{"type":"doc","content":[{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"摘要：隨着用於傳遞和記錄業務信息的管理文檔的廣泛使用，能夠魯棒且高效地從這些文檔中自動提取和理解內容的方法成爲一個迫切的需求。本次解讀的文章提出利用圖神經網絡來解決半結構化文檔中的實體識別（NER）和關係提取問題。","attrs":{}}]}],"attrs":{}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本文分享自華爲雲社區","attrs":{}},{"type":"link","attrs":{"href":"https://bbs.huaweicloud.com/blogs/273142?utm_source=infoq&utm_medium=bbs-ex&utm_campaign=ei&utm_content=content","title":"","type":null},"content":[{"type":"text","text":"《論文解讀系列十一：圖神經網絡應用於半結構化文檔的命名實體識別和關係提取》","attrs":{}}]},{"type":"text","text":"，原文作者：小菜鳥chg。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/28/2809ce6fad7b20d554f32200c42fbd56.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"摘要：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着用於傳遞和記錄業務信息的管理文檔的廣泛使用，能夠魯棒且高效地從這些文檔中自動提取和理解內容的方法成爲一個迫切的需求。此外，基於圖的表達方法對不同文檔模版的變化具有靈活的適應性，從而使得圖表達方式與這些管理文檔的半結構化特性非常契合。正因爲圖神經網絡（GNN）能夠很好地學習出文檔中數據元素間的關係，所以本次解讀的文章提出利用圖神經網絡來解決半結構化文檔中的實體識別（NER）和關係提取問題。經實驗驗證該文章提出的方法在單詞分組、實體分類、關係預測三個任務上取得了SOTA結果，同時在FUNSD（表單理解）和IEHHR（手寫婚姻檔案理解）兩個完全不同類別的數據集上取得的實驗結果進一步驗證了本次解讀文章所提出的方法的泛化性。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"1.方法","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"GNN被廣泛應用於NER和表格提取等任務中，本次解讀的文章在此基礎上提出將GNN應用於提取key-value對的任務中，不僅對文檔圖片中的實體進行分類，而且還會對實體間的關係進行預測。 ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"給定一個輸入文檔，模型需要完成的任務包括：（a）單詞分組：檢測文檔實體，即將相同語義的單詞進行分組；（b）實體分類：將檢測到的實體分爲預設的類別；（c）關係預測：發現實體間配對關係。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"（1）圖的構造","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本次解讀的文章提出構造兩張圖來表示文檔，並在此基礎上訓練三個不同的模型來解決對應的任務：單詞分組f_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"1、實體分類f_2","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"2、關係預測f_3","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"3。如圖1所示，文檔會被表示爲由OCR結果構造的圖G_1=(V_1,E_1)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"1=(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"1,","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":"1)，其中V_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"1是由OCR結果中每個單詞組成的節點集合；對每個單詞文本框左上角間的距離進行k","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"k","attrs":{}},{"type":"text","text":"-近鄰（取k=10","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"k","attrs":{}},{"type":"text","text":"=10）來生成邊E_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":"1，對各邊計算分數s=f_1 (G_1)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"s","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"1(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"1)，篩選出大於閾值\\tau","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"τ","attrs":{}},{"type":"text","text":"（FUNSD設爲0.65，IEHHR設爲0.9）的邊就可以得到單詞分組的結果。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/58/58c64bbbb28a6f3f815cbdcedc9b11cb.jpeg","alt":"圖1 單詞分組的圖結構構造示意圖","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"圖1 單詞分組的圖結構構造示意圖","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/cc/cc5a40964f0c1d551f59d7839f569e6e.jpeg","alt":"圖2 實體分類和關係預測的圖結構構造示意圖","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"圖2 實體分類和關係預測的圖結構構造示意圖","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如圖2所示，在G_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"1的基礎上得到實體（即各單詞分組）後，由每個實體構造得到圖G_2=(V_2,E_2)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"2=(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"2,","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":"2)，其中V_2","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"2表示由G_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"1篩選得到的實體集合，E_2","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":"2是由各實體節點間全連接得到的邊集合。由c=f_2 (G_2)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"c","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"2(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"2)得到實體分類結果；由s=f_3 (G_3)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"s","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"3(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"3)得到關係預測結果。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"（2）圖的計算","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本次解讀文章中的f_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"1, f_2","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"2, f_3","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"3由L","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"L","attrs":{}},{"type":"text","text":"個GAT層（graph attention network）作爲模型骨幹結構並經過訓練優化得到。 ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"給定G=(V,E)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"=(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":")。每個節點v_i","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"vi","attrs":{}},{"type":"text","text":"的初始化表達由h_i^0=x_i,y_i,w_i,h_i,w_{embed}]","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hi","attrs":{}},{"type":"text","text":"0=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"xi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"yi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"wi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"wembed","attrs":{}},{"type":"text","text":"]拼接得到，其中x_i,y_i,w_i,h_i","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"xi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"yi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"wi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hi","attrs":{}},{"type":"text","text":"是單詞文本框的左上角橫縱座標和文本框寬高，w_{embed}","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"wembed","attrs":{}},{"type":"text","text":"爲單詞的詞向量。根據GAT，每一對節點間計算出其注意力係數：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"\\alpha_{ij}=\\frac{exp(LeakyRelu(V[Wh_i||Wh_j]))}{∑_{k \\in N_{v_i}}exp(LeakyRelu(V[Wh_i ||Wh_k]))}","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"αij","attrs":{}},{"type":"text","text":"=∑","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"k","attrs":{}},{"type":"text","text":"∈","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Nvi","attrs":{}},{"type":"text","text":"","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"exp","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"LeakyRelu","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"[","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Whi","attrs":{}},{"type":"text","text":"∣∣","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Whk","attrs":{}},{"type":"text","text":"]))","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"exp","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"LeakyRelu","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"[","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Whi","attrs":{}},{"type":"text","text":"∣∣","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Whj","attrs":{}},{"type":"text","text":"]))","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"其中W","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"W","attrs":{}},{"type":"text","text":"和V","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"是學習的權重參數。每個節點採用K","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"K","attrs":{}},{"type":"text","text":"個attention head，將各head的輸出結果拼接得到l+1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"l","attrs":{}},{"type":"text","text":"+1層的隱狀態輸出h_i^{l+1}","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hil","attrs":{}},{"type":"text","text":"+1：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"h_i^{l+1}=g(h_i )=||_{k=1}^K\\sigma(∑_{j \\in N_i}α_{ij}^k W^k h_j^l )","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hil","attrs":{}},{"type":"text","text":"+1=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"g","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hi","attrs":{}},{"type":"text","text":")=∣∣","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"k","attrs":{}},{"type":"text","text":"=1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"K","attrs":{}},{"type":"text","text":"","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"σ","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"j","attrs":{}},{"type":"text","text":"∈","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Ni","attrs":{}},{"type":"text","text":"∑","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"αijk","attrs":{}},{"type":"text","text":"","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Wkhjl","attrs":{}},{"type":"text","text":")","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於實體分類，將每一個節點的隱狀態表示（即GAT的輸出結果）傳遞到MLP中，得到分類結果：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"c_i=σ(Wh_i^L)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"ci","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"σ","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"WhiL","attrs":{}},{"type":"text","text":")","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於關係預測，將每一對節點的隱狀態表示（即GAT的輸出結果）的差值傳遞到MLP中，得到關係預測分數：s_ij=\\sigma(W(|h_i^L-h_j^L |))","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"si","attrs":{}},{"type":"text","text":"","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"j","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"σ","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"W","attrs":{}},{"type":"text","text":"(∣","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hiL","attrs":{}},{"type":"text","text":"−","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hjL","attrs":{}},{"type":"text","text":"∣))","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"將實體分類看做是節點分類，關係預測看做是邊分類，所有任務都用CE損失函數進行優化：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"CE(y')=-(y∙log(y')+(1-y)∙log(1-y' ))","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"CE","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"′)=−(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"∙","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"log","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"′)+(1−","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":")∙","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"log","attrs":{}},{"type":"text","text":"(1−","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"′))","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"其中y","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"是真實標註，y'","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"′是預測分數。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"2.實驗結果","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/53/537ae38a29901063629c8bf509af53c3.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從FUNSD實驗結果表明，本次解讀文章提出的方法與LayoutLM相比較還有優化空間，原因可能在於FUNSD的數據量較小。從IEHHR實驗結果表明，該方法在表單識別的其他領域即手寫記錄理解上也具有一定的效果，體現了其泛化性。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https://bbs.huaweicloud.com/blogs?utm_source=infoq&utm_medium=bbs-ex&utm_campaign=ei&utm_content=content","title":"","type":null},"content":[{"type":"text","text":"點擊關注，第一時間瞭解華爲雲新鮮技術~","attrs":{}}]}]}]}
論文解讀丨圖神經網絡應用於半結構化文檔的命名實體識別和關係提取
{"type":"doc","content":[{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"摘要：隨着用於傳遞和記錄業務信息的管理文檔的廣泛使用，能夠魯棒且高效地從這些文檔中自動提取和理解內容的方法成爲一個迫切的需求。本次解讀的文章提出利用圖神經網絡來解決半結構化文檔中的實體識別（NER）和關係提取問題。","attrs":{}}]}],"attrs":{}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本文分享自華爲雲社區","attrs":{}},{"type":"link","attrs":{"href":"https://bbs.huaweicloud.com/blogs/273142?utm_source=infoq&utm_medium=bbs-ex&utm_campaign=ei&utm_content=content","title":"","type":null},"content":[{"type":"text","text":"《論文解讀系列十一：圖神經網絡應用於半結構化文檔的命名實體識別和關係提取》","attrs":{}}]},{"type":"text","text":"，原文作者：小菜鳥chg。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/28/2809ce6fad7b20d554f32200c42fbd56.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"摘要：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"隨着用於傳遞和記錄業務信息的管理文檔的廣泛使用，能夠魯棒且高效地從這些文檔中自動提取和理解內容的方法成爲一個迫切的需求。此外，基於圖的表達方法對不同文檔模版的變化具有靈活的適應性，從而使得圖表達方式與這些管理文檔的半結構化特性非常契合。正因爲圖神經網絡（GNN）能夠很好地學習出文檔中數據元素間的關係，所以本次解讀的文章提出利用圖神經網絡來解決半結構化文檔中的實體識別（NER）和關係提取問題。經實驗驗證該文章提出的方法在單詞分組、實體分類、關係預測三個任務上取得了SOTA結果，同時在FUNSD（表單理解）和IEHHR（手寫婚姻檔案理解）兩個完全不同類別的數據集上取得的實驗結果進一步驗證了本次解讀文章所提出的方法的泛化性。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"1.方法","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"GNN被廣泛應用於NER和表格提取等任務中，本次解讀的文章在此基礎上提出將GNN應用於提取key-value對的任務中，不僅對文檔圖片中的實體進行分類，而且還會對實體間的關係進行預測。 ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"給定一個輸入文檔，模型需要完成的任務包括：（a）單詞分組：檢測文檔實體，即將相同語義的單詞進行分組；（b）實體分類：將檢測到的實體分爲預設的類別；（c）關係預測：發現實體間配對關係。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"（1）圖的構造","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本次解讀的文章提出構造兩張圖來表示文檔，並在此基礎上訓練三個不同的模型來解決對應的任務：單詞分組f_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"1、實體分類f_2","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"2、關係預測f_3","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"3。如圖1所示，文檔會被表示爲由OCR結果構造的圖G_1=(V_1,E_1)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"1=(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"1,","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":"1)，其中V_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"1是由OCR結果中每個單詞組成的節點集合；對每個單詞文本框左上角間的距離進行k","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"k","attrs":{}},{"type":"text","text":"-近鄰（取k=10","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"k","attrs":{}},{"type":"text","text":"=10）來生成邊E_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":"1，對各邊計算分數s=f_1 (G_1)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"s","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"1(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"1)，篩選出大於閾值\\tau","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"τ","attrs":{}},{"type":"text","text":"（FUNSD設爲0.65，IEHHR設爲0.9）的邊就可以得到單詞分組的結果。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/58/58c64bbbb28a6f3f815cbdcedc9b11cb.jpeg","alt":"圖1 單詞分組的圖結構構造示意圖","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"圖1 單詞分組的圖結構構造示意圖","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/cc/cc5a40964f0c1d551f59d7839f569e6e.jpeg","alt":"圖2 實體分類和關係預測的圖結構構造示意圖","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":"center","origin":null},"content":[{"type":"text","text":"圖2 實體分類和關係預測的圖結構構造示意圖","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如圖2所示，在G_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"1的基礎上得到實體（即各單詞分組）後，由每個實體構造得到圖G_2=(V_2,E_2)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"2=(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"2,","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":"2)，其中V_2","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"2表示由G_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"1篩選得到的實體集合，E_2","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":"2是由各實體節點間全連接得到的邊集合。由c=f_2 (G_2)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"c","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"2(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"2)得到實體分類結果；由s=f_3 (G_3)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"s","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"3(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"3)得到關係預測結果。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"（2）圖的計算","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本次解讀文章中的f_1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"1, f_2","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"2, f_3","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"f","attrs":{}},{"type":"text","text":"3由L","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"L","attrs":{}},{"type":"text","text":"個GAT層（graph attention network）作爲模型骨幹結構並經過訓練優化得到。 ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"給定G=(V,E)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"G","attrs":{}},{"type":"text","text":"=(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"E","attrs":{}},{"type":"text","text":")。每個節點v_i","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"vi","attrs":{}},{"type":"text","text":"的初始化表達由h_i^0=x_i,y_i,w_i,h_i,w_{embed}]","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hi","attrs":{}},{"type":"text","text":"0=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"xi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"yi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"wi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"wembed","attrs":{}},{"type":"text","text":"]拼接得到，其中x_i,y_i,w_i,h_i","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"xi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"yi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"wi","attrs":{}},{"type":"text","text":",","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hi","attrs":{}},{"type":"text","text":"是單詞文本框的左上角橫縱座標和文本框寬高，w_{embed}","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"wembed","attrs":{}},{"type":"text","text":"爲單詞的詞向量。根據GAT，每一對節點間計算出其注意力係數：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"\\alpha_{ij}=\\frac{exp(LeakyRelu(V[Wh_i||Wh_j]))}{∑_{k \\in N_{v_i}}exp(LeakyRelu(V[Wh_i ||Wh_k]))}","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"αij","attrs":{}},{"type":"text","text":"=∑","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"k","attrs":{}},{"type":"text","text":"∈","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Nvi","attrs":{}},{"type":"text","text":"","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"exp","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"LeakyRelu","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"[","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Whi","attrs":{}},{"type":"text","text":"∣∣","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Whk","attrs":{}},{"type":"text","text":"]))","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"exp","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"LeakyRelu","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"[","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Whi","attrs":{}},{"type":"text","text":"∣∣","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Whj","attrs":{}},{"type":"text","text":"]))","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"其中W","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"W","attrs":{}},{"type":"text","text":"和V","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"V","attrs":{}},{"type":"text","text":"是學習的權重參數。每個節點採用K","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"K","attrs":{}},{"type":"text","text":"個attention head，將各head的輸出結果拼接得到l+1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"l","attrs":{}},{"type":"text","text":"+1層的隱狀態輸出h_i^{l+1}","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hil","attrs":{}},{"type":"text","text":"+1：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"h_i^{l+1}=g(h_i )=||_{k=1}^K\\sigma(∑_{j \\in N_i}α_{ij}^k W^k h_j^l )","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hil","attrs":{}},{"type":"text","text":"+1=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"g","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hi","attrs":{}},{"type":"text","text":")=∣∣","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"k","attrs":{}},{"type":"text","text":"=1","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"K","attrs":{}},{"type":"text","text":"","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"σ","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"j","attrs":{}},{"type":"text","text":"∈","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Ni","attrs":{}},{"type":"text","text":"∑","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"αijk","attrs":{}},{"type":"text","text":"","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"Wkhjl","attrs":{}},{"type":"text","text":")","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於實體分類，將每一個節點的隱狀態表示（即GAT的輸出結果）傳遞到MLP中，得到分類結果：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"c_i=σ(Wh_i^L)","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"ci","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"σ","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"WhiL","attrs":{}},{"type":"text","text":")","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"對於關係預測，將每一對節點的隱狀態表示（即GAT的輸出結果）的差值傳遞到MLP中，得到關係預測分數：s_ij=\\sigma(W(|h_i^L-h_j^L |))","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"si","attrs":{}},{"type":"text","text":"","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"j","attrs":{}},{"type":"text","text":"=","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"σ","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"W","attrs":{}},{"type":"text","text":"(∣","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hiL","attrs":{}},{"type":"text","text":"−","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"hjL","attrs":{}},{"type":"text","text":"∣))","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"將實體分類看做是節點分類，關係預測看做是邊分類，所有任務都用CE損失函數進行優化：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"CE(y')=-(y∙log(y')+(1-y)∙log(1-y' ))","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"CE","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"′)=−(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"∙","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"log","attrs":{}},{"type":"text","text":"(","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"′)+(1−","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":")∙","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"log","attrs":{}},{"type":"text","text":"(1−","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"′))","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"其中y","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"是真實標註，y'","attrs":{}},{"type":"text","marks":[{"type":"italic","attrs":{}}],"text":"y","attrs":{}},{"type":"text","text":"′是預測分數。","attrs":{}}]},{"type":"heading","attrs":{"align":null,"level":1},"content":[{"type":"text","text":"2.實驗結果","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/53/537ae38a29901063629c8bf509af53c3.jpeg","alt":null,"title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"從FUNSD實驗結果表明，本次解讀文章提出的方法與LayoutLM相比較還有優化空間，原因可能在於FUNSD的數據量較小。從IEHHR實驗結果表明，該方法在表單識別的其他領域即手寫記錄理解上也具有一定的效果，體現了其泛化性。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https://bbs.huaweicloud.com/blogs?utm_source=infoq&utm_medium=bbs-ex&utm_campaign=ei&utm_content=content","title":"","type":null},"content":[{"type":"text","text":"點擊關注，第一時間瞭解華爲雲新鮮技術~","attrs":{}}]}]}]}
發表評論
所有評論
還沒有人評論，想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章