{"id":10659,"date":"2022-12-28T15:07:00","date_gmt":"2022-12-28T07:07:00","guid":{"rendered":"http:\/\/139.9.1.231\/?p=10659"},"modified":"2022-11-28T15:59:51","modified_gmt":"2022-11-28T07:59:51","slug":"prompting-ernie-layout","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/12\/28\/prompting-ernie-layout\/","title":{"rendered":"Prompting&#8212;ERNIE-Layout"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"325\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-119-1024x325.png\" alt=\"\" class=\"wp-image-10665\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-119-1024x325.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-119-300x95.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-119-768x244.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-119.png 1076w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u7c7b\u522b\uff1a\u8de8\u6a21\u6001\u5927\u6a21\u578b\uff08\u7528\u4e8e\u6587\u6863\u5206\u7c7b\u3001\u4fe1\u606f\u62bd\u53d6\u3001\u6587\u6863\u95ee\u7b54\u7b49\uff09<\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\"><a href=\"https:\/\/arxiv.org\/abs\/2210.06155\" target=\"_blank\" rel=\"noreferrer noopener\">\u2022ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding (EMNLP 2022)<\/a><\/p>\n\n\n\n<p class=\"has-light-pink-background-color has-background\"><a href=\"https:\/\/arxiv.org\/abs\/2107.13586\" target=\"_blank\" rel=\"noreferrer noopener\"> \u2022Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing (arxiv 2021)<\/a><\/p>\n\n\n\n<p class=\"has-light-blue-background-color has-background\">\u6a21\u578b\u8bd5\u73a9\uff1a<a href=\"https:\/\/huggingface.co\/spaces\/PaddlePaddle\/ERNIE-Layout\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/huggingface.co\/spaces\/PaddlePaddle\/ERNIE-Layout<\/a><\/p>\n\n\n\n<p class=\"has-light-gray-background-color has-background\">Github: <a rel=\"noreferrer noopener\" href=\"https:\/\/github.com\/PaddlePaddle\/PaddleNLP\/tree\/develop\/applications\/document_intelligence\" target=\"_blank\">https:\/\/github.com\/PaddlePaddle\/PaddleNLP\/tree\/develop\/applications\/document_intelligence<\/a><\/p>\n\n\n\n<p>        \u968f\u7740\u4f17\u591a\u884c\u4e1a\u7684\u6570\u5b57\u5316\u8f6c\u578b\uff0c\u7535\u5b50\u6587\u6863\u7684\u7ed3\u6784\u5316\u5206\u6790\u548c\u5185\u5bb9\u63d0\u53d6\u6210\u4e3a\u4e00\u9879\u70ed\u95e8\u7684\u7814\u7a76\u8bfe\u9898\u3002\u7535\u5b50\u6587\u6863\u5305\u62ec\u626b\u63cf\u56fe\u50cf\u6587\u4ef6\u548c\u8ba1\u7b97\u673a\u751f\u6210\u7684\u6570\u5b57\u6587\u6863\u4e24\u5927\u7c7b\uff0c\u6d89\u53ca\u5355\u636e\u3001\u884c\u4e1a\u62a5\u544a\u3001\u5408\u540c\u3001\u96c7\u4f63\u534f\u8bae\u3001\u53d1\u7968\u3001\u7b80\u5386\u7b49\u591a\u79cd\u7c7b\u578b\u3002\u667a\u80fd\u6587\u6863\u7406\u89e3\u4efb\u52a1\u4ee5\u7406\u89e3\u683c\u5f0f\u3001\u5e03\u5c40\u3001\u5185\u5bb9\u591a\u79cd\u591a\u6837\u7684\u6587\u6863\u4e3a\u76ee\u6807\uff0c\u5305\u62ec\u4e86\u6587\u6863\u5206\u7c7b\u3001\u6587\u6863\u4fe1\u606f\u62bd\u53d6\u3001\u6587\u6863\u95ee\u7b54\u7b49\u4efb\u52a1\u3002\u4e0e\u7eaf\u6587\u672c\u6587\u6863\u4e0d\u540c\u7684\u662f\uff0c\u6587\u6863\u5305\u542b\u8868\u683c\u3001\u56fe\u7247\u7b49\u591a\u79cd\u5185\u5bb9\uff0c\u5305\u542b\u4e30\u5bcc\u7684\u89c6\u89c9\u4fe1\u606f\u3002\u56e0\u4e3a\u6587\u6863\u5185\u5bb9\u4e30\u5bcc\u3001\u5e03\u5c40\u590d\u6742\u3001\u5b57\u4f53\u6837\u5f0f\u591a\u6837\u3001\u6570\u636e\u5b58\u5728\u566a\u58f0\uff0c\u6587\u6863\u7406\u89e3\u4efb\u52a1\u6781\u5177\u6311\u6218\u6027\u3002\u968f\u7740ERNIE\u7b49\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5728NLP\u9886\u57df\u53d6\u5f97\u4e86\u5de8\u5927\u7684\u6210\u529f\uff0c\u4eba\u4eec\u5f00\u59cb\u5173\u6ce8\u5728\u6587\u6863\u7406\u89e3\u9886\u57df\u8fdb\u884c\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u3002\u767e\u5ea6\u63d0\u51fa\u8de8\u6a21\u6001\u6587\u6863\u7406\u89e3\u6a21\u578b ERNIE-Layout\uff0c\u9996\u6b21\u5c06\u5e03\u5c40\u77e5\u8bc6\u589e\u5f3a\u6280\u672f\u878d\u5165\u8de8\u6a21\u6001\u6587\u6863\u9884\u8bad\u7ec3\uff0c\u5728 4 \u9879\u6587\u6863\u7406\u89e3\u4efb\u52a1\u4e0a\u5237\u65b0\u4e16\u754c\u6700\u597d\u6548\u679c\uff0c\u767b\u9876 DocVQA \u699c\u9996\u3002\u540c\u65f6\uff0cERNIE-Layout \u5df2\u96c6\u6210\u81f3\u767e\u5ea6\u667a\u80fd\u6587\u6863\u5206\u6790\u5e73\u53f0 TextMind\uff0c\u52a9\u529b\u4f01\u4e1a\u6570\u5b57\u5316\u5347\u7ea7\u3002<\/p>\n\n\n\n<h2>\u539f\u7406\u4ecb\u7ecd<\/h2>\n\n\n\n<p>       \u5bf9\u6587\u6863\u7406\u89e3\u6765\u8bf4\uff0c\u6587\u6863\u4e2d\u7684\u6587\u5b57\u9605\u8bfb\u987a\u5e8f\u81f3\u5173\u91cd\u8981\uff0c\u76ee\u524d\u4e3b\u6d41\u7684\u57fa\u4e8e OCR\uff08Optical Character Recognition\uff0c\u6587\u5b57\u8bc6\u522b\uff09\u6280\u672f\u7684\u6a21\u578b\u5927\u591a\u9075\u5faa\u300c\u4ece\u5de6\u5230\u53f3\u3001\u4ece\u4e0a\u5230\u4e0b\u300d\u7684\u539f\u5219\uff0c\u7136\u800c\u5bf9\u4e8e\u6587\u6863\u4e2d\u5206\u680f\u3001\u6587\u672c\u56fe\u7247\u8868\u683c\u6df7\u6742\u7684\u590d\u6742\u5e03\u5c40\uff0c\u6839\u636e OCR \u7ed3\u679c\u83b7\u53d6\u7684\u9605\u8bfb\u987a\u5e8f\u591a\u6570\u60c5\u51b5\u4e0b\u90fd\u662f\u9519\u8bef\u7684\uff0c\u4ece\u800c\u5bfc\u81f4\u6a21\u578b\u65e0\u6cd5\u51c6\u786e\u5730\u8fdb\u884c\u6587\u6863\u5185\u5bb9\u7684\u7406\u89e3\u3002<\/p>\n\n\n\n<p>     \u800c\u4eba\u7c7b\u901a\u5e38\u4f1a\u6839\u636e\u6587\u6863\u7ed3\u6784\u548c\u5e03\u5c40\u8fdb\u884c\u5c42\u6b21\u5316\u5206\u5757\u9605\u8bfb\uff0c\u53d7\u6b64\u542f\u53d1\uff0c\u767e\u5ea6\u7814\u7a76\u8005\u63d0\u51fa\u5728\u6587\u6863\u9884\u8bad\u6a21\u578b\u4e2d\u5bf9\u9605\u8bfb\u987a\u5e8f\u8fdb\u884c\u6821\u6b63\u7684\u5e03\u5c40\u77e5\u8bc6\u589e\u5f3a\u521b\u65b0\u601d\u8def\u3002TextMind \u5e73\u53f0\u4e0a\u4e1a\u754c\u9886\u5148\u7684\u6587\u6863\u89e3\u6790\u5de5\u5177\uff08Document Parser\uff09\u80fd\u591f\u51c6\u786e\u8bc6\u522b\u6587\u6863\u4e2d\u7684\u5206\u5757\u4fe1\u606f\uff0c\u4ea7\u51fa\u6b63\u786e\u7684\u6587\u6863\u9605\u8bfb\u987a\u5e8f\uff0c\u5c06\u9605\u8bfb\u987a\u5e8f\u4fe1\u53f7\u878d\u5408\u5230\u6a21\u578b\u7684\u8bad\u7ec3\u4e2d\uff0c\u4ece\u800c\u589e\u5f3a\u5bf9\u5e03\u5c40\u4fe1\u606f\u7684\u6709\u6548\u5229\u7528\uff0c\u63d0\u5347\u6a21\u578b\u5bf9\u4e8e\u590d\u6742\u6587\u6863\u7684\u7406\u89e3\u80fd\u529b\u3002<\/p>\n\n\n\n<p>      \u57fa\u4e8e\u5e03\u5c40\u77e5\u8bc6\u589e\u5f3a\u6280\u672f\uff0c\u540c\u65f6\u4f9d\u6258\u6587\u5fc3 ERNIE\uff0c\u767e\u5ea6\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u878d\u5408\u6587\u672c\u3001\u56fe\u50cf\u3001\u5e03\u5c40\u7b49\u4fe1\u606f\u8fdb\u884c\u8054\u5408\u5efa\u6a21\u7684\u8de8\u6a21\u6001\u901a\u7528\u6587\u6863\u9884\u8bad\u7ec3\u6a21\u578b ERNIE-Layout\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0cERNIE-Layout \u521b\u65b0\u6027\u5730\u63d0\u51fa\u4e86\u9605\u8bfb\u987a\u5e8f\u9884\u6d4b\u548c\u7ec6\u7c92\u5ea6\u56fe\u6587\u5339\u914d\u4e24\u4e2a\u81ea\u76d1\u7763\u9884\u8bad\u7ec3\u4efb\u52a1\uff0c\u6709\u6548\u63d0\u5347\u6a21\u578b\u5728\u6587\u6863\u4efb\u52a1\u4e0a\u8de8\u6a21\u6001\u8bed\u4e49\u5bf9\u9f50\u80fd\u529b\u548c\u5e03\u5c40\u7406\u89e3\u80fd\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/bce.bdstatic.com\/doc\/ai-doc\/wenxin\/image%20%2814%29_59cc6c8.png\" alt=\"image (14).png\"\/><\/figure>\n\n\n\n<p class=\"has-text-align-center\">\u6587\u5fc3ERNIE-Layout\u4ee5\u6587\u5fc3ERNIE\u4e3a\u5e95\u5ea7\uff0c\u878d\u5408\u6587\u672c\u3001\u56fe\u50cf\u3001\u5e03\u5c40\u7b49\u4fe1\u606f\u8fdb\u884c\u8de8\u6a21\u6001\u8054\u5408\u5efa\u6a21\uff0c\u521b\u65b0\u6027\u5f15\u5165\u5e03\u5c40\u77e5\u8bc6\u589e\u5f3a\uff0c\u63d0\u51fa\u9605\u8bfb\u987a\u5e8f\u9884\u6d4b\u3001\u7ec6\u7c92\u5ea6\u56fe\u6587\u5339\u914d\u7b49\u81ea\u76d1\u7763\u9884\u8bad\u7ec3\u4efb\u52a1\uff0c\u5347\u7ea7\u7a7a\u95f4\u89e3\u8026\u6ce8\u610f\u529b\u673a\u5236\u3002\u8f93\u5165\u57fa\u4e8eVIMER-StrucTexT\u5927\u6a21\u578b\u63d0\u4f9b\u7684OCR\u7ed3\u679c\uff0c\u5728\u5404\u6570\u636e\u96c6\u4e0a\u6548\u679c\u53d6\u5f97\u5927\u5e45\u5ea6\u63d0\u5347\uff0c\u76f8\u5173\u5de5\u4f5c\u5df2\u88abEMNLP 2022 Findings \u4f1a\u8bae\u6536\u5f55\u3002<\/p>\n\n\n\n<p class=\"has-text-align-center\"><img src=\"https:\/\/paddlepaddle-static.cdn.bcebos.com\/paddle-wechat-image\/mmbiz.qpic.cn\/mmbiz_png\/sKia1FKFiafgjDyKlqSGnoibX8WUxoIfiaYwMVY08WKS0yN2b8n0xwZibx0PpWgXaG37rRqmng1EicVZsaJcbJUutDAg\/image\">     \u25b2 \u6587\u5fc3ERNIE-Layout \u6280\u672f\u6846\u67b6<\/p>\n\n\n\n<h3><strong>Embedding<\/strong><\/h3>\n\n\n\n<p>Embedding \u7684\u8f93\u5165\u5305\u62ec\uff1a\u6587\u672c\u7684<code>token_ids<\/code>\uff0c\u6587\u672c\u5185\u5bb9\u5bf9\u5e94\u7684 bounding box\uff08\u5305\u542b&nbsp;<code>x1, x2,y1,y2,h,w<\/code>\uff09\uff0c\u56fe\u7247\uff0c\u4ee5\u53ca\u56fe\u7247\u5bf9\u5e94\u7684 bounding box\u3002<\/p>\n\n\n\n<p>\u5176\u4e2d bounding box \u7684\u6570\u503c\u88ab\u8f6c\u6362\u5230&nbsp;<code>0-1000<\/code>&nbsp;\u8303\u56f4\u3002\u800c\u540e\u901a\u8fc7\u4e00\u4e2a Embedding \u6765\u5206\u522b\u8ba1\u7b97\u5f97\u5230\u5bf9\u5e94\u7684&nbsp;<code>x1_embedding, x2_embedding, y1_embedding<\/code>&nbsp;\u7b49\u7b49 6 \u4e2a embeddings\u3002<\/p>\n\n\n\n<p><strong>\u6587\u5b57 Embedding<\/strong><\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>embeddings = (input_embedings + position_embeddings + x1 + y1 + x2 +\n              y2 + h + w + token_type_embeddings)\n<em># x1, y1, x2 , y2 , h , w \uff1a bounding box \u5404\u4e2a\u503c\u5bf9\u5e94\u7684 embedding<\/em>\n\u200b\nembeddings = self.layer_norm(embeddings)\ntext_embeddings = self.dropout(embeddings)<\/code><\/pre>\n\n\n\n<ul><li>\u5176\u4e2d\u91c7\u7528\u53ef\u5b66\u4e60\u7684&nbsp;<code>position_embeddings<\/code>\u3002<\/li><li>\u91c7\u7528&nbsp;<a rel=\"noreferrer noopener\" href=\"https:\/\/github.com\/Layout-Parser\/%20layout-parser\" target=\"_blank\">Layout-Parser<\/a>&nbsp;\u5bf9\u56fe\u7247\u4e2d\u7684\u6587\u672c\u5185\u5bb9\uff0c\u6839\u636e\u9605\u8bfb\u987a\u5e8f\u8fdb\u884c\u6392\u5e8f\uff0c\u5b89\u6392\u5bf9\u5e94\u7684&nbsp;<code>position_ids<\/code>\u3002<\/li><li><strong>Layout Embedding<\/strong>:the OCR tool provides its 2D coordinates with the width and height of the bounding box<\/li><\/ul>\n\n\n\n<p><strong>\u56fe\u50cf Embedding<\/strong><\/p>\n\n\n\n<p>\u56fe\u7247\u88ab\u8f6c\u6362\u6210&nbsp;<code>224* 224<\/code>&nbsp;\u7684\u683c\u5f0f\uff0c\u7ecf\u8fc7 backbone \u7f16\u7801\u540e\uff0c\u5206\u5272\u6210\u4e86&nbsp;<code>7*7<\/code>&nbsp;\u4e2a patch\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>x = self.visual(image)  <em># x &#91;batch, 49, 256]<\/em>\nvisual_embeddings = self.visual_act_fn(self.visual_proj(x)  <em># batch, 49, hidden_size<\/em><\/code><\/pre>\n\n\n\n<p>\u4e0e\u6587\u672c Embedding \u76f8\u540c\uff0c<code>visual_embeddings<\/code>&nbsp;\u9700\u8981\u518d\u52a0\u4e0a&nbsp;<code>position_embeddings<\/code>,&nbsp;<code>token_type_embeddigns<\/code>,&nbsp;<code>bbox_embeddigns<\/code>&nbsp;\u7b49\uff0c\u5f97\u5230\u6700\u7ec8\u56fe\u50cf embedding\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"536\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-120-1024x536.png\" alt=\"\" class=\"wp-image-10685\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-120-1024x536.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-120-300x157.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-120-768x402.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-120-1536x804.png 1536w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-120.png 1810w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3><strong>\u9884\u8bad\u7ec3<\/strong><\/h3>\n\n\n\n<ul><li><strong>Reading Order Prediction\uff1a<\/strong>\u5bf9\u6587\u5b57\u90e8\u5206\uff0c\u5224\u65adtoken\u4e4b\u95f4\u7684\u5148\u540e\u9605\u8bfb\u987a\u5e8f\u3002\u53ef\u4ee5\u901a\u8fc7\u9605\u8bfb\u987a\u5e8f\u6784\u5efa\u4e00\u4e2a\u5305\u542b 01 \u7684\u90bb\u63a5\u77e9\u9635\uff0c\u800c\u540e\u4e0e attention matrix \u8ba1\u7b97\u4ea4\u53c9\u71b5\u3002<\/li><li><strong>Replaced Region Prediction\uff1a<\/strong>\u5bf9\u4e8e\u56fe\u7247\u90e8\u5206\uff0c\u6709 10% \u7684\u6982\u7387\u66ff\u6362\u56fe\u7247 patch\uff0c\u901a\u8fc7 cls \u4f4d\u7f6e\u7684\u7f16\u7801\u5224\u65ad\u54ea\u4e9b patch \u88ab\u66ff\u6362\u4e86<\/li><li><strong>Masked Visual-Language Modeling<\/strong>\uff1a\u7c7b\u4f3c MLM\uff0c\u53ea\u662f\u8fd9\u6b21\u6211\u4eec\u53ef\u4ee5\u7528\u56fe\u7247\u90e8\u5206\u7684embedding\u4fe1\u606f\u6765\u9884\u6d4b\u88ab\u906e\u76d6\u7684\u6587\u5b57\u5185\u5bb9\u3002<\/li><li><strong>Text-Image Alignment<\/strong>\uff1a\u968f\u610f\u8986\u76d6\u4e00\u4e9b\u6587\u5b57\uff0c\u7136\u540e\u7528\u4e00\u4e2a\u7ebf\u6027\u5c42\u8fdb\u884c\u5206\u7c7b\u4efb\u52a1\uff0c\u5224\u65ad\u6587\u5b57\u662f\u5426\u88ab\u8986\u76d6\u4f4f\u4e86\u3002<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-121-1024x573.png\" alt=\"\" class=\"wp-image-10686\" width=\"690\" height=\"386\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-121-1024x573.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-121-300x168.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-121-768x430.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-121-1536x860.png 1536w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-121.png 1836w\" sizes=\"(max-width: 690px) 100vw, 690px\" \/><\/figure>\n\n\n\n<p class=\"has-text-align-center\">  \u6587\u5fc3ERNIE-mmLayout\u4e3a\u8fdb\u4e00\u6b65\u63a2\u7d22\u4e0d\u540c\u7c92\u5ea6\u5143\u7d20\u5173\u7cfb\u5bf9\u6587\u6863\u7406\u89e3\u7684\u4ef7\u503c\uff0c\u5728\u6587\u5fc3ERNIE-Layout\u7684\u57fa\u7840\u4e0a\u5f15\u5165\u57fa\u4e8eGNN\u7684\u591a\u7c92\u5ea6\u3001\u591a\u6a21\u6001Transformer\u5c42\uff0c\u5b9e\u73b0\u6587\u6863\u56fe\u805a\u5408\uff08Document Graph Aggregation\uff09\u8868\u793a\u3002\u6700\u7ec8\uff0c\u5728\u591a\u4e2a\u4fe1\u606f\u62bd\u53d6\u4efb\u52a1\u4e0a\u4ee5\u66f4\u5c11\u7684\u6a21\u578b\u53c2\u6570\u91cf\u8d85\u8fc7SOTA\u6210\u7ee9\uff0c\u76f8\u5173\u8bba\u6587\u88abACM MM 2022\u4f1a\u8bae\u6536\u5f55&nbsp;\u3002<img src=\"https:\/\/paddlepaddle-static.cdn.bcebos.com\/paddle-wechat-image\/mmbiz.qpic.cn\/mmbiz_jpg\/sKia1FKFiafgjDyKlqSGnoibX8WUxoIfiaYwR1vSgvOyypALxHlGd0jdNF8gwrzoVvXbD3FHGkZIko12UIo3FZZjibQ\/image\">\u25b2 \u6587\u5fc3ERNIE-mmLayout \u6280\u672f\u6846\u67b6<\/p>\n\n\n\n<p>Ernie-layout \u6574\u4f53\u91c7\u7528 Transformer Encoder \u67b6\u6784\uff0c\u7279\u70b9\u5728\u4e8e\uff1a<\/p>\n\n\n\n<ul><li>\u501f\u9274\u4e86 DeBERTa \u7684\u89e3\u8026\u6ce8\u610f\u529b\uff0c\u4f9d\u9760\u989d\u5916\u7684&nbsp;<a href=\"https:\/\/github.com\/Layout-Parser\/%20layout-parser\" target=\"_blank\" rel=\"noreferrer noopener\">Layout-Parser<\/a>&nbsp;\u6765\u8bbe\u8ba1 position_ids\u3002<\/li><li>\u540c\u65f6\u5bf9\u6587\u6863\u56fe\u7247\u53ca\u6587\u6863\u4e2d\u7684\u6587\u5b57\u8fdb\u884c\u7f16\u7801\uff0c\u5e76\u8bbe\u8ba1\u4e864\u79cd\u56fe\u6587\u7ed3\u5408\u7684\u9884\u8bad\u7ec3\u65b9\u5f0f\u3002<\/li><li>\u9700\u8981\u4f9d\u9760\u989d\u5916\u7684 OCR \u5de5\u5177\u6765\u83b7\u5f97\u56fe\u7247\u4e2d\u7684\u6587\u5b57\u5185\u5bb9\uff0c\u53ca\u5176\u5bf9\u5e94\u4f4d\u7f6e\u4fe1\u606f\u3002<\/li><\/ul>\n\n\n\n<p>\u4ee5\u4e0b\u662f\u6587\u6863\u667a\u80fd\u6280\u672f\u7684\u4e00\u4e9b\u5e94\u7528\u573a\u666f\u5c55\u793a\uff1a<\/p>\n\n\n\n<ul><li>\u53d1\u7968\u62bd\u53d6\u95ee\u7b54<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image\"><a href=\"https:\/\/user-images.githubusercontent.com\/40840292\/196118171-fd3e49a0-b9f1-4536-a904-c48f709a2dec.png\" target=\"_blank\" rel=\"noreferrer noopener\"><img src=\"https:\/\/user-images.githubusercontent.com\/40840292\/196118171-fd3e49a0-b9f1-4536-a904-c48f709a2dec.png\" alt=\"\"\/><\/a><\/figure>\n\n\n\n<ul><li>\u6d77\u62a5\u62bd\u53d6\u95ee\u7b54<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image\"><a href=\"https:\/\/user-images.githubusercontent.com\/40840292\/195610368-04230855-62de-439e-b708-2c195b70461f.png\" target=\"_blank\" rel=\"noreferrer noopener\"><img src=\"https:\/\/user-images.githubusercontent.com\/40840292\/195610368-04230855-62de-439e-b708-2c195b70461f.png\" alt=\"\"\/><\/a><\/figure>\n\n\n\n<ul><li>\u7f51\u9875\u62bd\u53d6\u95ee\u7b54<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image\"><a href=\"https:\/\/user-images.githubusercontent.com\/40840292\/195611613-bdbe692e-d7f2-4a2b-b548-1a933463b0b9.png\" target=\"_blank\" rel=\"noreferrer noopener\"><img src=\"https:\/\/user-images.githubusercontent.com\/40840292\/195611613-bdbe692e-d7f2-4a2b-b548-1a933463b0b9.png\" alt=\"\"\/><\/a><\/figure>\n\n\n\n<ul><li>\u8868\u683c\u62bd\u53d6\u95ee\u7b54<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image\"><a href=\"https:\/\/user-images.githubusercontent.com\/40840292\/195610692-8367f1c8-32c2-4b5d-9514-a149795cf609.png\" target=\"_blank\" rel=\"noreferrer noopener\"><img src=\"https:\/\/user-images.githubusercontent.com\/40840292\/195610692-8367f1c8-32c2-4b5d-9514-a149795cf609.png\" alt=\"\"\/><\/a><\/figure>\n\n\n\n<ul><li>\u8bd5\u5377\u62bd\u53d6\u95ee\u7b54<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image\"><a href=\"https:\/\/user-images.githubusercontent.com\/40840292\/195823294-d891d95a-2ef8-4519-be59-0fedb96c00de.png\" target=\"_blank\" rel=\"noreferrer noopener\"><img src=\"https:\/\/user-images.githubusercontent.com\/40840292\/195823294-d891d95a-2ef8-4519-be59-0fedb96c00de.png\" alt=\"\"\/><\/a><\/figure>\n\n\n\n<ul><li>\u82f1\u6587\u7968\u636e\u591a\u8bed\u79cd\uff08\u4e2d\u3001\u82f1\u3001\u65e5\u3001\u6cf0\u3001\u897f\u73ed\u7259\u3001\u4fc4\u8bed\uff09\u62bd\u53d6\u95ee\u7b54<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image\"><a href=\"https:\/\/user-images.githubusercontent.com\/40840292\/195610820-7fb88608-b317-45fc-a6ab-97bf3b20a4ac.png\" target=\"_blank\" rel=\"noreferrer noopener\"><img src=\"https:\/\/user-images.githubusercontent.com\/40840292\/195610820-7fb88608-b317-45fc-a6ab-97bf3b20a4ac.png\" alt=\"\"\/><\/a><\/figure>\n\n\n\n<ul><li>\u4e2d\u6587\u7968\u636e\u591a\u8bed\u79cd\uff08\u4e2d\u7b80\u3001\u4e2d\u7e41\u3001\u82f1\u3001\u65e5\u3001\u6cd5\u8bed\uff09\u62bd\u53d6\u95ee\u7b54<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image\"><a href=\"https:\/\/user-images.githubusercontent.com\/40840292\/195611075-9323ce9f-134b-4657-ab1c-f4892075d909.png\" target=\"_blank\" rel=\"noreferrer noopener\"><img src=\"https:\/\/user-images.githubusercontent.com\/40840292\/195611075-9323ce9f-134b-4657-ab1c-f4892075d909.png\" alt=\"\"\/><\/a><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"586\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-122-1024x586.png\" alt=\"\" class=\"wp-image-10688\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-122-1024x586.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-122-300x172.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-122-768x439.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-122-1536x879.png 1536w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-122.png 1662w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2><strong>Visual Prompting<\/strong><\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"973\" height=\"497\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-123.png\" alt=\"\" class=\"wp-image-10689\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-123.png 973w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-123-300x153.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-123-768x392.png 768w\" sizes=\"(max-width: 973px) 100vw, 973px\" \/><\/figure>\n\n\n\n<p>(a)Fine-tuning adapts the entire model parameters. <\/p>\n\n\n\n<p>(b)Linear probes adapt the model outputs (usually activations at the penultimate layer) by learning a linear layer. <\/p>\n\n\n\n<p>(c)Prompting adapts the (downstream) dataset by reformulating the input and\/or output.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"793\" height=\"149\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-124.png\" alt=\"\" class=\"wp-image-10693\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-124.png 793w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-124-300x56.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-124-768x144.png 768w\" sizes=\"(max-width: 793px) 100vw, 793px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"655\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-125-1024x655.png\" alt=\"\" class=\"wp-image-10696\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-125-1024x655.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-125-300x192.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-125-768x491.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-125.png 1311w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p class=\"has-light-pink-background-color has-background\"><strong><em>\u76f8\u5173\u8bba\u6587\uff1a<a rel=\"noreferrer noopener\" href=\"https:\/\/github.com\/thunlp\/PromptPapers\" target=\"_blank\">https:\/\/github.com\/thunlp\/PromptPapers<\/a><\/em><\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"379\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-126-1024x379.png\" alt=\"\" class=\"wp-image-10697\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-126-1024x379.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-126-300x111.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-126-768x284.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/11\/image-126.png 1102w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>\u7c7b\u522b\uff1a\u8de8\u6a21\u6001\u5927\u6a21\u578b\uff08\u7528\u4e8e\u6587\u6863\u5206\u7c7b\u3001\u4fe1\u606f\u62bd\u53d6\u3001\u6587\u6863\u95ee\u7b54\u7b49\uff09 \u2022ERNIE-Layout: Layout Know &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2022\/12\/28\/prompting-ernie-layout\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">Prompting&#8212;ERNIE-Layout<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[21],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/10659"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=10659"}],"version-history":[{"count":35,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/10659\/revisions"}],"predecessor-version":[{"id":10702,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/10659\/revisions\/10702"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=10659"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=10659"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=10659"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}