{"id":7812,"date":"2024-07-05T15:01:01","date_gmt":"2024-07-05T07:01:01","guid":{"rendered":""},"modified":"2024-07-05T15:01:01","modified_gmt":"2024-07-05T07:01:01","slug":"bert\u9884\u8bad\u7ec3\u52a0lstm_pytorch\u9884\u8bad\u7ec3\u6a21\u578b","status":"publish","type":"post","link":"https:\/\/mushiming.com\/7812.html","title":{"rendered":"bert\u9884\u8bad\u7ec3\u52a0lstm_pytorch\u9884\u8bad\u7ec3\u6a21\u578b"},"content":{"rendered":"

\n <\/path> \n<\/svg> <\/p>\n

 \u6587\u7ae0\u8f6c\u8f7d\uff5c\u667a\u6e90\u793e\u533a \u672c\u671f\u8d21\u732e\u8005\uff5c\u7533\u5fb7\u5468 \u7fdf\u73c2 \u5434\u65b0\u521a <\/code><\/pre>\n
\u5173\u4e8e\u5468\u520a<\/strong><\/h3>\n
\u672c\u671f\u5468\u520a\uff0c\u6211\u4eec\u9009\u62e9\u4e8612\u7bc7\u9884\u8bad\u7ec3\u76f8\u5173\u7684\u8bba\u6587\uff0c\u6d89\u53ca\u53e5\u5b50\u8868\u793a\u3001\u53d8\u6362\u5668\u7ed3\u6784\u4f18\u5316\u3001\u6570\u636e\u589e\u5f3a\u3001\u7f51\u7edc\u7ed3\u6784\u4f18\u5316\u3001\u52a8\u6001\u795e\u7ecf\u7f51\u7edc\u3001\u6a21\u578b\u538b\u7f29\u3001\u56fe\u9884\u8bad\u7ec3\u6a21\u578b\u3001\u96f6\u6837\u672c\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u3001\u5fae\u8c03\u63a2\u7d22\u3001\u56e0\u679c\u63a8\u7406\u3001\u5e7b\u60f3\u751f\u6210\u3001\u5206\u5b50\u7ed3\u6784\u9605\u8bfb\u7406\u89e3\u7684\u63a2\u7d22\u3002\u6b64\u5916\uff0c\u5728\u7814\u7a76\u52a8\u6001\u65b9\u9762\uff0c\u6211\u4eec\u9009\u62e9\u4e861\u7bc7\u9884\u8bad\u7ec3\u8d44\u8baf\uff0c\u5c06\u4ecb\u7ecd\u591a\u4efb\u52a1\u63d0\u793a\u5b66\u4e60\u65b9\u9762\u7684\u4e00\u4e9b\u6700\u65b0\u5185\u5bb9\u3002<\/p>\n
\u8bba\u6587\u63a8\u8350<\/strong><\/h3>\n
\u6807\u9898\uff1a\u5317\u5927\uff5cSGPT: GPT Sentence Embeddings for Semantic Search\uff08SGPT\uff1a\u7528\u4e8e\u8bed\u4e49\u7684 GPT \u53e5\u5b50\u5d4c\u5165\u641c\u7d22\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aNiklas Muennighoff<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u4ecb\u7ecd\u4e86\u4e00\u79cd\u7528\u4e8e\u68c0\u7d22\u7684\u6587\u672c\u8868\u793a\u7b97\u6cd5\u3002GPT\u53d8\u6362\u5668\u662f\u53ef\u7528\u7684\u6700\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u4f46\u8bed\u4e49\u641c\u7d22\u662f\u4ee5 BERT \u53d8\u6362\u5668\u4e3a\u4e3b\u3002\u4f5c\u8005\u63d0\u51fa\u4e86SGPT-BE\u548cSGPT-CE\u7528\u4e8e\u5c06GPT\u6a21\u578b\u4f5c\u4e3a\u53cc\u7f16\u7801\u5668\u6216\u4ea4\u53c9\u7f16\u7801\u5668\u5e94\u7528\u4e8e\u5bf9\u79f0\u6216\u975e\u5bf9\u79f0\u641c\u7d22\u3002SGPT-BE\u901a\u8fc7\u5bf9\u6bd4\u4ea7\u751f\u8bed\u4e49\u4e0a\u6709\u610f\u4e49\u7684\u53e5\u5b50\u5d4c\u5165\u4ec5\u5bf9\u504f\u7f6e\u5f20\u91cf\u8fdb\u884c\u5fae\u8c03\u548c\u4e00\u79cd\u65b0\u9896\u7684\u6c60\u5316\u65b9\u6cd5\u3002\u4e00\u4e2a58\u4ebf\u7684\u53c2\u6570SGPT-BE\u6bd4\u6700\u4f73\u53ef\u7528\u53e5\u5b50\u5d4c\u5165\u7684\u6027\u80fd\u9ad8\u51fa6%BEIR\u4e0a\u6700\u65b0\u7684\u6700\u65b0\u6280\u672f\u3002\u5b83\u4f18\u4e8e\u540c\u65f6\u63d0\u51fa\u7684OpenAI-175B Davinci\u7aef\u70b9\u7684\u5d4c\u5165\uff0c\u53ef\u5fae\u8c03 250,000\u500d\u4ee5\u4e0a\u53c2\u6570\u3002SGPT-CE\u4f7f\u7528\u6765\u81eaGPT\u6a21\u578b\u7684\u5bf9\u6570\u6982\u7387\uff0c\u65e0\u9700\u4efb\u4f55\u5fae\u8c03\u300261\u4ebf\u53c2\u6570\u7684SGPT-CE\u5728BEIR\u4e0a\u8bbe\u7f6e\u4e86\u65e0\u76d1\u7763\u7684\u6700\u5148\u8fdb\u6280\u672f\uff0c\u57287\u4e2a\u6570\u636e\u96c6\u4e0a\u51fb\u8d25\u4e86\u6709\u76d1\u7763\u7684\u6700\u5148\u8fdb\u6280\u672f\uff0c\u4f46\u5728\u5176\u4ed6\u6570\u636e\u96c6\u4e0a\u663e\u8457\u843d\u540e\u3002\u4f5c\u8005\u5c55\u793a\u4e86\u5982\u4f55\u901a\u8fc7\u8c03\u6574\u63d0\u793a\u6765\u7f13\u89e3\u8fd9\u79cd\u60c5\u51b5\u3002SGPT-BE\u548cSGPT-CE\u6027\u80fd\u968f\u6a21\u578b\u5927\u5c0f\u800c\u53d8\u5316\uff0c\u7136\u800c\uff0c\u589e\u52a0\u4e86\u5e94\u8003\u8651\u5ef6\u8fdf\u3001\u5b58\u50a8\u548c\u8ba1\u7b97\u6210\u672c\u3002<\/p>\n
\u4ee3\u7801\u4e0b\u8f7d\uff1ahttps:\/\/github.com\/Muennighoff\/sgpt<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aSGPT: GPT Sentence Embeddings for Semantic Search<\/p>\n
\u6807\u9898\uff1a\u5eb7\u5948\u5c14\u3001\u8c37\u6b4c\uff5cTransformer Quality in Linear Time(\u7ebf\u6027\u65f6\u95f4\u7684\u53d8\u6362\u5668\u8d28\u91cf)<\/strong><\/p>\n
\u4f5c\u8005\uff1aWeizhe Hua, Zihang Dai, Hanxiao Liu, Quoc V. Le<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u6539\u8fdb\u7684\u53d8\u6362\u5668\u7ed3\u6784\u3002\u4f5c\u8005\u91cd\u65b0\u5ba1\u89c6\u4e86\u53d8\u6362\u5668\u4e2d\u7684\u8bbe\u8ba1\u9009\u62e9\uff0c\u5e76\u63d0\u51fa\u89e3\u51b3\u5904\u7406\u957f\u5e8f\u5217\u5f31\u70b9\u7684\u65b9\u6cd5\u3002\u9996\u5148\uff0c\u4f5c\u8005\u63d0\u51fa\u4e00\u4e2a\u7b80\u5355\u5c42\u547d\u540d\u4e3a\u95e8\u63a7\u6ce8\u610f\u529b\u5355\u5143\uff0c\u5176\u4e2d\u5141\u8bb8\u4f7f\u7528\u8f83\u5f31\u7684\u5355\u5934\u6ce8\u610f\u529b\uff0c\u540c\u65f6\u5c06\u8d28\u91cf\u635f\u5931\u964d\u81f3\u6700\u4f4e\u3002\u7136\u540e\u4f5c\u8005\u63d0\u51fa\u4e00\u79cd\u4e92\u8865\u7684\u7ebf\u6027\u8fd1\u4f3c\u65b9\u6cd5\u5230\u8fd9\u4e2a\u65b0\u5c42\uff0c\u5b83\u662f\u52a0\u901f\u5668\u53cb\u597d\u7684\u548c\u9ad8\u5ea6\u7684\u8d28\u91cf\u7ade\u4e89\u529b\u3002\u6a21\u578b\u547d\u540d\u4e3aFLASH\uff08\u5355\u5934\u5feb\u901f\u7ebf\u6027\u6ce8\u610f\u529b\uff09\uff0c\u4e24\u4e2a\u6539\u8fdb\u53d8\u6362\u5668\u77ed\u5e8f\u5217\uff08512\uff09\u548c\u957f\u5e8f\u5217\uff088K\uff09\u4e0a\u4e0b\u6587\u957f\u5ea6\u5b9e\u73b0\u5339\u914d\u7684\u56f0\u60d1\u5ea6\uff0c\u5b9e\u73b0\u5728Wiki-40B\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347\u9ad8\u8fbe4.9\u00d7\uff0c\u5728\u81ea\u56de\u5f52\u8bed\u8a00\u5efa\u6a21PG-19\u4e0a\u63d0\u534712.1x\uff0c\u548c\u5c4f\u853d\u8bed\u8a00\u5efa\u6a21C4\u4e0a\u76844.8\u00d7\u8bad\u7ec3\u901f\u5ea6\u63d0\u5347\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aTransformer Quality in Linear Time<\/p>\n
\u6807\u9898\uff1a\u9ea6\u8003\u745e\u5927\u5b66\u3001\u5fae\u8f6f\uff5cPromDA: Prompt-based Data Augmentation for Low-Resource NLU Task(PromDA\uff1a\u57fa\u4e8e\u63d0\u793a\u7684\u4f4e\u8d44\u6e90NLU\u6570\u636e\u589e\u5f3a)<\/strong><\/p>\n
\u4f5c\u8005\uff1aYufei Wang,Can Xu, Daxin Jiang\u7b49<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u63d0\u51fa\u4e00\u79cd\u4f4e\u8d44\u6e90\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u4efb\u52a1\u7684\u6570\u636e\u589e\u5f3a\u65b9\u6cd5\u3002\u4f5c\u8005\u63d0\u51fa\u4e86\u5728\u51bb\u7ed3\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b (PLM)\u4e0b\uff0c\u57fa\u4e8e\u63d0\u793a\u7684\u6570\u636e\u589e\u5f3a\u6a21\u578b\uff08PromDA\uff09\u53ea\u8bad\u7ec3\u5c0f\u89c4\u6a21\u7684\u8f6f\u63d0\u793a\uff08\u5373\u4e00\u7ec4\u53ef\u8bad\u7ec3\u7684\u5411\u91cf\uff09\u3002\u8fd9\u907f\u514d\u4eba\u5de5\u6536\u96c6\u672a\u6807\u8bb0\u7684\u57df\u5185\u6570\u636e\u5e76\u4fdd\u6301\u751f\u6210\u7684\u5408\u6210\u6570\u636e\u7684\u8d28\u91cf\u3002\u6b64\u5916\uff0cPromDA\u901a\u8fc7\u4e24\u4e2a\u4e0d\u540c\u7684\u751f\u6210\u5408\u6210\u6570\u636e\u4f7f\u7528 NLU \u6a21\u578b\u67e5\u770b\u5e76\u8fc7\u6ee4\u6389\u4f4e\u8d28\u91cf\u7684\u6570\u636e\u3002\u56db\u4e2a\u57fa\u51c6\u7684\u5b9e\u9a8c\u8868\u660e\uff0c\u7531PromDA\u6210\u529f\u63d0\u5347\u4e86NLU\u6a21\u578b\u7684\u6027\u80fd\uff0c\u8fd9\u4e9b\u6a21\u578b\u59cb\u7ec8\u4f18\u4e8e\u51e0\u4e2a\u7ade\u4e89\u57fa\u7ebf\u6a21\u578b\uff0c\u5305\u62ec\u6700\u5148\u8fdb\u7684\u534a\u76d1\u7763\u6a21\u578b\u4f7f\u7528\u672a\u6807\u8bb0\u7684\u57df\u5185\u6570\u636e\u3002\u6765\u81eaPromDA\u7684\u5408\u6210\u6570\u636e\u4e5f\u4e0e\u672a\u6807\u8bb0\u7684\u57df\u5185\u6570\u636e\u4e92\u8865\uff0cNLU\u6a21\u578b\u53ef\u4ee5\u8fdb\u4e00\u6b65\u6539\u8fdb\u4ed6\u4eec\u7ed3\u5408\u8d77\u6765\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aPromDA: Prompt-based Data Augmentation for Low-Resource NLU Task<\/p>\n
\u6807\u9898\uff1a\u82f1\u7279\u5c14\uff5cTRIMBERT: TAILORING BERT FOR TRADE-OFFS\uff08TRIMBERT\uff1a\u4e3a\u6743\u8861\u53d6\u820d\u91cf\u8eab\u5b9a\u5236 BERT\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aSharath Nittur Sridhar, Anthony Sarah, Sairam Sundaresan<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u4ecb\u7ecd\u4e86\u4e00\u79cd\u65b0\u7684BERT\u7f51\u7edc\u7ed3\u6784\u6539\u8fdb\u3002\u57fa\u4e8eBERT \u7684\u6a21\u578b\u5728\u89e3\u51b3\u5404\u79cd\u81ea\u7136\u8bed\u8a00\u95ee\u9898\u4e0a\u5df2\u7ecf\u5904\u7406\uff08NLP\uff09\u4efb\u52a1\u975e\u5e38\u6210\u529f\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u8bb8\u591a\u8fd9\u4e9b\u5927\u578b\u6a21\u578b\u9700\u8981\u5927\u91cf\u7684\u8ba1\u7b97\u8d44\u6e90\u548c\/\u6216\u65f6\u95f4\u6765\u8fdb\u884c\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\uff0c\u8fd9\u9650\u5236\u4e86\u66f4\u5e7f\u6cdb\u7684\u53ef\u91c7\u7528\u6027\u3002\u5c3d\u7ba1\u81ea\u6ce8\u610f\u529b\u5c42\u5df2\u7ecf\u8fc7\u5145\u5206\u7814\u7a76\uff0c\u5728\u6587\u732e\u4e2d\u4ecd\u7136\u7f3a\u5931\u8ddf\u968f\u81ea\u6ce8\u610f\u529b\u5c42\u7684\u4e2d\u95f4\u5c42\u7684\u7814\u7a76\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u4f5c\u8005\u8868\u660e\u51cf\u5c11BERT-BASE\u4e2d\u7684\u4e2d\u95f4\u5c42\u6570\u91cf\u5bfc\u81f4\u4e0b\u6e38\u4efb\u52a1\u7684\u5fae\u8c03\u7cbe\u5ea6\u635f\u5931\u6700\u5c0f\uff0c\u540c\u65f6\u663e\u8457\u51cf\u5c11\u6a21\u578b\u5927\u5c0f\u548c\u8bad\u7ec3\u65f6\u95f4\u3002\u4f5c\u8005\u8fdb\u4e00\u6b65\u51cf\u8f7b\u4e86\u4e24\u4e2a\u5173\u952e\u74f6\u9888\uff0c\u901a\u8fc7\u7528\u8ba1\u7b97\u66f4\u7b80\u5355\u7684\u66ff\u4ee3\u65b9\u6cd5\u66ff\u6362\u81ea\u6ce8\u610f\u5c42\u4e2d\u7684\u6240\u6709softmax\u64cd\u4f5c\u5e76\u5220\u9664\u6240\u6709layernorm\u64cd\u4f5c\u7684\u4e00\u534a\u3002\u8fd9\u8fdb\u4e00\u6b65\u51cf\u5c11\u8bad\u7ec3\u65f6\u95f4\uff0c\u540c\u65f6\u4fdd\u6301\u9ad8\u6c34\u5e73\u7684\u5fae\u8c03\u7cbe\u5ea6\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aTRIMBERT: TAILORING BERT FOR TRADE-OFFS<\/p>\n
\u6807\u9898\uff1a\u52a0\u5229\u798f\u5c3c\u4e9a\u5927\u5b66 | A Survey on Dynamic Neural Networks for Natural Language Processing\uff08NLP\u52a8\u6001\u795e\u7ecf\u7f51\u7edc\u7efc\u8ff0\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aCanwen Xu, Julian McAuley<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u662f\u52a8\u6001\u795e\u7ecf\u7f51\u7edc\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\u7684\u7efc\u8ff0\u3002\u6709\u6548\u5730\u7f29\u653e\u5927\u578bTransformer \u6a21\u578b\u662f\u81ea\u7136\u8bed\u8a00\u5904\u7406\u6700\u65b0\u8fdb\u5c55\u7684\u4e3b\u8981\u9a71\u52a8\u529b\u3002\u52a8\u6001\u795e\u7ecf\u7f51\u7edc\u662f\u4e00\u4e2a\u65b0\u5174\u7684\u7814\u7a76\u65b9\u5411\uff0c\u5b83\u80fd\u591f\u6839\u636e\u8f93\u5165\u52a8\u6001\u8c03\u6574\u795e\u7ecf\u7f51\u7edc\u7684\u8ba1\u7b97\u8def\u5f84\uff0c\u4ece\u800c\u5728\u8ba1\u7b97\u91cf\u548c\u65f6\u95f4\u4e0a\u5b9e\u73b0\u4e9a\u7ebf\u6027\u589e\u957f\u3002\u52a8\u6001\u795e\u7ecf\u7f51\u7edc\u53ef\u80fd\u662f\u4e00\u4e2a\u5f88\u6709\u524d\u666f\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u53ef\u4ee5\u89e3\u51b3\u4e0d\u65ad\u589e\u957f\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u7684\u53c2\u6570\u6570\u91cf\uff0c\u65e2\u53ef\u4ee5\u4f7f\u7528\u6570\u4e07\u4ebf\u4e2a\u53c2\u6570\u8fdb\u884c\u6a21\u578b\u9884\u8bad\u7ec3\uff0c\u53c8\u53ef\u4ee5\u5728\u79fb\u52a8\u8bbe\u5907\u4e0a\u8fdb\u884c\u66f4\u5feb\u7684\u63a8\u7406\u3002\u5728\u8fd9\u7bc7\u7efc\u8ff0\u4e2d\uff0c\u4f5c\u8005\u603b\u7ed3\u4e86\u4e09\u79cd\u52a8\u6001\u795e\u7ecf\u7f51\u7edc\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\u7684\u8fdb\u5c55\uff1a\u7565\u8bfb\uff08skimming\uff09\u3001\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08mixture of experts\uff09\u548c\u65e9\u671f\u9000\u51fa\u63a8\u7406\uff08early exit\uff09\u3002\u4f5c\u8005\u8fd8\u5f3a\u8c03\u4e86\u52a8\u6001\u795e\u7ecf\u7f51\u7edc\u76ee\u524d\u9762\u4e34\u7684\u6311\u6218\u548c\u672a\u6765\u7814\u7a76\u7684\u65b9\u5411\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aA Survey on Dynamic Neural Networks for Natural Language Processing<\/p>\n
\u6807\u9898\uff1a\u52a0\u5229\u798f\u5c3c\u4e9a\u5927\u5b66 | A Survey on Model Compression for Natural Language Processing\uff08NLP\u6a21\u578b\u538b\u7f29\u7efc\u8ff0\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aCanwen Xu, Julian McAuley<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u7efc\u8ff0\u7740\u91cd\u5728NLP\u6a21\u578b\u538b\u7f29\u7684\u63a8\u7406\u9636\u6bb5\u3002\u968f\u7740Transformer \u548c\u9884\u8bad\u7ec3\u6280\u672f\u7b49\u65b0\u4f53\u7cfb\u7ed3\u6784\u7684\u53d1\u5c55\uff0c\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u7684\u5e94\u7528\u53d6\u5f97\u4e86\u91cd\u5927\u8fdb\u5c55\u3002\u7136\u800c\uff0cTransformer\u7684\u9ad8\u80fd\u8017\u548c\u957f\u63a8\u7406\u5ef6\u8fdf\u963b\u788d\u4e86NLP\u8fdb\u5165\u66f4\u5e7f\u6cdb\u7684\u573a\u666f\uff0c\u5305\u62ec\u8fb9\u7f18\u548c\u79fb\u52a8\u8ba1\u7b97\u3002\u6709\u6548\u7684NLP\u7814\u7a76\u7684\u76ee\u7684\u662f\u5168\u9762\u8003\u8651\u8ba1\u7b97\uff0c\u65f6\u95f4\u548c\u78b3\u6392\u653e\u7684\u6574\u4e2a\u751f\u547d\u5468\u671f\u7684NLP\uff0c\u5305\u62ec\u6570\u636e\u51c6\u5907\uff0c\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u3002\u5728\u672c\u6b21\u7efc\u8ff0\u4e2d\uff0c\u4f5c\u8005\u5c06\u91cd\u70b9\u653e\u5728\u63a8\u7406\u9636\u6bb5\uff0c\u5e76\u56de\u987eNLP\u6a21\u578b\u538b\u7f29\u7684\u73b0\u72b6\uff0c\u5305\u62ec\u57fa\u51c6\u3001\u6307\u6807\u548c\u65b9\u6cd5\uff0c\u6700\u540e\u4f5c\u8005\u8fd8\u6982\u8ff0\u4e86\u76ee\u524d\u7684\u969c\u788d\u548c\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aA Survey on Model Compression for Natural Language Processing<\/p>\n
\u6807\u9898\uff1a\u897f\u6e56\u5927\u5b66\u3001\u4e2d\u79d1\u9662\u7b49 | A Survey of Pretraining on Graphs: Taxonomy, Methods, and Applications\uff08\u56fe\u9884\u8bad\u7ec3\u7efc\u8ff0\uff1a\u5206\u7c7b\u3001\u65b9\u6cd5\u548c\u5e94\u7528\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aJun Xia, Yanqiao Zhu, Yuanqi Du\u7b49<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u662f\u7b2c\u4e00\u4e2a\u56fe\u9884\u8bad\u7ec3\u7684\u7efc\u8ff0\u3002\u50cfBERT\u8fd9\u6837\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff08PLM\uff09\u5df2\u7ecf\u5f7b\u5e95\u6539\u53d8\u4e86\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u7684\u683c\u5c40\u3002\u53d7\u5176\u6269\u6563\u7684\u542f\u53d1\uff0c\u4eba\u4eec\u5bf9\u9884\u8bad\u7ec3\u56fe\u6a21\u578b\uff08PGM\uff09\u8fdb\u884c\u4e86\u5927\u91cf\u7684\u7814\u7a76\u3002\u7531\u4e8ePGMs\u5f3a\u5927\u7684\u6a21\u578b\u67b6\u6784\uff0c\u53ef\u4ee5\u4ece\u5927\u91cf\u6807\u8bb0\u548c\u672a\u6807\u8bb0\u7684\u56fe\u5f62\u6570\u636e\u4e2d\u83b7\u53d6\u4e30\u5bcc\u7684\u77e5\u8bc6\u3002\u9690\u5f0f\u7f16\u7801\u5728\u6a21\u578b\u53c2\u6570\u4e2d\u7684\u77e5\u8bc6\u53ef\u4ee5\u4f7f\u5404\u79cd\u4e0b\u6e38\u4efb\u52a1\u53d7\u76ca\uff0c\u5e76\u6709\u52a9\u4e8e\u7f13\u89e3\u56fe\u5b66\u4e60\u7684\u51e0\u4e2a\u57fa\u672c\u95ee\u9898\u3002\u5728\u8fd9\u7bc7\u8bba\u6587\u4e2d\uff0c\u4f5c\u8005\u63d0\u4f9b\u4e86\u7b2c\u4e00\u4e2a\u5168\u9762PGMs\u7684\u7efc\u8ff0\u3002\u4f5c\u8005\u9996\u5148\u4ecb\u7ecd\u4e86\u56fe\u8868\u793a\u5b66\u4e60\u7684\u5c40\u9650\u6027\uff0c\u7136\u540e\u4ecb\u7ecd\u4e86\u56fe\u9884\u8bad\u7ec3\u7684\u52a8\u673a\u3002\u7136\u540e\uff0c\u6839\u636e\u5206\u7c7b\u6cd5\u4ece\u56db\u4e2a\u4e0d\u540c\u7684\u89d2\u5ea6\u5bf9\u73b0\u6709PGM\u8fdb\u884c\u4e86\u7cfb\u7edf\u5206\u7c7b\u3002\u63a5\u4e0b\u6765\uff0c\u4f5c\u8005\u4ecb\u7ecdPGMs\u5728\u793e\u4ea4\u63a8\u8350\u548c\u836f\u7269\u53d1\u73b0\u65b9\u9762\u7684\u5e94\u7528\u3002\u6700\u540e\uff0c\u4f5c\u8005\u6982\u8ff0\u4e86\u591a\u4e2a\u6709\u524d\u9014\u7684\u7814\u7a76\u65b9\u5411\u3001\u4f5c\u4e3a\u672a\u6765\u7814\u7a76\u7684\u6307\u5bfc\u53c2\u8003\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aA Survey of Pretraining on Graphs: Taxonomy, Methods, and Applications<\/p>\n
\u6807\u9898\uff1a\u4f0a\u5229\u8bfa\u4f0a\u5927\u5b66\u9999\u69df\u5206\u6821|Generating Training Data with Language Models:Towards Zero-Shot Language Understanding\uff08\u7528\u8bed\u8a00\u6a21\u578b\u751f\u6210\u8bad\u7ec3\u6570\u636e\uff1a\u8d70\u5411\u96f6\u6837\u672c\u8bed\u8a00\u7406\u89e3\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aYu Meng\uff0c Jiaxin Huang \uff0cYu Zhang\u7b49<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u7efc\u5408\u8fd0\u7528\u4e24\u7c7b\u4e0d\u540c\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u96f6\u6837\u672c\u5b66\u4e60\u7684\u63a2\u7d22\u3002\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b (PLM) \u5728\u5404\u79cd\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u5353\u8d8a\u7684\u6027\u80fd\uff1a\u5355\u5411 PLM\uff08\u4f8b\u5982 GPT\uff09\u4ee5\u5176\u5353\u8d8a\u7684\u6587\u672c\u751f\u6210\u80fd\u529b\u800c\u95fb\u540d\uff1b\u53cc\u5411 PLM\uff08\u4f8b\u5982BERT\uff09\u4e00\u76f4\u662f\u81ea\u7136\u8bed\u8a00\u7406\u89e3\uff08NLU\uff09\u4efb\u52a1\u7684\u4e3b\u8981\u9009\u62e9\u3002\u867d\u7136\u8fd9\u4e24\u79cd\u7c7b\u578b\u7684\u6a21\u578b\u90fd\u53d6\u5f97\u4e86\u6709\u5e0c\u671b\u7684\u5c0f\u6837\u672c\u5b66\u4e60\u6027\u80fd\uff0c\u4f46\u5b83\u4eec\u5728\u96f6\u6837\u672c\u5b66\u4e60\u65b9\u9762\u7684\u6f5c\u529b\u5c1a\u672a\u5f97\u5230\u5145\u5206\u5f00\u53d1\u3002\u5728\u672c\u6587\u4e2d\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u7684\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u4f7f\u7528\u4e24\u79cd\u7c7b\u578b\u7684 PLM \u5bf9 NLU \u4efb\u52a1\u8fdb\u884c\u5b8c\u5168\u96f6\u6837\u672c\u5b66\u4e60\uff0c\u800c\u4e0d\u9700\u8981\u4efb\u4f55\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u6570\u636e\uff1a\u5355\u5411 PLM \u751f\u6210\u7531\u63d0\u793a\u5f15\u5bfc\u7684\u7c7b\u6761\u4ef6\u6587\u672c\uff0c\u7528\u4f5c\u8bad\u7ec3\u7528\u4e8e\u5fae\u8c03\u53cc\u5411 PLM \u7684\u6570\u636e\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aGenerating Training Data with Language Models:Towards Zero-Shot Language Understanding<\/p>\n
\u6807\u9898\uff1a\u65af\u5766\u798f | Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution\uff08\u5fae\u8c03\u4f1a\u626d\u66f2\u9884\u8bad\u7ec3\u7279\u5f81\uff0c\u5e76\u5728\u5206\u5e03\u5916\u8868\u73b0\u4e0d\u4f73\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aAnanya Kumar, Percy Liang \u7b49<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u4e3aICLR2022 oral\u8bba\u6587\uff0c\u4ecb\u7ecd\u4e86\u4e00\u79cd\u65b0\u7684\u5fae\u8c03\u7b56\u7565\u3002\u5f53\u628a\u9884\u8bad\u7ec3\u7684\u6a21\u578b\u8fc1\u79fb\u5230\u4e0b\u6e38\u4efb\u52a1\u65f6\uff0c\u4e24\u79cd\u6d41\u884c\u7684\u65b9\u6cd5\u662f\u5b8c\u5168\u5fae\u8c03\uff08\u66f4\u65b0\u6240\u6709\u7684\u6a21\u578b\u53c2\u6570\uff09\u548c\u7ebf\u6027\u63a2\u6d4b\uff08\u53ea\u66f4\u65b0\u6700\u540e\u7684\u7ebf\u6027\u5c42\uff09\u3002\u4f17\u6240\u5468\u77e5\uff0c\u5fae\u8c03\u4f1a\u83b7\u5f97\u66f4\u597d\u7684\u5206\u5e03\u5185\uff08ID\uff09\u51c6\u786e\u7387\uff0c\u7136\u800c\uff0c\u672c\u6587\u53d1\u73b0\uff0c\u5f53\u9884\u8bad\u7ec3\u7684\u7279\u5f81\u5f88\u597d\u4e14\u5206\u5e03\u504f\u79fb\u8f83\u5927\u65f6\uff0c\u5373\u5206\u5e03\u5916\uff08OOD\uff09\u65f6\uff0c\u5fae\u8c03\u53ef\u4ee5\u8fbe\u5230\u6bd4\u7ebf\u6027\u63a2\u6d4b\u66f4\u5dee\u7684\u7cbe\u5ea6\u3002\u572810\u4e2a\u5206\u5e03\u504f\u79fb\u6570\u636e\u96c6\u4e0a\uff0c\u5fae\u8c03\u5e73\u5747\u83b7\u5f97\u6bd4\u7ebf\u6027\u63a2\u6d4b\u9ad82%\u7684ID\u51c6\u786e\u7387\uff0c\u4f46OOD\u51c6\u786e\u7387\u4f4e7%\u3002\u5f53\u672c\u6587\u7528\u56fa\u5b9a\u6216\u968f\u673a\u7684\u5934\u521d\u59cb\u5316\u65f6\uff0c\u5fae\u8c03\u7684OOD\u8bef\u5dee\u5f88\u9ad8\uff0c\u8fd9\u662f\u56e0\u4e3a\u5728\u5fae\u8c03\u5b66\u4e60\u5934\u7684\u540c\u65f6\uff0c\u795e\u7ecf\u7f51\u7edc\u7684\u4e0b\u5c42\u540c\u65f6\u53d1\u751f\u53d8\u5316\uff0c\u5e76\u626d\u66f2\u4e86\u9884\u8bad\u7ec3\u7684\u7279\u5f81\u3002\u672c\u6587\u7684\u5206\u6790\u8868\u660e\uff0c\u7b80\u5355\u7684\u4e24\u6b65\u7b56\u7565\uff0c\u5148\u7ebf\u6027\u63a2\u6d4b\u7136\u540e\u5b8c\u5168\u5fae\u8c03\uff08LP-FT\uff09\u53ef\u88ab\u7528\u4f5c\u5fae\u8c03\u7684\u542f\u53d1\u5f0f\u65b9\u6cd5\uff0c\u5b83\u7ed3\u5408\u4e86\u5fae\u8c03\u548c\u7ebf\u6027\u63a2\u6d4b\u7684\u4f18\u70b9\u3002\u4ece\u7ecf\u9a8c\u4e0a\u770bLP-FT\u5728\u8bc4\u4ef7\u6570\u636e\u96c6\u4e0a\u7684\u8868\u73b0\u4f18\u4e8e\u5fae\u8c03\u548c\u7ebf\u6027\u63a2\u6d4b\uff0cID\u4e0a\u6bd4\u5168\u5fae\u8c03\u597d1%\uff0cOOD\u4e0a\u6bd4\u5168\u5fae\u8c03\u597d10%\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aFine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution<\/p>\n
\u6807\u9898\uff1a\u7f8e\u56fdSIFT\u516c\u53f8 | From Unstructured Text to Causal Knowledge Graphs: A Transformer-Based Approach\uff08\u4ece\u975e\u7ed3\u6784\u5316\u6587\u672c\u5230\u56e0\u679c\u77e5\u8bc6\u56fe\u8c31\uff1a\u57fa\u4e8eTransformer\u7684\u65b9\u6cd5\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aScott Friedman, Sonja Schmer-Galunder \u7b49<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u5229\u7528\u79d1\u5b66\u6587\u672c\u9884\u8bad\u7ec3\u8fdb\u884c\u4e86\u56e0\u679c\u77e5\u8bc6\u56fe\u8c31\u7684\u6784\u5efa\u3002\u672c\u6587\u5b9a\u6027\u56e0\u679c\u5173\u7cfb\u8868\u8fbe\u4e86\u4e16\u754c\u4e0a\u79bb\u6563\u6216\u8fde\u7eed\u7684\u76f8\u4e92\u4f5c\u7528\u3001\u4f9d\u8d56\u6027\u3001\u65f6\u95f4\u548c\u5355\u8c03\u6027\u7684\u7ea6\u675f\u3002\u63d0\u53d6\u548c\u8868\u793a\u8fd9\u4e9b\u4e0d\u540c\u7684\u56e0\u679c\u5173\u7cfb\u5bf9\u4e8e\u5728\u4ece\u79d1\u5b66\u53d1\u73b0\u5230\u793e\u4f1a\u79d1\u5b66\u7b49\u9886\u57df\u8fd0\u4f5c\u7684\u8ba4\u77e5\u7cfb\u7edf\u6765\u8bf4\u81f3\u5173\u91cd\u8981\u3002\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u4e2a\u57fa\u4e8eTransformer\u7684NLP\u67b6\u6784\uff0c\u5b83\u7ed3\u5408SciBERT\u7b49\u9884\u8bad\u7ec3\u6a21\u578b\u63d0\u53d6\u77e5\u8bc6\u56fe\u8c31\u4fe1\u606f\uff0c\u83b7\u5f97\u5305\u62ec\u7528\u8bed\u8a00\u63cf\u8ff0\u7684\u53d8\u91cf\u6216\u56e0\u7d20\u3001\u8fd9\u4e9b\u53d8\u91cf\u4e0a\u7684\u5b9a\u6027\u56e0\u679c\u5173\u7cfb\u3001\u9650\u5236\u8fd9\u4e9b\u56e0\u679c\u5173\u7cfb\u7684\u9650\u5b9a\u8bcd\u548c\u91cf\u7ea7\u4ee5\u53ca\u5728\u5927\u578b\u672c\u4f53\u4e2d\u5b9a\u4f4d\u6bcf\u4e2a\u63d0\u53d6\u8282\u70b9\u7684\u8bcd\u4e49\u3002\u8fd9\u4e2a\u7684\u67b6\u6784\u5e76\u975e\u662f\u4e00\u4e2a\u8ba4\u77e5\u7cfb\u7edf\uff0c\u7136\u800c\u5b83\u53ef\u4ee5\u5728\u73b0\u5b9e\u4e16\u754c\u7684\u9886\u57df\u4e2d\u51c6\u786e\u5730\u63d0\u53d6\u77e5\u8bc6\u56fe\u8c31\uff0c\u5e76\u4e14\u5176\u4ea7\u751f\u7684\u77e5\u8bc6\u56fe\u8c31\u5bf9\u4e8e\u8fdb\u884c\u57fa\u4e8e\u56fe\u8c31\u7684\u63a8\u7406\u7684\u8ba4\u77e5\u7cfb\u7edf\u5177\u5907\u5b9e\u7528\u6027\u3002\u672c\u6587\u540c\u65f6\u4e5f\u5c55\u793a\u4e86\u8fd9\u79cd\u65b9\u6cd5\u5904\u7406\u6765\u81ea\u5b66\u672f\u51fa\u7248\u7269\u3001\u65b0\u95fb\u6587\u7ae0\u548c\u793e\u4f1a\u5a92\u4f53\u7684\u6587\u672c\u8f93\u5165\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aFrom Unstructured Text to Causal Knowledge Graphs: A Transformer-Based Approach<\/p>\n
\u6807\u9898\uff1a\u6e2f\u79d1\u5927 | Survey of Hallucination in Natural Language Generation\uff08\u81ea\u7136\u8bed\u8a00\u751f\u6210\u4e2d\u7684\u5e7b\u89c9\u7efc\u8ff0\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aZiwei Ji, Pascale Fung \u7b49<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u5bf9\u9884\u8bad\u7ec3\u751f\u6210\u6a21\u578b\u4e2d\u5e7b\u89c9\u4fe1\u606f\u8fdb\u884c\u4e86\u7efc\u8ff0\u3002\u7684\u8fd1\u5e74\u6765\uff0c\u7531\u4e8e\u57fa\u4e8eTransformer\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff0c\u81ea\u7136\u8bed\u8a00\u751f\u6210\u5f97\u5230\u4e86\u6307\u6570\u7ea7\u7684\u6539\u5584\u3002\u8fd9\u79cd\u8fdb\u6b65\u4f7f\u5f97\u81ea\u7136\u8bed\u8a00\u751f\u6210\u66f4\u52a0\u6d41\u7545\u548c\u8fde\u8d2f\uff0c\u81ea\u7136\u800c\u7136\u5730\u5e26\u52a8\u4e86\u4e0b\u6e38\u4efb\u52a1\u7684\u53d1\u5c55\u3002\u7136\u800c\u8fd9\u79cd\u751f\u6210\u5305\u62ec\u5e7b\u89c9\u6587\u672c\uff0c\u5e7b\u89c9\u662f\u57fa\u4e8e\u795e\u7ecf\u7684\u81ea\u7136\u8bed\u8a00\u751f\u6210\u7684\u4e00\u4e2a\u4f2a\u547d\u9898\uff0c\u7531\u4e8e\u5b83\u4eec\u770b\u8d77\u6765\u5f88\u6d41\u7545\uff0c\u56e0\u6b64\u4f1a\u5bf9\u7528\u6237\u4ea7\u751f\u8bef\u5bfc\u3002\u672c\u6587\u5bf9NLG\u7684\u5e7b\u89c9\u95ee\u9898\u7684\u7814\u7a76\u8fdb\u5c55\u548c\u6311\u6218\u505a\u4e86\u4e00\u4e2a\u5e7f\u6cdb\u7684\u6982\u8ff0\uff0c\u5206\u6790\u4e86\u9020\u6210\u5e7b\u89c9\u7684\u5404\u79cd\u56e0\u7d20\uff0c\u5305\u62ec\u5608\u6742\u7684\u6570\u636e\u3001\u9519\u8bef\u7684\u53c2\u6570\u5316\u77e5\u8bc6\u3001\u4e0d\u6b63\u786e\u7684\u6ce8\u610f\u529b\u673a\u5236\u3001\u4e0d\u6070\u5f53\u7684\u8bad\u7ec3\u7b56\u7565\u548c\u63a8\u7406\u66b4\u9732\u504f\u5dee\u7b49\u3002\u4f5c\u8005\u8868\u660e\u5b58\u5728\u4e24\u7c7b\u5e7b\u89c9\uff0c\u5373\u5185\u5728\u7684\u5e7b\u89c9\u548c\u5916\u5728\u7684\u5e7b\u89c9\uff0c\u5b83\u4eec\u9700\u8981\u7528\u4e0d\u540c\u7684\u7f13\u89e3\u7b56\u7565\u6765\u5bf9\u5f85\u3002\u672c\u6587\u4e5f\u5728\u76f8\u5e94\u4e0b\u6e38\u4efb\u52a1\u4e2d\uff0c\u5305\u62ec\u6458\u8981\u603b\u7ed3\u3001\u5bf9\u8bdd\u751f\u6210\u3001\u751f\u6210\u6027\u95ee\u9898\u56de\u7b54\u3001\u6570\u636e\u5230\u6587\u672c\u751f\u6210\u548c\u673a\u5668\u7ffb\u8bd1\u7b49\u8fdb\u884c\u4e86\u9488\u5bf9\u5e7b\u89c9\u7684\u5177\u4f53\u5b9e\u4f8b\u5206\u6790\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aSurvey of Hallucination in Natural Language Generation<\/p>\n
\u6807\u9898\uff1a\u6e05\u534e | A deep-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals\uff08\u7406\u89e3\u529b\u53ef\u4e0e\u4eba\u7c7b\u4e13\u4e1a\u4eba\u58eb\u76f8\u5ab2\u7f8e\u7684\u8fde\u63a5\u5206\u5b50\u7ed3\u6784\u548c\u751f\u7269\u533b\u5b66\u6587\u672c\u7684\u6df1\u5ea6\u5b66\u4e60\u7cfb\u7edf\uff09<\/strong><\/p>\n
\u4f5c\u8005\uff1aZhiyuan Liu & Maosong Sun \u7b49<\/p>\n
\u7b80\u4ecb\uff1a\u672c\u6587\u4ecb\u7ecd\u4e86\u4e00\u4e2a\u591a\u6a21\u6001\u5206\u5b50\u8868\u5f81\u5b66\u4e60\u7cfb\u7edf\u3002\u4e3a\u4e86\u52a0\u901f\u751f\u7269\u533b\u5b66\u7814\u7a76\u8fdb\u7a0b\uff0c\u7814\u7a76\u8005\u5f00\u59cb\u901a\u8fc7\u9605\u8bfb\u5927\u89c4\u6a21\u7684\u751f\u7269\u533b\u5b66\u6570\u636e\u6765\u81ea\u52a8\u83b7\u53d6\u5206\u5b50\u5b9e\u4f53\u7684\u77e5\u8bc6\u3002\u53d7\u4eba\u7c7b\u4ece\u5206\u5b50\u7ed3\u6784\u548c\u751f\u7269\u533b\u5b66\u6587\u672c\u4fe1\u606f\u7684\u591a\u529f\u80fd\u9605\u8bfb\u4e2d\u5b66\u4e60\u6df1\u5ea6\u5206\u5b50\u77e5\u8bc6\u7684\u542f\u53d1\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u4e2a\u77e5\u8bc6\u578b\u673a\u5668\u9605\u8bfb\u7cfb\u7edf\uff0c\u5728\u4e00\u4e2a\u7edf\u4e00\u7684\u6df1\u5ea6\u5b66\u4e60\u9884\u8bad\u7ec3\u6846\u67b6\u4e2d\u8854\u63a5\u8fd9\u4e24\u7c7b\u4fe1\u606f\uff0c\u5b83\u89e3\u51b3\u4e86\u73b0\u6709\u7684\u673a\u5668\u9605\u8bfb\u6a21\u578b\u53ea\u80fd\u5206\u522b\u5904\u7406\u4e0d\u540c\u7c7b\u578b\u7684\u6570\u636e\u7684\u95ee\u9898\uff0c\u4ece\u800c\u5b9e\u73b0\u5bf9\u5206\u5b50\u5b9e\u4f53\u7684\u5168\u9762\u5f7b\u5e95\u7684\u7406\u89e3\u3002\u901a\u8fc7\u5728\u4e0d\u540c\u4fe1\u606f\u6e90\u5185\u548c\u4e0d\u540c\u4fe1\u606f\u6e90\u4e4b\u95f4\u4ee5\u9884\u8bad\u7ec3\u65e0\u76d1\u7763\u7684\u65b9\u5f0f\u638c\u63e1\u5143\u77e5\u8bc6\uff0c\u672c\u6587\u7684\u7cfb\u7edf\u53ef\u4ee5\u4fc3\u8fdb\u5404\u79cd\u73b0\u5b9e\u4e16\u754c\u7684\u751f\u7269\u533b\u5b66\u5e94\u7528\uff0c\u5305\u62ec\u5206\u5b50\u7279\u6027\u9884\u6d4b\u3001\u751f\u7269\u533b\u5b66\u5173\u7cfb\u63d0\u53d6\u7b49\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u672c\u6587\u7684\u7cfb\u7edf\u5728\u5206\u5b50\u7279\u6027\u7684\u7406\u89e3\u80fd\u529b\u4e0a\u751a\u81f3\u8d85\u8fc7\u4e86\u4eba\u7c7b\u4e13\u4e1a\u4eba\u58eb\uff0c\u540c\u65f6\u4e5f\u63ed\u793a\u4e86\u5176\u5728\u4fc3\u8fdb\u672a\u6765\u81ea\u52a8\u836f\u7269\u53d1\u73b0\u548c\u8bb0\u5f55\u65b9\u9762\u7684\u5de8\u5927\u6f5c\u529b\u3002<\/p>\n
\u8bba\u6587\u4e0b\u8f7d\uff1aA deep-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals<\/p>\n
AMiner\u5e73\u53f0\u6536\u5f55\u8d85\u8fc7 1.3 \u4ebf\u5b66\u8005\u30013.2 \u4ebf\u7bc7\u8bba\u6587\u30014 \u5343\u591a\u4e07\u4e2a\u4e13\u5229\u3001\u5c06\u8fd1 1 \u4e07\u4e2a\u6570\u636e\u96c6\u3001\u8d85\u8fc7 100 \u4e2a\u5f00\u653e\u7b97\u6cd5\u4f9b\u79d1\u6280\u5de5\u4f5c\u8005\u514d\u8d39\u68c0\u7d22\u4f7f\u7528\uff0c\u52a9\u529b\u79d1\u6280\u521b\u65b0\u3002<\/strong><\/em>
 \"bert\u9884\u8bad\u7ec3\u52a0lstm_pytorch\u9884\u8bad\u7ec3\u6a21\u578b
 \"bert\u9884\u8bad\u7ec3\u52a0lstm_pytorch\u9884\u8bad\u7ec3\u6a21\u578b<\/p>\n","protected":false},"excerpt":{"rendered":"bert\u9884\u8bad\u7ec3\u52a0lstm_pytorch\u9884\u8bad\u7ec3\u6a21\u578b\u6587\u7ae0\u8f6c\u8f7d\uff5c\u667a\u6e90\u793e\u533a\u672c\u671f\u8d21\u732e\u8005\uff5c\u7533\u5fb7\u5468\u7fdf\u73c2\u5434\u65b0\u521a\u5173\u4e8e\u5468\u520a\u672c\u671f\u5468\u520a\uff0c\u6211\u4eec\u9009\u62e9\u4e8612\u7bc7\u9884\u8bad....","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"_links":{"self":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/7812"}],"collection":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/comments?post=7812"}],"version-history":[{"count":0,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/posts\/7812\/revisions"}],"wp:attachment":[{"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/media?parent=7812"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/categories?post=7812"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mushiming.com\/wp-json\/wp\/v2\/tags?post=7812"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}