\u7684\u300c8x7B\u300d\u610f\u5473\u8457\uff0c\u5728\u5176Transformer\u7684\u6bcf\u500bMoE\u5c64\u4e2d\uff0c\u90fd\u5305\u542b8\u500b\u5c08\u5bb6\uff0c\u6bcf\u500b\u5c08\u5bb6\u7684\u53c3\u6578\u898f\u6a21\u7d04\u70ba70\u5104\u3002\u7136\u800c\uff0c\u5c0d\u65bc\u6bcf\u500b\u8a5e\u5143\uff0c\u5176\u8def\u7531\u5668\u53ea\u6703\u9078\u64c72\u500b\u5c08\u5bb6\u4f86\u9032\u884c\u8655\u7406 <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<p>\u9019\u4f7f\u5f97Mixtral\u7684\u7e3d\u53c3\u6578\u9054\u5230\u4e86\u7d04470\u5104\uff0c\u4f46\u5176\u55ae\u6b21\u63a8\u7406\u7684\u6709\u6548\u8a08\u7b97\u53c3\u6578\u50c5\u7d04\u70ba130\u5104 <sup><\/sup>\u3002\u5176\u7d50\u679c\u662f\uff0cMixtral\u7684\u63a8\u7406\u901f\u5ea6\u8207\u4e00\u500b130\u5104\u53c3\u6578\u7684\u5bc6\u96c6\u6a21\u578b\u76f8\u7576\uff0c\u4f46\u5176\u6027\u80fd\u537b\u5728\u591a\u500b\u57fa\u6e96\u6e2c\u8a66\u4e2d\u8d85\u8d8a\u4e86\u898f\u6a21\u66f4\u5927\u7684\u5bc6\u96c6\u6a21\u578b\uff0c\u5982Llama 2 70B\uff0c\u751a\u81f3\u5728\u67d0\u4e9b\u65b9\u9762\u53ef\u4ee5\u8207GPT-3.5\u76f8\u5ab2\u7f8e <sup><\/sup>\u3002Mixtral\u7684\u6210\u529f\uff0c\u7279\u5225\u662f\u5176\u958b\u6e90\u7684\u7279\u6027\uff0c\u6709\u529b\u5730\u8b49\u660e\u4e86MoE\u662f\u901a\u5f80\u66f4\u5f37\u5927\u3001\u66f4\u9ad8\u6548LLM\u7684\u53ef\u884c\u8def\u5f91\uff0c\u4e5f\u9810\u793a\u8457\u672a\u4f86\u67b6\u69cb\u7684\u4e3b\u6d41\u8da8\u52e2\u3002\u95dc\u65bcGPT-4\u4e5f\u63a1\u7528\u4e86MoE\u67b6\u69cb\u7684\u50b3\u8a00\uff0c\u96d6\u7136\u672a\u7d93\u5b98\u65b9\u8b49\u5be6\uff0c\u4f46\u9032\u4e00\u6b65\u5f37\u5316\u4e86\u9019\u4e00\u8da8\u52e2\u7684\u53ef\u4fe1\u5ea6 <sup><\/sup>\u3002 &nbsp;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7d50\u8ad6\uff1a\u4e0d\u673d\u7684\u907a\u7522\u8207\u878d\u5408\u7684\u672a\u4f86<\/h3>\n\n\n\n<p>\u56de\u9867\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u5f9e2017\u5e74\u81f3\u4eca\u7684\u6f14\u5316\u6b77\u7a0b\uff0c\u4e09\u5927\u738b\u671d\u5404\u81ea\u7559\u4e0b\u4e86\u4e0d\u673d\u7684\u907a\u7522\uff0c\u4e26\u5171\u540c\u6307\u5411\u4e00\u500b\u878d\u5408\u7684\u672a\u4f86\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7de8\u78bc\u5668-Only\u738b\u671d<\/strong>\u7559\u4e0b\u4e86<strong>\u6df1\u5ea6\u4e0a\u4e0b\u6587\u7406\u89e3<\/strong>\u7684\u907a\u7522\u3002\u5b83\u6240\u958b\u5275\u7684MLM\u9810\u8a13\u7df4\u76ee\u6a19\u548c\u300c\u9810\u8a13\u7df4-\u5fae\u8abf\u300d\u7bc4\u5f0f\uff0c\u81f3\u4eca\u4ecd\u5728NLU\u9818\u57df\u767c\u63ee\u8457\u91cd\u8981\u4f5c\u7528\u3002\u5b83\u6559\u6703\u4e86\u6211\u5011\uff0c\u4e00\u500b\u771f\u6b63\u7406\u89e3\u8a9e\u8a00\u7684\u6a21\u578b\u9700\u8981\u5177\u5099\u96d9\u5411\u5be9\u8996\u6587\u672c\u7684\u80fd\u529b\u3002<\/li>\n\n\n\n<li><strong>\u89e3\u78bc\u5668-Only\u738b\u671d<\/strong>\u7559\u4e0b\u4e86<strong>\u5f37\u5927\u751f\u6210\u80fd\u529b<\/strong>\u548c<strong>\u898f\u6a21\u6548\u61c9<\/strong>\u7684\u907a\u7522\u3002\u5b83\u901a\u904e\u4e0a\u4e0b\u6587\u5b78\u7fd2\u5fb9\u5e95\u6539\u8b8a\u4e86\u4eba\u6a5f\u4ea4\u4e92\u7684\u65b9\u5f0f\uff0c\u4e26\u63ed\u793a\u4e86\u7576\u6a21\u578b\u898f\u6a21\u9054\u5230\u4e00\u5b9a\u7a0b\u5ea6\u5f8c\uff0c\u6703\u6e67\u73fe\u51fa\u4ee4\u4eba\u9a5a\u8a1d\u7684\u667a\u80fd\u3002\u5b83\u5c07LLM\u5f9e\u5be6\u9a57\u5ba4\u63a8\u5411\u4e86\u6578\u5341\u5104\u666e\u901a\u7528\u6236\u3002<\/li>\n\n\n\n<li><strong>\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u738b\u671d<\/strong>\u7559\u4e0b\u4e86<strong>\u9748\u6d3b\u6027<\/strong>\u548c<strong>\u5e8f\u5217\u8f49\u63db<\/strong>\u7684\u907a\u7522\u3002\u5b83\u5c55\u793a\u4e86\u5982\u4f55\u901a\u904e\u4e00\u500b\u7d71\u4e00\u7684\u6846\u67b6\u4f86\u89e3\u6c7a\u591a\u6a23\u5316\u7684NLP\u554f\u984c\uff0c\u4e26\u5728\u9700\u8981\u7cbe\u78ba\u6620\u5c04\u8f38\u5165\u8207\u8f38\u51fa\u7684\u4efb\u52d9\u4e0a\u5b9a\u7fa9\u4e86SOTA\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5c55\u671b\u672a\u4f86\uff0c\u6a21\u578b\u7684\u67b6\u69cb\u8a2d\u8a08\u5c07\u4e0d\u518d\u662f\u6559\u689d\u5f0f\u5730\u9078\u64c7\u4e09\u8005\u4e4b\u4e00\u3002\u76f8\u53cd\uff0c\u6211\u5011\u5c07\u770b\u5230\u66f4\u6df1\u5c64\u6b21\u7684\u878d\u5408\u3002\u672a\u4f86\u7684\u67b6\u69cb\u5f88\u53ef\u80fd\u6703\u5728\u89e3\u78bc\u5668-Only\u7684\u57fa\u790e\u4e0a\uff0c\u878d\u5165\u66f4\u591a\u7406\u89e3\u80fd\u529b\uff0c\u6216\u8005\u5728\u7de8\u78bc\u5668-\u89e3\u78bc\u5668\u7684\u6846\u67b6\u4e0b\uff0c\u63a2\u7d22\u66f4\u9ad8\u6548\u7684\u4ea4\u4e92\u65b9\u5f0f\u3002\u800c\u50cf**\u7a00\u758f\u6027\uff08MoE\uff09**\u9019\u6a23\u7684\u539f\u5247\uff0c\u5c07\u6210\u70ba\u69cb\u5efa\u9019\u4e00\u5207\u7684\u57fa\u790e\u5c64\u3002\u5f9e\u5bc6\u96c6\u5230\u7a00\u758f\u7684\u8f49\u8b8a\uff0c\u4e26\u975e\u5c0d\u904e\u53bb\u7684\u5426\u5b9a\uff0c\u800c\u662f\u5c0d\u7e2e\u653e\u5b9a\u5f8b\u6210\u529f\u6240\u5f15\u767c\u7684\u6311\u6230\u7684\u5fc5\u7136\u56de\u61c9\u3002\u9019\u662f\u4e00\u6b21\u5411\u8457\u66f4\u9ad8\u6548\u7387\u3001\u66f4\u5f37\u6a21\u584a\u5316\u548c\u66f4\u53ef\u6301\u7e8c\u64f4\u5c55\u65b9\u5411\u7684\u6f14\u9032\u3002\u4e09\u5927\u738b\u671d\u7684\u907a\u7522\u5c07\u88ab\u7e7c\u627f\uff0c\u4e26\u5728\u4e00\u500b\u66f4\u70ba\u9ad8\u6548\u548c\u5f37\u5927\u7684\u65b0\u67b6\u69cb\u7bc4\u5f0f\u4e2d\u5f97\u5230\u7d71\u4e00\uff0c\u7e7c\u7e8c\u63a8\u52d5\u8457\u901a\u7528\u4eba\u5de5\u667a\u80fd\u7684\u908a\u754c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5c0e\u8ad6\uff1a\u65b0\u7d00\u5143\u7684\u9ece\u660e\u8207\u5049\u5927\u7684\u5206\u6b67 Transformer\u4e4b\u524d\u7684\u4e16\u754c\uff1a\u8a18\u61b6\u8207\u4e26\u884c\u5316\u7684 &hellip; <a href=\"https:\/\/ouyangminwei.com\/index.php\/2025\/07\/28\/encoder-decoder-encoder-decoder-only-1\/\">\u95b1\u8b80\u5168\u6587 <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[1],"tags":[],"post_format":[],"class_list":["post-874","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_edit_lock":"1752830359:1","_edit_last":"1","_aioseo_title":"#post_title #separator_sa #site_title","_aioseo_description":"#post_excerpt","_aioseo_keywords":"","_aioseo_og_title":"","_aioseo_og_description":"","_aioseo_og_article_section":"","_aioseo_og_article_tags":"","_aioseo_twitter_title":"","_aioseo_twitter_description":"","_oembed_2544c1d0cb3503ab4c4d558c3b3c8873":"","_oembed_time_2544c1d0cb3503ab4c4d558c3b3c8873":"","_oembed_99481806ecbe6ce4ee46f8588d320993":"","_oembed_db663acf973e82e6d9d80df71945dfb8":"","_oembed_16cdfab488f57db73586f4286af2704f":"","_wp_old_slug":"","_links":{"self":[{"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/posts\/874","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/comments?post=874"}],"version-history":[{"count":1,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/posts\/874\/revisions"}],"predecessor-version":[{"id":875,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/posts\/874\/revisions\/875"}],"wp:attachment":[{"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/media?parent=874"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/categories?post=874"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/tags?post=874"},{"taxonomy":"post_format","embeddable":true,"href":"https:\/\/ouyangminwei.com\/index.php\/wp-json\/wp\/v2\/post_format?post=874"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}