-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathkg2text-Notes.html
782 lines (618 loc) · 43.6 KB
/
kg2text-Notes.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
<!DOCTYPE html>
<html lang="zh-CN,default">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=2">
<meta name="theme-color" content="#222">
<meta name="generator" content="Hexo 4.2.1">
<link rel="apple-touch-icon" sizes="180x180" href="/images/apple-touch-icon-next.png">
<link rel="icon" type="image/png" sizes="32x32" href="/images/favicom-32x32-logo.png">
<link rel="icon" type="image/png" sizes="16x16" href="/images/favicon-16x16-logo.png">
<link rel="mask-icon" href="/images/logo.svg" color="#222">
<meta name="google-site-verification" content="ROUQoZovui7E3zu9iZWkJdCuxBSmH6dcEV2qncmZ4DQ">
<meta name="msvalidate.01" content="A5DD278FA75034DFB543D13A5DEF0F94">
<meta name="baidu-site-verification" content="h7T48zL37w">
<link rel="stylesheet" href="/css/main.css">
<link rel="stylesheet" href="/lib/font-awesome/css/all.min.css">
<script id="hexo-configurations">
var NexT = window.NexT || {};
var CONFIG = {"hostname":"blog.charfole.top","root":"/","scheme":"Gemini","version":"7.8.0","exturl":true,"sidebar":{"position":"left","display":"post","padding":18,"offset":12,"onmobile":false},"copycode":{"enable":false,"show_result":false,"style":null},"back2top":{"enable":true,"sidebar":true,"scrollpercent":true},"bookmark":{"enable":false,"color":"#222","save":"auto"},"fancybox":false,"mediumzoom":false,"lazyload":false,"pangu":false,"comments":{"style":"tabs","active":null,"storage":true,"lazyload":false,"nav":null},"algolia":{"hits":{"per_page":10},"labels":{"input_placeholder":"Search for Posts","hits_empty":"We didn't find any results for the search: ${query}","hits_stats":"${hits} results found in ${time} ms"}},"localsearch":{"enable":true,"trigger":"auto","top_n_per_article":1,"unescape":false,"preload":false},"motion":{"enable":true,"async":false,"transition":{"post_block":"fadeIn","post_header":"slideDownIn","post_body":"slideDownIn","coll_header":"slideLeftIn","sidebar":"slideUpIn"}},"path":"search.xml"};
</script>
<meta name="description" content="Paper Information Title: Modeling Global and Local Node Contexts for Text Generation from Knowledge GraphsLinks: https://arxiv.org/abs/2001.11003Date: 2020.06.22Comments: Transactions of the Associati">
<meta property="og:type" content="article">
<meta property="og:title" content="KG2Text-Notes">
<meta property="og:url" content="https://blog.charfole.top/kg2text-Notes.html">
<meta property="og:site_name" content="Charfole's Blog">
<meta property="og:description" content="Paper Information Title: Modeling Global and Local Node Contexts for Text Generation from Knowledge GraphsLinks: https://arxiv.org/abs/2001.11003Date: 2020.06.22Comments: Transactions of the Associati">
<meta property="og:locale" content="zh_CN">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310214058323.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210306154921113.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310215230970.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310215245017.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210307195339134.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210309210539790.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310115529028.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310115637199.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310155325418.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310161944010.png">
<meta property="og:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310205641461.png">
<meta property="article:published_time" content="2021-03-14T12:05:10.000Z">
<meta property="article:modified_time" content="2023-09-21T09:33:05.653Z">
<meta property="article:author" content="Charfole">
<meta property="article:tag" content="Paper">
<meta property="article:tag" content="Notes">
<meta name="twitter:card" content="summary">
<meta name="twitter:image" content="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310214058323.png">
<link rel="canonical" href="https://blog.charfole.top/kg2text-Notes.html">
<script id="page-configurations">
// https://hexo.io/docs/variables.html
CONFIG.page = {
sidebar: "",
isHome : false,
isPost : true,
lang : 'zh-CN'
};
</script>
<title>KG2Text-Notes | Charfole's Blog</title>
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-174017535-1"></script>
<script>
if (CONFIG.hostname === location.hostname) {
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-174017535-1');
}
</script>
<noscript>
<style>
.use-motion .brand,
.use-motion .menu-item,
.sidebar-inner,
.use-motion .post-block,
.use-motion .pagination,
.use-motion .comments,
.use-motion .post-header,
.use-motion .post-body,
.use-motion .collection-header { opacity: initial; }
.use-motion .site-title,
.use-motion .site-subtitle {
opacity: initial;
top: initial;
}
.use-motion .logo-line-before i { left: initial; }
.use-motion .logo-line-after i { right: initial; }
</style>
</noscript>
<link rel="alternate" href="/atom.xml" title="Charfole's Blog" type="application/atom+xml">
</head>
<body itemscope itemtype="http://schema.org/WebPage">
<div class="container use-motion">
<div class="headband"></div>
<header class="header" itemscope itemtype="http://schema.org/WPHeader">
<div class="header-inner"><div class="site-brand-container">
<div class="site-nav-toggle">
<div class="toggle" aria-label="切换导航栏">
<span class="toggle-line toggle-line-first"></span>
<span class="toggle-line toggle-line-middle"></span>
<span class="toggle-line toggle-line-last"></span>
</div>
</div>
<div class="site-meta">
<a href="/" class="brand" rel="start">
<span class="logo-line-before"><i></i></span>
<h1 class="site-title">Charfole's Blog</h1>
<span class="logo-line-after"><i></i></span>
</a>
</div>
<div class="site-nav-right">
<div class="toggle popup-trigger">
<i class="fa fa-search fa-fw fa-lg"></i>
</div>
</div>
</div>
<nav class="site-nav">
<ul id="menu" class="main-menu menu">
<li class="menu-item menu-item-home">
<a href="/" rel="section"><i class="fa fa-home fa-fw"></i>首页</a>
</li>
<li class="menu-item menu-item-archives">
<a href="/archives/" rel="section"><i class="fa fa-archive fa-fw"></i>归档</a>
</li>
<li class="menu-item menu-item-tags">
<a href="/tags/" rel="section"><i class="fa fa-tags fa-fw"></i>标签</a>
</li>
<li class="menu-item menu-item-categories">
<a href="/categories/" rel="section"><i class="fa fa-th fa-fw"></i>分类</a>
</li>
<li class="menu-item menu-item-search">
<a role="button" class="popup-trigger"><i class="fa fa-search fa-fw"></i>搜索
</a>
</li>
</ul>
</nav>
<div class="search-pop-overlay">
<div class="popup search-popup">
<div class="search-header">
<span class="search-icon">
<i class="fa fa-search"></i>
</span>
<div class="search-input-container">
<input autocomplete="off" autocapitalize="off"
placeholder="搜索..." spellcheck="false"
type="search" class="search-input">
</div>
<span class="popup-btn-close">
<i class="fa fa-times-circle"></i>
</span>
</div>
<div id="search-result">
<div id="no-result">
<i class="fa fa-spinner fa-pulse fa-5x fa-fw"></i>
</div>
</div>
</div>
</div>
</div>
</header>
<div class="reading-progress-bar"></div>
<main class="main">
<div class="main-inner">
<div class="content-wrap">
<div class="content post posts-expand">
<article itemscope itemtype="http://schema.org/Article" class="post-block" lang="zh-CN">
<link itemprop="mainEntityOfPage" href="https://blog.charfole.top/kg2text-Notes.html">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="image" content="/images/avatar.gif">
<meta itemprop="name" content="Charfole">
<meta itemprop="description" content="Carpe diem">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="Charfole's Blog">
</span>
<header class="post-header">
<h1 class="post-title" itemprop="name headline">
KG2Text-Notes
</h1>
<div class="post-meta">
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar"></i>
</span>
<span class="post-meta-item-text">发表于</span>
<time title="创建时间：2021-03-14 20:05:10" itemprop="dateCreated datePublished" datetime="2021-03-14T20:05:10+08:00">2021-03-14</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-calendar-check"></i>
</span>
<span class="post-meta-item-text">更新于</span>
<time title="修改时间：2023-09-21 17:33:05" itemprop="dateModified" datetime="2023-09-21T17:33:05+08:00">2023-09-21</time>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-folder"></i>
</span>
<span class="post-meta-item-text">分类于</span>
<span itemprop="about" itemscope itemtype="http://schema.org/Thing">
<a href="/categories/%E7%A7%91%E7%A0%94/" itemprop="url" rel="index"><span itemprop="name">科研</span></a>
</span>
</span>
<span class="post-meta-item" title="阅读次数" id="busuanzi_container_page_pv" style="display: none;">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="post-meta-item-text">阅读次数：</span>
<span id="busuanzi_value_page_pv"></span>
</span>
<span class="post-meta-item">
<span class="post-meta-item-icon">
<i class="far fa-comment"></i>
</span>
<span class="post-meta-item-text">Valine：</span>
<a title="valine" href="/kg2text-Notes.html#valine-comments" itemprop="discussionUrl">
<span class="post-comments-count valine-comment-count" data-xid="/kg2text-Notes.html" itemprop="commentCount"></span>
</a>
</span>
<br>
<span class="post-meta-item" title="本文字数">
<span class="post-meta-item-icon">
<i class="far fa-file-word"></i>
</span>
<span class="post-meta-item-text">本文字数：</span>
<span>5k</span>
</span>
<span class="post-meta-item" title="阅读时长">
<span class="post-meta-item-icon">
<i class="far fa-clock"></i>
</span>
<span class="post-meta-item-text">阅读时长 ≈</span>
<span>5 分钟</span>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h3 id="Paper-Information"><a href="#Paper-Information" class="headerlink" title="Paper Information"></a>Paper Information</h3><hr>
<p><strong>Title</strong>: Modeling Global and Local Node Contexts for Text Generation from Knowledge Graphs<br><strong>Links</strong>: <span class="exturl" data-url="aHR0cHM6Ly9hcnhpdi5vcmcvYWJzLzIwMDEuMTEwMDM=">https://arxiv.org/abs/2001.11003<i class="fa fa-external-link-alt"></i></span><br><strong>Date</strong>: 2020.06.22<br><strong>Comments</strong>: Transactions of the Association for Computational Linguistics (TACL)<br><strong>Subjects</strong>: KG、AI<br><strong>Index Terms</strong>: Knowledge Graphs, Text Generation<br><strong>Authors</strong>: Leonardo F. R. Ribeiro, Yue Zhang, Claire Gardent, Iryna Gurevych<br><a id="more"></a></p>
<h3 id="Notes"><a href="#Notes" class="headerlink" title="Notes"></a>Notes</h3><hr>
<h3 id="Abstract"><a href="#Abstract" class="headerlink" title="Abstract"></a>Abstract</h3><ul>
<li>融合了 Global Node Encode 和 Local Node Encode 来构造新的神经网络，从而更好地学习上下文节点嵌入。</li>
<li>运用 <span class="exturl" data-url="aHR0cHM6Ly9naXRodWIuY29tL3Jpa2R6L0dyYXBoV3JpdGVyL3RyZWUvbWFzdGVyL2RhdGE=">AGENDA<i class="fa fa-external-link-alt"></i></span> 数据集和 <span class="exturl" data-url="aHR0cHM6Ly93ZWJubGctY2hhbGxlbmdlLmxvcmlhLmZyL2NoYWxsZW5nZV8yMDE3Lw==">WEBNLG<i class="fa fa-external-link-alt"></i></span> 数据集进行实验，各项评价指标得到了提升。</li>
</ul>
<hr>
<h3 id="Introduction"><a href="#Introduction" class="headerlink" title="Introduction"></a>Introduction</h3><ul>
<li>图到文本的生成指的是根据输入的图结构生成对应的自然语言文本，图可以指的是语义表示或知识图谱，文本包括单一的句子或包含多行句子的完整文本，<strong>而本文的任务是根据知识图谱生成完整文本。</strong></li>
<li><p>Encode：</p>
<ul>
<li>Global Node Encode：优点为考虑了大量的上下文节点，但因为将所有节点都看作和其他节点简单相连而忽略了图的拓扑结构。</li>
<li>Local Node Encode：将每个节点的相邻节点情况，即图的拓扑结构考虑到其中，但其缺点是难以构造图中相距较远节点的关系。<img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310214058323.png" alt="image-20210310214058323"></li>
</ul>
</li>
<li><p>主要贡献</p>
<ul>
<li>首次融合了 Global Node Encode 和 Local Node Encode 来构建 graph-to-text 模型。</li>
<li>首次提出了一个将 Global Node Encode 和 Local Node Encode 进行组合的 GAT 架构。</li>
</ul>
</li>
</ul>
<hr>
<h3 id="Related-Work"><a href="#Related-Work" class="headerlink" title="Related Work"></a>Related Work</h3><ul>
<li><strong>AMR-to-Text</strong>：AMR 代表 Abstract Meaning Representation Graphs，是图的其中一种，具体的例子可参照下图，将 AMR 转成文本的已经有多个研究，使用的方法包括但不限于：GNN、GCN、LSTM 等。</li>
</ul>
<p> <img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210306154921113.png" alt="image-20210306154921113" style="zoom: 60%;"></p>
<ul>
<li><p><strong>KG-to-Text</strong>：知识图谱和 AMR 相比更加稀疏，有着更庞大数量的关系且没有固定的拓扑结构；不同数据集的知识图谱会有着较大的差异，这使得生成文本的过程更加困难。该任务常用的方法包括但不限于：LSTM/GRU、GNN、GCN、Transformer。</p>
</li>
<li><p><strong>加入图的全局信息</strong>：为了更好地完成 graph-to-text 的工作，越来越多的研究加入了全局的节点信息，大部分的工作都是通过扩展图的结构，在图中加入一个全局节点来完成的。</p>
</li>
</ul>
<hr>
<h3 id="Graph-to-Text-Model"><a href="#Graph-to-Text-Model" class="headerlink" title="Graph-to-Text Model"></a>Graph-to-Text Model</h3><p>论文的这一部分包含以下内容：</p>
<ol>
<li>如何将输入的数据转换成关系图。</li>
<li>描述如何使用 GAT 构建 graph encoders。</li>
<li>描述将 global encoder 和 local encoder 结合的方法。</li>
<li>描述 decode 和训练模型的过程。</li>
</ol>
<h4 id="Graph-Preparation"><a href="#Graph-Preparation" class="headerlink" title="Graph Preparation"></a>Graph Preparation</h4><p>每个 KG 都是一个由带权边组成的有向图，它的表达形式为：$G_e = (V_e, ε_e, R)$，其实体节点的表示为 $e∈V_e$，其带权边集为 $(e_h, r, e_t)∈ε_e$，代表了实体 $e_h$ 和 $e_t$ 存在关系 $r∈R$。</p>
<p>有一点和其他方法不同的是，这里将图中的<strong>实体集看成是一组节点的集合</strong>，其中每个组成实体的符号 (token) 都是一个节点。例如，KG 中有一个实体是 “node embedding”，那么该实体则由两个节点组成，分别为 node 和 embedding。各节点之间的边的关系对应着其所属实体之间的边的关系，即边 $(u,r,v)∈ε$ 当且仅当存在一条边 $(e<em>h, r, e_t)∈ε_e$，且 $u∈e_h$， $v∈e_t$ 。节点 $v$ 可以表示为一个嵌入 (embedding)：$h_v^0∈R^{d</em>{v}}$。</p>
<p>这一种图的表示方法有着很好的表达能力，但它也有一个副作用，便是消去了原实体中的单词顺序信息，为了避免该种影响，应该在 embedding 中同时加入对应 token 的位置信息。</p>
<h4 id="Graph-Neural-Networks-GNN"><a href="#Graph-Neural-Networks-GNN" class="headerlink" title="Graph Neural Networks (GNN)"></a>Graph Neural Networks (GNN)</h4><p>GNN 的工作原理为：通过学习节点的上下文节点表示和其边信息，通过信息传播机制来迭代更新当前节点的 embedding。第 $l$ 层 GNN 关于 $v$ 的上下文节点表示的公式为：<br><img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310215230970.png" alt="image-20210310215230970"></p>
<p>其中 $AGGR^{l}(.)$ 是 $l$ 层上的聚合函数 (aggregation function)，$r<em>{uv}$ 代表了 $u$ 和 $v$ 之间的关系，$N(v)$ 是 $v$ 的所有上下文节点集合，即那些与 $v$ 相邻的节点。我们可以将得到$h</em>{N_{(v)}}^{(l)}$ 用于更新第 $l$ 层节点 $v$ 的表示，公式为：<br><img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310215245017.png" alt="image-20210310215245017"></p>
<p>在 $L$ 次迭代后，一个节点的表示包含了当前迭代中的上下文节点信息。$AGGR^{l}(.)$ 和 $COMBINE^{l}(.)$ 函数的选择根据 GNN 的不同而不同，常见的 $AGGR^{l}(.)$ 是对 $N(v)$ 求和，而$COMBINE^{l}(.)$ 函数则通常对表示向量进行拼接 (concatenation) 。</p>
<h4 id="Global-Graph-Encoder"><a href="#Global-Graph-Encoder" class="headerlink" title="Global Graph Encoder"></a>Global Graph Encoder</h4><p>全局图编码器在更新每个节点的表示时需要考虑全图的节点，这里采用了注意力机制作为消息传递机制，并将其扩展成为 GAT 结构。该编码器的公式如下所示：</p>
<script type="math/tex; mode=display">
h_{N_{(V)}} = \sum_{u∈V} a_{vu}W_{g}h_u</script><p>其中 $W<em>g$ 是模型的参数，注意力权重 $a</em>{vu}$ 的公式为：</p>
<script type="math/tex; mode=display">
a_{vu} = \frac {exp(e_{vu})} {\sum_{k∈V}exp(e_{vk})}</script><p>其中 $e_{vu}$ 用于权衡节点 $u$ 对 $v$ 的重要性，其公式为：</p>
<script type="math/tex; mode=display">
e_{vu} = ((W_qH_v)\top(W_kh_u))/d_z</script><p>为了捕捉到节点之间的不同关系，一共设计了 $K$ 个独立的全局卷积，将其计算完毕后进行拼接，有</p>
<script type="math/tex; mode=display">
\hat{h}{_{N(v)}} = {||}_{k=1}^K h_{N(v)}^{(k)}</script><p>最后，$COMBINE^{l}(.)$ 函数由 LayerNorm 和 FFN 结构组成，其公式推导为：</p>
<script type="math/tex; mode=display">
\hat{h_v}=LayerNorm(\hat{h}{_{N(v)}}+h_v),</script><script type="math/tex; mode=display">
h_v^{global} = FFN(\hat{h_v}+\hat{h}_{N(v)}+h_v)</script><h4 id="Local-Graph-Encoder"><a href="#Local-Graph-Encoder" class="headerlink" title="Local Graph Encoder"></a>Local Graph Encoder</h4><p>全局编码层中没有考虑到节点之间的边的信息和图的结构，为了补充这些信息，需要结合局部编码器到模型当中。其 $AGGR^{l}(.)$ 函数为：</p>
<script type="math/tex; mode=display">
h_{N(v)} = \sum_{u∈N(v)}a_{vu}W_rh_u</script><p>其中 $W<em>r$ 代表了节点 $u$ 和 $v$ 之间的关系，注意力函数 $a</em>{vu}$ 的计算公式为：</p>
<script type="math/tex; mode=display">
a_{vu} = \frac {exp(e_{vu})} {\sum_{k∈N(v)}exp(e_{vk})}</script><p>$e_{vu}$ 的计算公式为：</p>
<script type="math/tex; mode=display">
e_{vu} = σ(a\top[W_vh_v||W_rh_u])</script><p>其中 σ 是激活函数，$a$ 和 $W<em>v$ 是模型参数。同样地，这里将 $K$ 个拼接起来，得到 $\hat{h}{</em>{N(v)}}$。那么$COMBINE^{l}(.)$ 函数的定义为：</p>
<script type="math/tex; mode=display">
h_v^{local} = RNN (h_v,\hat{h}_{N(v)})</script><h4 id="Combining-Global-and-Local-Encodings"><a href="#Combining-Global-and-Local-Encodings" class="headerlink" title="Combining Global and Local Encodings"></a>Combining Global and Local Encodings</h4><p>直观地来说，合并两种编码器一般有两种方法，第一种方法是并行结构，也就是将全局和局部节点得到的表示进行拼接。第二种方法是级联的结构，首先得到一个全局编码器的表示，随后将其作为局部编码器的输入。</p>
<p>这两种方法的每一层都只包含单一的一种编码器，现在提出设想：将两种编码器在层内进行合并，再重复一定的次数从而得到最终表示。层内合并的方法也是类似的，有并行合并和级联合并两种方法，因此总共有四种模型结构，它们如下图所示：</p>
<p><img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210307195339134.png" alt="image-20210307195339134"></p>
<h4 id="Decoder-and-Training"><a href="#Decoder-and-Training" class="headerlink" title="Decoder and Training"></a>Decoder and Training</h4><p>Decoder 用于根据 Encoder 所学习到的图表示来生成对应的文本，这里的 Decoder 使用的是著名论文 <span class="exturl" data-url="aHR0cHM6Ly9hcnhpdi5vcmcvYWJzLzE3MDYuMDM3NjI=">“Attention is all you need”<i class="fa fa-external-link-alt"></i></span> 中的 Decoder 结构。在训练中的其中一个挑战是需要生成包含多行文本的输出，因此在训练过程加入了 length penalty。</p>
<hr>
<h3 id="Data-and-Preprocessing"><a href="#Data-and-Preprocessing" class="headerlink" title="Data and Preprocessing"></a>Data and Preprocessing</h3><h4 id="AGENDA"><a href="#AGENDA" class="headerlink" title="AGENDA"></a>AGENDA</h4><p>该数据集包含了12个顶级 AI 会议的论文摘要，每个样本都包含了论文的标题、论文摘要和其对应的知识图谱。在预处理中同时将标题中的每个 token 当作一个 node，和知识图谱中的所有 node 合并构成图。</p>
<h4 id="WebNLG"><a href="#WebNLG" class="headerlink" title="WebNLG"></a>WebNLG</h4><p>该数据集包含了从 DBPedia 抽取出来的知识图谱，该数据集中包含了较多的边数目，为了防止维度爆炸，作者使用了正则化来定义模型的关系权重。此外，作者还用到了 Levi Transformation，将关系边转化为一个结点，将关系对应的两个节点与其相连。</p>
<p>下图是这两个数据集的概览：</p>
<p><img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210309210539790.png" alt="image-20210309210539790"></p>
<hr>
<h3 id="Experiments"><a href="#Experiments" class="headerlink" title="Experiments"></a>Experiments</h3><p>项目使用 PyTorch Geometric 和 OpenNMT-py 进行实验，Adam 优化器的参数选取分别为 $β_1 = 0.1$ 和 $β_2 =0.2$ ，学习率随着训练的轮次而逐渐上升。作者采用了 byte pair encoding 的方法，从而使实体词汇变成更小的 sub-words。</p>
<p>在评估方面，使用的 metrics 有：BLEU、METEOR、CHRF++ 等。在层次结构模型中，层次的选择从 {2，4，6} 中选择，在普通的并行和级联模型中，全局编码层和局部编码层的层数分别从 {2，4，6} 和 {1，2，3} 中选择，隐藏编码器的维度从 {256，384，448} 中选择。</p>
<h4 id="Results-on-AGENDA"><a href="#Results-on-AGENDA" class="headerlink" title="Results on AGENDA"></a>Results on AGENDA</h4><p><img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310115529028.png" alt="image-20210310115529028"></p>
<h4 id="Results-on-WebNLG"><a href="#Results-on-WebNLG" class="headerlink" title="Results on WebNLG"></a>Results on WebNLG</h4><p>从 AGENDA 数据集的实验可以看出 CGE 的模型效果更好，因此在 WebNLG 实验中没有再用到 PGE。</p>
<p><img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310115637199.png" alt="image-20210310115637199"></p>
<h4 id="Development-Experiments"><a href="#Development-Experiments" class="headerlink" title="Development Experiments"></a>Development Experiments</h4><p>经过实验，作者发现了以下几个结论：</p>
<ul>
<li>编码器的层数越多，效果越好</li>
<li>编码器的向量维度并不是越高越好，因不同模型而异</li>
<li>对于一些模型来说，进一步增加编码器的层数和参数数量能进一步提升模型性能</li>
</ul>
<p><img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310155325418.png" alt="image-20210310155325418"></p>
<h4 id="Ablation-Study"><a href="#Ablation-Study" class="headerlink" title="Ablation Study"></a>Ablation Study</h4><p>在这一部分中，作者将模型中的某些部分消去，从而查看去除这些部分的模型结果是怎样的。具体的结果如下图所示：</p>
<p><img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310161944010.png" alt="image-20210310161944010"></p>
<h4 id="Impact-of-the-Graph-Structure-and-Ouput-Length"><a href="#Impact-of-the-Graph-Structure-and-Ouput-Length" class="headerlink" title="Impact of the Graph Structure and Ouput Length"></a>Impact of the Graph Structure and Ouput Length</h4><ul>
<li>在 AGENDA 数据集中，图的规模越大，模型的效果越好；而 WebNLG 的结果则恰恰相反，规模越大，效果反而越差。</li>
<li>当生成的句子越长的时候，模型的效果会下降，尽管加入了 length penalty，生成得到的文本长度与原文本仍有一定的差距，这也是作者后续所要优化的方面。</li>
</ul>
<h4 id="Human-Evaluation"><a href="#Human-Evaluation" class="headerlink" title="Human Evaluation"></a>Human Evaluation</h4><p>为了进一步评估生成文本的质量，作者聘请了人类对 baseline、本文提出的模型所生成的文本和原来的文本进行评估，评估的两个方面为文本的适当性和流畅性，结果如下图所示，并得出了以下的结论：</p>
<ul>
<li>本文提出的模型的人工评测效果在各类的样本中都比 baseline 要好。</li>
<li>当图的规模增大时，模型的评测效果变差。</li>
</ul>
<p>下图是其中一个 KG 和其对应的三种文本：</p>
<p><img src="https://charfole-blog.oss-cn-shenzhen.aliyuncs.com/image/image-20210310205641461.png" alt="image-20210310205641461"></p>
<h4 id="Additiional-Experiments"><a href="#Additiional-Experiments" class="headerlink" title="Additiional Experiments"></a>Additiional Experiments</h4><p>在这一部分，作者对模型的一些其他部分进行进一步的探究，探究所得出的结论如下所示：</p>
<ul>
<li><strong>Sharing vocabulary</strong> 策略对<strong>小数据集</strong>更为重要，而 <strong>length penalty</strong> 则对<strong>大数据集</strong>更为重要。</li>
<li>全局编码器中注意力权重最大的点往往是较远的结点，而局部编码器则是较近的结点，这和认知是相符的。</li>
</ul>
<hr>
<h3 id="Conclusion"><a href="#Conclusion" class="headerlink" title="Conclusion"></a>Conclusion</h3><p>在本论文中，作者提出了一个结合注意力机制的 KG-to-Text 神经网络结构，同时在模型中结合了全局编码器和局部编码器来改善文本的生成。经过实验分析，在级联结构下构建层次模型结合两类编码器，可以达到 state-of-the-art。同时，作者提出了以下几点未来的研究方向：</p>
<ol>
<li>进一步研究全局和局部编码器的组合方式来提升效果。</li>
<li>尝试在图中结合预训练的上下文词嵌入。</li>
<li>继续探究长文本生成和原文本差距较大的原因。</li>
</ol>
<h3 id="Reproduction"><a href="#Reproduction" class="headerlink" title="Reproduction"></a>Reproduction</h3><p>论文作者已提供了该研究的<span class="exturl" data-url="aHR0cHM6Ly9naXRodWIuY29tL1VLUExhYi9rZzJ0ZXh0">代码<i class="fa fa-external-link-alt"></i></span>，但仓库中依然缺少数据集。</p>
<p>本人基于作者的指引，下载了AGENDA数据集进行试验与debug，使其能够在<span class="exturl" data-url="aHR0cHM6Ly93d3cua2FnZ2xlLmNvbS9jaGFyZm9sZS9rZzJ0ZXh0LW5vdGVib29r">Kaggle平台<i class="fa fa-external-link-alt"></i></span>上成功运行。</p>
</div>
<footer class="post-footer">
<div class="post-tags">
<a href="/tags/Paper/" rel="tag"># Paper</a>
<a href="/tags/Notes/" rel="tag"># Notes</a>
</div>
<div class="post-nav">
<div class="post-nav-item">
<a href="/terminal-proxy.html" rel="prev" title="各类终端走代理的设置方法">
<i class="fa fa-chevron-left"></i> 各类终端走代理的设置方法
</a></div>
<div class="post-nav-item">
<a href="/lanqiao-review.html" rel="next" title="蓝桥杯踩坑题">
蓝桥杯踩坑题 <i class="fa fa-chevron-right"></i>
</a></div>
</div>
</footer>
</article>
</div>
<div class="comments" id="valine-comments"></div>
<script>
window.addEventListener('tabs:register', () => {
let { activeClass } = CONFIG.comments;
if (CONFIG.comments.storage) {
activeClass = localStorage.getItem('comments_active') || activeClass;
}
if (activeClass) {
let activeTab = document.querySelector(`a[href="#comment-${activeClass}"]`);
if (activeTab) {
activeTab.click();
}
}
});
if (CONFIG.comments.storage) {
window.addEventListener('tabs:click', event => {
if (!event.target.matches('.tabs-comment .tab-content .tab-pane')) return;
let commentClass = event.target.classList[1];
localStorage.setItem('comments_active', commentClass);
});
}
</script>
</div>
<div class="toggle sidebar-toggle">
<span class="toggle-line toggle-line-first"></span>
<span class="toggle-line toggle-line-middle"></span>
<span class="toggle-line toggle-line-last"></span>
</div>
<aside class="sidebar">
<div class="sidebar-inner">
<ul class="sidebar-nav motion-element">
<li class="sidebar-nav-toc">
文章目录
</li>
<li class="sidebar-nav-overview">
站点概览
</li>
</ul>
<!--noindex-->
<div class="post-toc-wrap sidebar-panel">
<div class="post-toc motion-element"><ol class="nav"><li class="nav-item nav-level-3"><a class="nav-link" href="#Paper-Information"><span class="nav-number">1.</span> <span class="nav-text">Paper Information</span></a></li><li class="nav-item nav-level-3"><a class="nav-link" href="#Notes"><span class="nav-number">2.</span> <span class="nav-text">Notes</span></a></li><li class="nav-item nav-level-3"><a class="nav-link" href="#Abstract"><span class="nav-number">3.</span> <span class="nav-text">Abstract</span></a></li><li class="nav-item nav-level-3"><a class="nav-link" href="#Introduction"><span class="nav-number">4.</span> <span class="nav-text">Introduction</span></a></li><li class="nav-item nav-level-3"><a class="nav-link" href="#Related-Work"><span class="nav-number">5.</span> <span class="nav-text">Related Work</span></a></li><li class="nav-item nav-level-3"><a class="nav-link" href="#Graph-to-Text-Model"><span class="nav-number">6.</span> <span class="nav-text">Graph-to-Text Model</span></a><ol class="nav-child"><li class="nav-item nav-level-4"><a class="nav-link" href="#Graph-Preparation"><span class="nav-number">6.1.</span> <span class="nav-text">Graph Preparation</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Graph-Neural-Networks-GNN"><span class="nav-number">6.2.</span> <span class="nav-text">Graph Neural Networks (GNN)</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Global-Graph-Encoder"><span class="nav-number">6.3.</span> <span class="nav-text">Global Graph Encoder</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Local-Graph-Encoder"><span class="nav-number">6.4.</span> <span class="nav-text">Local Graph Encoder</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Combining-Global-and-Local-Encodings"><span class="nav-number">6.5.</span> <span class="nav-text">Combining Global and Local Encodings</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Decoder-and-Training"><span class="nav-number">6.6.</span> <span class="nav-text">Decoder and Training</span></a></li></ol></li><li class="nav-item nav-level-3"><a class="nav-link" href="#Data-and-Preprocessing"><span class="nav-number">7.</span> <span class="nav-text">Data and Preprocessing</span></a><ol class="nav-child"><li class="nav-item nav-level-4"><a class="nav-link" href="#AGENDA"><span class="nav-number">7.1.</span> <span class="nav-text">AGENDA</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#WebNLG"><span class="nav-number">7.2.</span> <span class="nav-text">WebNLG</span></a></li></ol></li><li class="nav-item nav-level-3"><a class="nav-link" href="#Experiments"><span class="nav-number">8.</span> <span class="nav-text">Experiments</span></a><ol class="nav-child"><li class="nav-item nav-level-4"><a class="nav-link" href="#Results-on-AGENDA"><span class="nav-number">8.1.</span> <span class="nav-text">Results on AGENDA</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Results-on-WebNLG"><span class="nav-number">8.2.</span> <span class="nav-text">Results on WebNLG</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Development-Experiments"><span class="nav-number">8.3.</span> <span class="nav-text">Development Experiments</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Ablation-Study"><span class="nav-number">8.4.</span> <span class="nav-text">Ablation Study</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Impact-of-the-Graph-Structure-and-Ouput-Length"><span class="nav-number">8.5.</span> <span class="nav-text">Impact of the Graph Structure and Ouput Length</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Human-Evaluation"><span class="nav-number">8.6.</span> <span class="nav-text">Human Evaluation</span></a></li><li class="nav-item nav-level-4"><a class="nav-link" href="#Additiional-Experiments"><span class="nav-number">8.7.</span> <span class="nav-text">Additiional Experiments</span></a></li></ol></li><li class="nav-item nav-level-3"><a class="nav-link" href="#Conclusion"><span class="nav-number">9.</span> <span class="nav-text">Conclusion</span></a></li><li class="nav-item nav-level-3"><a class="nav-link" href="#Reproduction"><span class="nav-number">10.</span> <span class="nav-text">Reproduction</span></a></li></ol></div>
</div>
<!--/noindex-->
<div class="site-overview-wrap sidebar-panel">
<div class="site-author motion-element" itemprop="author" itemscope itemtype="http://schema.org/Person">
<p class="site-author-name" itemprop="name">Charfole</p>
<div class="site-description" itemprop="description">Carpe diem</div>
</div>
<div class="site-state-wrap motion-element">
<nav class="site-state">
<div class="site-state-item site-state-posts">
<a href="/archives/">
<span class="site-state-item-count">14</span>
<span class="site-state-item-name">日志</span>
</a>
</div>
<div class="site-state-item site-state-categories">
<a href="/categories/">
<span class="site-state-item-count">8</span>
<span class="site-state-item-name">分类</span></a>
</div>
<div class="site-state-item site-state-tags">
<a href="/tags/">
<span class="site-state-item-count">29</span>
<span class="site-state-item-name">标签</span></a>
</div>
</nav>
</div>
<div class="links-of-author motion-element">
<span class="links-of-author-item">
<span class="exturl" data-url="aHR0cHM6Ly9naXRodWIuY29tL2NoYXJmb2xl" title="GitHub → https://github.com/charfole"><i class="fab fa-github fa-fw"></i>GitHub</span>
</span>
<span class="links-of-author-item">
<span class="exturl" data-url="bWFpbHRvOmNoYXJmb2xlQDE2My5jb20=" title="E-Mail → mailto:[email protected]"><i class="fa fa-envelope fa-fw"></i>E-Mail</span>
</span>
<span class="links-of-author-item">
<a href="/atom.xml" title="RSS → /atom.xml"><i class="fa fa-rss fa-fw"></i>RSS</a>
</span>
</div>
<div class="cc-license motion-element" itemprop="license">
<span class="exturl cc-opacity" data-url="aHR0cHM6Ly9jcmVhdGl2ZWNvbW1vbnMub3JnL2xpY2Vuc2VzL2J5LW5jLXNhLzQuMC8="><img src="/images/cc-by-nc-sa.svg" alt="Creative Commons"></span>
</div>
</div>
<div class="back-to-top motion-element">
<i class="fa fa-arrow-up"></i>
<span>0%</span>
</div>
</div>
</aside>
<div id="sidebar-dimmer"></div>
</div>
</main>
<footer class="footer">
<div class="footer-inner">
<div class="copyright">
© 2020 –
<span itemprop="copyrightYear">2023</span>
<span class="with-love">
<i class="fa fa-heart"></i>
</span>
<span class="author" itemprop="copyrightHolder">Charfole</span>
<span class="post-meta-divider">|</span>
<span class="post-meta-item-icon">
<i class="fa fa-chart-area"></i>
</span>
<span class="post-meta-item-text">站点总字数：</span>
<span title="站点总字数">61k</span>
<span class="post-meta-divider">|</span>
<span class="post-meta-item-icon">
<i class="fa fa-coffee"></i>
</span>
<span class="post-meta-item-text">站点阅读时长 ≈</span>
<span title="站点阅读时长">55 分钟</span>
</div>
<div class="powered-by">由 <span class="exturl theme-link" data-url="aHR0cHM6Ly9oZXhvLmlv">Hexo</span> & <span class="exturl theme-link" data-url="aHR0cHM6Ly90aGVtZS1uZXh0Lm9yZw==">NexT.Gemini</span> 强力驱动
</div>
<div class="busuanzi-count">
<script async src="https://busuanzi.ibruce.info/busuanzi/2.3/busuanzi.pure.mini.js"></script>
<span class="post-meta-item" id="busuanzi_container_site_uv" style="display: none;">
<span class="post-meta-item-icon">
<i class="fa fa-user"></i>
</span>
<span class="site-uv" title="总访客量">
<span id="busuanzi_value_site_uv"></span>
</span>
</span>
<span class="post-meta-divider">|</span>
<span class="post-meta-item" id="busuanzi_container_site_pv" style="display: none;">
<span class="post-meta-item-icon">
<i class="fa fa-eye"></i>
</span>
<span class="site-pv" title="总访问量">
<span id="busuanzi_value_site_pv"></span>
</span>
</span>
</div>
</div>
</footer>
</div>
<script src="/lib/anime.min.js"></script>
<script src="/lib/velocity/velocity.min.js"></script>
<script src="/lib/velocity/velocity.ui.min.js"></script>
<script src="/js/utils.js"></script>
<script src="/js/motion.js"></script>
<script src="/js/schemes/pisces.js"></script>
<script src="/js/next-boot.js"></script>
<script>
(function(){
var canonicalURL, curProtocol;
//Get the <link> tag
var x=document.getElementsByTagName("link");
//Find the last canonical URL
if(x.length > 0){
for (i=0;i<x.length;i++){
if(x[i].rel.toLowerCase() == 'canonical' && x[i].href){
canonicalURL=x[i].href;
}
}
}
//Get protocol
if (!canonicalURL){
curProtocol = window.location.protocol.split(':')[0];
}
else{
curProtocol = canonicalURL.split(':')[0];
}
//Get current URL if the canonical URL does not exist
if (!canonicalURL) canonicalURL = window.location.href;
//Assign script content. Replace current URL with the canonical URL
!function(){var e=/([http|https]:\/\/[a-zA-Z0-9\_\.]+\.baidu\.com)/gi,r=canonicalURL,t=document.referrer;if(!e.test(r)){var n=(String(curProtocol).toLowerCase() === 'https')?"https://sp0.baidu.com/9_Q4simg2RQJ8t7jm9iCKT-xh_/s.gif":"//api.share.baidu.com/s.gif";t?(n+="?r="+encodeURIComponent(document.referrer),r&&(n+="&l="+r)):r&&(n+="?l="+r);var i=new Image;i.src=n}}(window);})();
</script>
<script src="/js/local-search.js"></script>
<script>
if (typeof MathJax === 'undefined') {
window.MathJax = {
loader: {
source: {
'[tex]/amsCd': '[tex]/amscd',
'[tex]/AMScd': '[tex]/amscd'
}
},
tex: {
inlineMath: {'[+]': [['$', '$']]},
tags: 'ams'
},
options: {
renderActions: {
findScript: [10, doc => {
document.querySelectorAll('script[type^="math/tex"]').forEach(node => {
const display = !!node.type.match(/; *mode=display/);
const math = new doc.options.MathItem(node.textContent, doc.inputJax[0], display);
const text = document.createTextNode('');
node.parentNode.replaceChild(text, node);
math.start = {node: text, delim: '', n: 0};
math.end = {node: text, delim: '', n: 0};
doc.math.push(math);
});
}, '', false],
insertedScript: [200, () => {
document.querySelectorAll('mjx-container').forEach(node => {
let target = node.parentNode;
if (target.nodeName.toLowerCase() === 'li') {
target.parentNode.classList.add('has-jax');
}
});
}, '', false]
}
}
};
(function () {
var script = document.createElement('script');
script.src = '//cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js';
script.defer = true;
document.head.appendChild(script);
})();
} else {
MathJax.startup.document.state(0);
MathJax.texReset();
MathJax.typeset();
}
</script>
<script>
NexT.utils.loadComments(document.querySelector('#valine-comments'), () => {
NexT.utils.getScript('//unpkg.com/valine/dist/Valine.min.js', () => {
var GUEST = ['nick', 'mail', 'link'];
var guest = 'nick,mail,link';
guest = guest.split(',').filter(item => {
return GUEST.includes(item);
});
new Valine({
el : '#valine-comments',
verify : false,
notify : false,
appId : 'hXI9oKO1Olitf9gLi5jb1ciq-gzGzoHsz',
appKey : 'JrOSBT9CBoGf3OzREeMchSkt',
placeholder: "Just go go",
avatar : 'mm',
meta : guest,
pageSize : '10' || 10,
visitor : false,
lang : '' || 'zh-cn',
path : location.pathname,
recordIP : false,
serverURLs : ''
});
}, window.Valine);
});
</script>
</body>
</html>