-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path9244fc05ea6b.html
581 lines (393 loc) · 36.9 KB
/
9244fc05ea6b.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
<!DOCTYPE html>
<html lang=zh>
<head>
<meta charset="utf-8">
<meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1">
<meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1, minimum-scale=1, user-scalable=no, minimal-ui">
<meta name="renderer" content="webkit">
<meta http-equiv="Cache-Control" content="no-transform" />
<meta http-equiv="Cache-Control" content="no-siteapp" />
<meta name="apple-mobile-web-app-capable" content="yes">
<meta name="apple-mobile-web-app-status-bar-style" content="black">
<meta name="format-detection" content="telephone=no,email=no,adress=no">
<!-- Color theme for statusbar -->
<meta name="theme-color" content="#000000" />
<!-- 强制页面在当前窗口以独立页面显示,防止别人在框架里调用页面 -->
<meta http-equiv="window-target" content="_top" />
<title>如何从零开始构建一个网络讨论帖分类模型？ | chadqiu's blog</title>
<meta name="description" content="Motivation前几天搭建了一个对牛客网每天最新的工作信息进行爬取的程序，见牛客网爬虫，但从网上爬取下来的帖子有很多不是工作信息，需要把这部分干扰信息给排除掉，否则很影响使用心情。之前使用关键词与正则表达式进行了简单过滤，但总是有一些漏网之鱼，且容易误伤，如果能训练一个NLP分类模型来进行过滤，那就再好不过了，正好本人的研究方向是NLP，就想试着构建一个模型玩玩了。 数据准备但一般情况下要训练">
<meta property="og:type" content="article">
<meta property="og:title" content="如何从零开始构建一个网络讨论帖分类模型？">
<meta property="og:url" content="http://chadqiu.github.io/9244fc05ea6b.html">
<meta property="og:site_name" content="chadqiu">
<meta property="og:description" content="Motivation前几天搭建了一个对牛客网每天最新的工作信息进行爬取的程序，见牛客网爬虫，但从网上爬取下来的帖子有很多不是工作信息，需要把这部分干扰信息给排除掉，否则很影响使用心情。之前使用关键词与正则表达式进行了简单过滤，但总是有一些漏网之鱼，且容易误伤，如果能训练一个NLP分类模型来进行过滤，那就再好不过了，正好本人的研究方向是NLP，就想试着构建一个模型玩玩了。 数据准备但一般情况下要训练">
<meta property="og:locale" content="zh_CN">
<meta property="og:image" content="http://chadqiu.github.io/images/newcoder_data.png">
<meta property="og:image" content="http://chadqiu.github.io/images/newcoder_f1.png">
<meta property="article:published_time" content="2023-02-03T00:56:52.000Z">
<meta property="article:modified_time" content="2024-11-26T16:26:22.004Z">
<meta property="article:author" content="John Doe">
<meta property="article:tag" content="huggingface">
<meta property="article:tag" content="python">
<meta property="article:tag" content="AI">
<meta property="article:tag" content="NLP">
<meta name="twitter:card" content="summary">
<meta name="twitter:image" content="http://chadqiu.github.io/images/newcoder_data.png">
<!-- Canonical links -->
<link rel="canonical" href="http://chadqiu.github.io/9244fc05ea6b.html">
<link rel="alternate" href="/atom.xml" title="chadqiu" type="application/atom+xml">
<link rel="icon" href="images/avatar.jpg" type="image/x-icon">
<link rel="stylesheet" href="/css/style.css">
<meta name="generator" content="Hexo 6.3.0"></head>
<body class="main-center theme-green" itemscope itemtype="http://schema.org/WebPage">
<header class="header" itemscope itemtype="http://schema.org/WPHeader">
<div class="slimContent">
<div class="navbar-header">
<div class="profile-block text-center">
<a id="avatar" href="https://github.com/chadqiu" target="_blank">
<img class="img-circle img-rotate" src="/images/logo.jpg" width="200" height="200">
</a>
<h2 id="name" class="hidden-xs hidden-sm">chadqiu</h2>
<h3 id="title" class="hidden-xs hidden-sm hidden-md">Developer & Researcher</h3>
<small id="location" class="text-muted hidden-xs hidden-sm"><i class="icon icon-map-marker"></i> Shanghai, China</small>
</div>
<div class="search" id="search-form-wrap">
<form class="search-form sidebar-form">
<div class="input-group">
<input type="text" class="search-form-input form-control" placeholder="搜索" />
<span class="input-group-btn">
<button type="submit" class="search-form-submit btn btn-flat" onclick="return false;"><i class="icon icon-search"></i></button>
</span>
</div>
</form>
<div class="ins-search">
<div class="ins-search-mask"></div>
<div class="ins-search-container">
<div class="ins-input-wrapper">
<input type="text" class="ins-search-input" placeholder="想要查找什么..." x-webkit-speech />
<button type="button" class="close ins-close ins-selectable" data-dismiss="modal" aria-label="Close"><span aria-hidden="true">×</span></button>
</div>
<div class="ins-section-wrapper">
<div class="ins-section-container"></div>
</div>
</div>
</div>
</div>
<button class="navbar-toggle collapsed" type="button" data-toggle="collapse" data-target="#main-navbar" aria-controls="main-navbar" aria-expanded="false">
<span class="sr-only">Toggle navigation</span>
<span class="icon-bar"></span>
<span class="icon-bar"></span>
<span class="icon-bar"></span>
</button>
</div>
<nav id="main-navbar" class="collapse navbar-collapse" itemscope itemtype="http://schema.org/SiteNavigationElement" role="navigation">
<ul class="nav navbar-nav main-nav ">
<li class="menu-item menu-item-home">
<a href="/.">
<i class="icon icon-home-fill"></i>
<span class="menu-title">首页</span>
</a>
</li>
<li class="menu-item menu-item-archives">
<a href="/archives">
<i class="icon icon-archives-fill"></i>
<span class="menu-title">归档</span>
</a>
</li>
<li class="menu-item menu-item-categories">
<a href="/categories">
<i class="icon icon-folder"></i>
<span class="menu-title">分类</span>
</a>
</li>
<li class="menu-item menu-item-tags">
<a href="/tags">
<i class="icon icon-tags"></i>
<span class="menu-title">标签</span>
</a>
</li>
<li class="menu-item menu-item-repository">
<a href="/repository">
<i class="icon icon-project"></i>
<span class="menu-title">项目</span>
</a>
</li>
<li class="menu-item menu-item-links">
<a href="/links">
<i class="icon icon-friendship"></i>
<span class="menu-title">友链</span>
</a>
</li>
<li class="menu-item menu-item-about">
<a href="/about">
<i class="icon icon-cup-fill"></i>
<span class="menu-title">关于</span>
</a>
</li>
</ul>
<ul class="social-links">
<li><a href="https://github.com/chadqiu" target="_blank" title="Github" data-toggle=tooltip data-placement=top><i class="icon icon-github"></i></a></li>
<li><a href="/atom.xml" target="_blank" title="Rss" data-toggle=tooltip data-placement=top><i class="icon icon-rss"></i></a></li>
</ul>
</nav>
</div>
</header>
<aside class="sidebar" itemscope itemtype="http://schema.org/WPSideBar">
<div class="slimContent">
<div class="widget">
<h3 class="widget-title">公告</h3>
<div class="widget-body">
<div id="board">
<div class="content">
<p>欢迎交流与分享经验!</p>
</div>
</div>
</div>
</div>
<div class="widget">
<h3 class="widget-title">分类</h3>
<div class="widget-body">
<ul class="category-list"><li class="category-list-item"><a class="category-list-link" href="/categories/AI/">AI</a><span class="category-list-count">3</span></li><li class="category-list-item"><a class="category-list-link" href="/categories/tools-%E7%88%AC%E8%99%AB/">tools - 爬虫</a><span class="category-list-count">3</span></li><li class="category-list-item"><a class="category-list-link" href="/categories/%E8%AE%A1%E7%AE%97%E6%9C%BA%E5%9F%BA%E7%A1%80-%E6%95%B0%E6%8D%AE%E7%BB%93%E6%9E%84%E4%B8%8E%E7%AE%97%E6%B3%95/">计算机基础 - 数据结构与算法</a><span class="category-list-count">1</span></li></ul>
</div>
</div>
<div class="widget">
<h3 class="widget-title">标签</h3>
<div class="widget-body">
<ul class="tag-list" itemprop="keywords"><li class="tag-list-item"><a class="tag-list-link" href="/tags/AI/" rel="tag">AI</a><span class="tag-list-count">2</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/Java/" rel="tag">Java</a><span class="tag-list-count">1</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/NLP/" rel="tag">NLP</a><span class="tag-list-count">2</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/dataset/" rel="tag">dataset</a><span class="tag-list-count">1</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/huggingface/" rel="tag">huggingface</a><span class="tag-list-count">3</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/python/" rel="tag">python</a><span class="tag-list-count">6</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/train/" rel="tag">train</a><span class="tag-list-count">1</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/%E4%BA%8C%E5%88%86%E6%90%9C%E7%B4%A2/" rel="tag">二分搜索</a><span class="tag-list-count">1</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/%E7%88%AC%E8%99%AB/" rel="tag">爬虫</a><span class="tag-list-count">3</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/%E7%89%9B%E5%AE%A2%E7%BD%91/" rel="tag">牛客网</a><span class="tag-list-count">1</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/%E7%AE%97%E6%B3%95/" rel="tag">算法</a><span class="tag-list-count">1</span></li><li class="tag-list-item"><a class="tag-list-link" href="/tags/%E8%B0%B7%E6%AD%8C%E5%AD%A6%E6%9C%AF/" rel="tag">谷歌学术</a><span class="tag-list-count">1</span></li></ul>
</div>
</div>
<div class="widget">
<h3 class="widget-title">标签云</h3>
<div class="widget-body tagcloud">
<a href="/tags/AI/" style="font-size: 13.33px;">AI</a> <a href="/tags/Java/" style="font-size: 13px;">Java</a> <a href="/tags/NLP/" style="font-size: 13.33px;">NLP</a> <a href="/tags/dataset/" style="font-size: 13px;">dataset</a> <a href="/tags/huggingface/" style="font-size: 13.67px;">huggingface</a> <a href="/tags/python/" style="font-size: 14px;">python</a> <a href="/tags/train/" style="font-size: 13px;">train</a> <a href="/tags/%E4%BA%8C%E5%88%86%E6%90%9C%E7%B4%A2/" style="font-size: 13px;">二分搜索</a> <a href="/tags/%E7%88%AC%E8%99%AB/" style="font-size: 13.67px;">爬虫</a> <a href="/tags/%E7%89%9B%E5%AE%A2%E7%BD%91/" style="font-size: 13px;">牛客网</a> <a href="/tags/%E7%AE%97%E6%B3%95/" style="font-size: 13px;">算法</a> <a href="/tags/%E8%B0%B7%E6%AD%8C%E5%AD%A6%E6%9C%AF/" style="font-size: 13px;">谷歌学术</a>
</div>
</div>
<div class="widget">
<h3 class="widget-title">归档</h3>
<div class="widget-body">
<ul class="archive-list"><li class="archive-list-item"><a class="archive-list-link" href="/archives/2025/02/">二月 2025</a><span class="archive-list-count">1</span></li><li class="archive-list-item"><a class="archive-list-link" href="/archives/2025/01/">一月 2025</a><span class="archive-list-count">5</span></li><li class="archive-list-item"><a class="archive-list-link" href="/archives/2024/12/">十二月 2024</a><span class="archive-list-count">5</span></li><li class="archive-list-item"><a class="archive-list-link" href="/archives/2024/11/">十一月 2024</a><span class="archive-list-count">1</span></li><li class="archive-list-item"><a class="archive-list-link" href="/archives/2023/02/">二月 2023</a><span class="archive-list-count">3</span></li><li class="archive-list-item"><a class="archive-list-link" href="/archives/2023/01/">一月 2023</a><span class="archive-list-count">5</span></li></ul>
</div>
</div>
<div class="widget">
<h3 class="widget-title">最新文章</h3>
<div class="widget-body">
<ul class="recent-post-list list-unstyled no-thumbnail">
<li>
<div class="item-inner">
<p class="item-category">
</p>
<p class="item-title">
<a href="/d8613ad6db3e.html" class="title">Z哥25-02-09直播</a>
</p>
<p class="item-date">
<time datetime="2025-02-10T00:59:45.000Z" itemprop="datePublished">2025-02-10</time>
</p>
</div>
</li>
<li>
<div class="item-inner">
<p class="item-category">
</p>
<p class="item-title">
<a href="/814464067ab6.html" class="title">Z哥25-01-27直播</a>
</p>
<p class="item-date">
<time datetime="2025-01-28T14:24:20.000Z" itemprop="datePublished">2025-01-28</time>
</p>
</div>
</li>
<li>
<div class="item-inner">
<p class="item-category">
</p>
<p class="item-title">
<a href="/49dc4032c5df.html" class="title">Z哥25-01-24直播</a>
</p>
<p class="item-date">
<time datetime="2025-01-25T12:05:29.000Z" itemprop="datePublished">2025-01-25</time>
</p>
</div>
</li>
<li>
<div class="item-inner">
<p class="item-category">
</p>
<p class="item-title">
<a href="/30138261f7a5.html" class="title">Z哥25-01-19直播</a>
</p>
<p class="item-date">
<time datetime="2025-01-22T01:22:26.000Z" itemprop="datePublished">2025-01-22</time>
</p>
</div>
</li>
<li>
<div class="item-inner">
<p class="item-category">
</p>
<p class="item-title">
<a href="/842fcfb8d8d1.html" class="title">Z哥25-01-12直播</a>
</p>
<p class="item-date">
<time datetime="2025-01-13T01:09:32.000Z" itemprop="datePublished">2025-01-13</time>
</p>
</div>
</li>
</ul>
</div>
</div>
</div>
</aside>
<aside class="sidebar sidebar-toc collapse in " id="collapseToc" itemscope itemtype="http://schema.org/WPSideBar">
<div class="slimContent">
<nav id="toc" class="article-toc">
<h3 class="toc-title">文章目录</h3>
<ol class="toc"><li class="toc-item toc-level-2"><a class="toc-link" href="#Motivation"><span class="toc-number">1.</span> <span class="toc-text">Motivation</span></a></li><li class="toc-item toc-level-2"><a class="toc-link" href="#%E6%95%B0%E6%8D%AE%E5%87%86%E5%A4%87"><span class="toc-number">2.</span> <span class="toc-text">数据准备</span></a></li><li class="toc-item toc-level-2"><a class="toc-link" href="#%E6%A8%A1%E5%9E%8B%E4%B8%8E%E8%AE%AD%E7%BB%83"><span class="toc-number">3.</span> <span class="toc-text">模型与训练</span></a></li><li class="toc-item toc-level-2"><a class="toc-link" href="#%E9%A2%84%E6%B5%8B%E8%BF%87%E6%BB%A4"><span class="toc-number">4.</span> <span class="toc-text">预测过滤</span></a></li></ol>
</nav>
</div>
</aside>
<main class="main" role="main">
<div class="content">
<article id="post-如何从零开始构建一个网络讨论帖分类模型" class="article article-type-post" itemscope itemtype="http://schema.org/BlogPosting">
<div class="article-header">
<h1 class="article-title" itemprop="name">
如何从零开始构建一个网络讨论帖分类模型？
</h1>
<div class="article-meta">
<span class="article-date">
<i class="icon icon-calendar-check"></i>
<a href="/9244fc05ea6b.html" class="article-date">
<time datetime="2023-02-03T00:56:52.000Z" itemprop="datePublished">2023-02-03</time>
</a>
</span>
<span class="article-category">
<i class="icon icon-folder"></i>
<a class="article-category-link" href="/categories/AI/">AI</a>
</span>
<span class="article-tag">
<i class="icon icon-tags"></i>
<a class="article-tag-link-link" href="/tags/AI/" rel="tag">AI</a>, <a class="article-tag-link-link" href="/tags/NLP/" rel="tag">NLP</a>, <a class="article-tag-link-link" href="/tags/huggingface/" rel="tag">huggingface</a>, <a class="article-tag-link-link" href="/tags/python/" rel="tag">python</a>
</span>
<span class="post-comment"><i class="icon icon-comment"></i> <a href="/9244fc05ea6b.html#comments" class="article-comment-link">评论</a></span>
</div>
</div>
<div class="article-entry marked-body" itemprop="articleBody">
<h2 id="Motivation"><a href="#Motivation" class="headerlink" title="Motivation"></a>Motivation</h2><p>前几天搭建了一个对牛客网每天最新的工作信息进行爬取的程序，见<a href="https://chadqiu.github.io/f06a19b2ce94.html">牛客网爬虫</a>，但从网上爬取下来的帖子有很多不是工作信息，需要把这部分干扰信息给排除掉，否则很影响使用心情。之前使用关键词与正则表达式进行了简单过滤，但总是有一些漏网之鱼，且容易误伤，如果能训练一个NLP分类模型来进行过滤，那就再好不过了，正好本人的研究方向是NLP，就想试着构建一个模型玩玩了。</p>
<h2 id="数据准备"><a href="#数据准备" class="headerlink" title="数据准备"></a>数据准备</h2><p>但一般情况下要训练一个NLP模型需要几千到几万条有标注好的数据，而本项目没有现成的数据，这也是构建模型最困难的地方了。通过爬虫，获取了4万条左右的历史数据，包含id、用户昵称、标题、正文等内容，如下图所示，但没有标签。通过观察，可以把这些帖子大致分成 【招聘信息、经验贴、求助帖】三类，接下来就该考虑如何进行标注了。<br><img src="/images/newcoder_data.png" alt="牛客帖子数据"></p>
<p>人工标注太费时费力了，而且非常的不优雅，我们还是希望找到一个自动标注的方法，这里首先想到的就是最近两年在学术界比较火的few-shot、zero-shot技术了，且一般模型越大，效果越好。目前能访问到的大模型有： <a target="_blank" rel="noopener" href="https://openai.com/">openAI</a>的GPT3及最近大火的chatGPT，<a target="_blank" rel="noopener" href="https://wenxin.baidu.com/ernie3">百度文心</a>的 ERNIE 3.0大模型，已经一些机构开源在<a target="_blank" rel="noopener" href="https://huggingface.co/models">huggingface</a> 和 <a target="_blank" rel="noopener" href="https://modelscope.cn/studios">魔搭社区</a>的大模型，我使用prompt进行了一轮zero-shot尝试。<br>prompt格式示例如下：</p>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br></pre></td><td class="code"><pre><span class="line"></span><br><span class="line">请问下面文本属于 招聘信息、 经验贴、 求助贴 三者中的哪一类？</span><br><span class="line">秋招大结局（泪目了）。家人们泪目了，一波三折之后获得的小奖状，已经准备春招了，没想到被捞啦，嗐，总之是有个结果，还是很开心的[掉小珍珠了][掉小珍珠了]</span><br><span class="line"></span><br><span class="line">请问下面文本属于哪一类帖子？</span><br><span class="line">秋招大结局（泪目了）。家人们泪目了，一波三折之后获得的小奖状，已经准备春招了，没想到被捞啦，嗐，总之是有个结果，还是很开心的[掉小珍珠了][掉小珍珠了]</span><br><span class="line">选项：招聘信息, 经验贴, 求助贴</span><br><span class="line">答案：</span><br></pre></td></tr></table></figure>
<p>经过一轮测试，发现他们的效果如下： chatGPT > 百度文心 >> others<br>chatGPT表现较好，绝大本分都预测的比较准确，百度文心也基本可用，大部分都能答正确，之后就准备使用API来调用这两个大模型来标数据了，但百度文心每天只能访问200次，我很快超出次数限制，现阶段还不能直接付费购买服务，只能填合作申请表，然后等待。<br>chatGPT不对中国用户开放，无法直接注册账户，通过特殊方法也是可以注册上的。前段时间翻墙后还能正常访问chatGPT的页面，但现在访问不了了，API在国内也访问不了，但可以采用“东数西算”的思想，把数据拿到国外的服务器上计算就行了，最简单的方法就是使用google的colab，免费创建一个notebook，并把数据传到google drive 或 GitHub，然后访问openAI的api。调用api需要先到<a target="_blank" rel="noopener" href="https://platform.openai.com/account/api-keys">官网</a>上申请一个API key，然后再调用，使用pyhton调用API的代码如下：</p>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br></pre></td><td class="code"><pre><span class="line"></span><br><span class="line"><span class="keyword">import</span> openai</span><br><span class="line">openai.api_key = <span class="string">"your api key"</span></span><br><span class="line"></span><br><span class="line">s = <span class="string">'''请问下面文本属于哪一类帖子？</span></span><br><span class="line"><span class="string">viv0社招。 #春招# 有匹配岗位 有意向大佬欢迎＋微g1r4ffe内推 ...viv0社招开启，岗位多多hc多多。博士应聘专家岗位有1年以上工作经验即可 #社招#</span></span><br><span class="line"><span class="string">选项：招聘信息, 经验贴, 求助贴</span></span><br><span class="line"><span class="string">答案：'''</span></span><br><span class="line"></span><br><span class="line">rst = openai.Completion.create(</span><br><span class="line"> model=<span class="string">"text-davinci-003"</span>, </span><br><span class="line"> prompt= s,</span><br><span class="line"> max_tokens=<span class="number">15</span>,</span><br><span class="line"> temperature=<span class="number">0</span></span><br><span class="line">)</span><br><span class="line"></span><br><span class="line"><span class="built_in">print</span>(rst[<span class="string">'choices'</span>][<span class="number">0</span>][<span class="string">"text"</span>])</span><br><span class="line"></span><br><span class="line"><span class="comment"># output: 招聘信息</span></span><br></pre></td></tr></table></figure>
<p>直接进去还没有chatGPT的API，但有 text-davinci-003 这一强大的模型，它基于GPT3大模型，使用了跟chatGPT相似的instruction训练，亲测效果很好，跟chatGPT差不多，甚至可以说就是chatGPT了。最终，用API标注了500条左右的数据，然后又人工标注了100条数据作为测试集。</p>
<h2 id="模型与训练"><a href="#模型与训练" class="headerlink" title="模型与训练"></a>模型与训练</h2><p>训练的基本策略为使用伪标签技术，即先使用少量数据训练一个模型，让这个模型去标数据，然后用其标注的数据集进行训练，最后结果往往会超过原来那个标注的模型。<br>由于500条数据仍然很小，属于few-shot的范围了，因此希望使用尽量大的模型，一般模型越大，表现往往越好，大模型的few-shot能力也强，我在AutoDL上租了个24GB显存的A5000GPU，最大也就能训练1.3B大小的模型，但经过一系列实验后发现，居然是roberta-large表现最好，在我那个100数据的小测试集上F1 score超过了90%，然后用它对剩下的3万多条数据进行预测，生成标注数据集，最后使用该数据集训练一个新模型。<br>由于后期要在cpu上运行，因此希望使用尽量小的模型，这里选择了腾讯的 uer/chinese_roberta_L-4_H-512 模型进行训练，训练结果出人意料的好(也许是测试集太小，不准确)，如下图所示：<br><img src="/images/newcoder_f1.png"></p>
<p>训练完成后的模型在roberta4h512文件夹中，可通过huggingface本地读取，读取示例如下：</p>
<figure class="highlight plaintext"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br></pre></td><td class="code"><pre><span class="line">from transformers import AutoTokenizer, AutoModelForSequenceClassification</span><br><span class="line">model_name = "roberta4h512"</span><br><span class="line">model = AutoModelForSequenceClassification.from_pretrained(model_name)</span><br><span class="line">tokenizer = AutoTokenizer.from_pretrained(model_name)</span><br><span class="line"></span><br></pre></td></tr></table></figure>
<p>模型训练代码： <a target="_blank" rel="noopener" href="https://github.com/chadqiu/newcoder-crawler/blob/main/bert_train.py">bert_train.py</a><br>使用训练好的模型进行伪标签数据生成的代码：<a target="_blank" rel="noopener" href="https://github.com/chadqiu/newcoder-crawler/blob/main/predict.py">predict.py</a><br>模型训练细节见 <a href="https://chadqiu.github.io/e819d4a7ec80.html">如何使用huggingface的trainer训练模型？</a></p>
<h2 id="预测过滤"><a href="#预测过滤" class="headerlink" title="预测过滤"></a>预测过滤</h2><p>我们把爬回来的帖子中预测为招聘信息的帖子留下来，其他的过滤掉即可。爬虫程序一天执行一次，可以采用类似懒加载的方式加载模型，为了性能，需要分batch进行计算, 实测在cpu下183条数据需要6.5s左右，平均每条数据推理时间在36ms左右。预测代码如下：</p>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br></pre></td><td class="code"><pre><span class="line"><span class="keyword">from</span> transformers <span class="keyword">import</span> AutoTokenizer, AutoModelForSequenceClassification</span><br><span class="line"></span><br><span class="line"><span class="keyword">def</span> <span class="title function_">_batch_generate</span>(<span class="params">texts, model, tokenizer, id2label = {<span class="number">0</span>: <span class="string">'招聘信息'</span>, <span class="number">1</span>: <span class="string">'经验贴'</span>, <span class="number">2</span>: <span class="string">'求助贴'</span>}, max_length = <span class="number">128</span></span>):</span><br><span class="line"> inputs = tokenizer( texts, return_tensors=<span class="string">"pt"</span>, max_length=<span class="number">128</span>, padding=<span class="literal">True</span>, truncation=<span class="literal">True</span>)</span><br><span class="line"> outputs = model(**inputs).logits.argmax(-<span class="number">1</span>).tolist()</span><br><span class="line"> <span class="keyword">return</span> [id2label[x] <span class="keyword">for</span> x <span class="keyword">in</span> outputs]</span><br><span class="line"></span><br><span class="line"><span class="keyword">def</span> <span class="title function_">model_predict</span>(<span class="params">text_list, model = <span class="literal">None</span>, tokenizer = <span class="literal">None</span>, model_name = <span class="string">"roberta4h512"</span>, batch_size = <span class="number">4</span></span>):</span><br><span class="line"> <span class="keyword">if</span> <span class="keyword">not</span> text_list: <span class="keyword">return</span> []</span><br><span class="line"> <span class="keyword">if</span> <span class="keyword">not</span> model:</span><br><span class="line"> model = AutoModelForSequenceClassification.from_pretrained(model_name)</span><br><span class="line"> <span class="keyword">if</span> <span class="keyword">not</span> tokenizer:</span><br><span class="line"> tokenizer = AutoTokenizer.from_pretrained(model_name)</span><br><span class="line"> model.<span class="built_in">eval</span>()</span><br><span class="line"> result, start = [], <span class="number">0</span></span><br><span class="line"> <span class="keyword">while</span>(start < <span class="built_in">len</span>(text_list)):</span><br><span class="line"> result.extend(_batch_generate(text_list[start : start + batch_size], model, tokenizer))</span><br><span class="line"> start += batch_size</span><br><span class="line"> <span class="keyword">return</span> result</span><br><span class="line"></span><br></pre></td></tr></table></figure>
<p>使用示例如下：</p>
<figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br></pre></td><td class="code"><pre><span class="line"></span><br><span class="line">ss = [</span><br><span class="line"> <span class="string">'秋招大结局（泪目了）。家人们泪目了，一波三折之后获得的小奖状，已经准备春招了，没想到被捞啦，嗐，总之是有个结果，还是很开心的[掉小珍珠了][掉小珍珠了]'</span>,</span><br><span class="line"> <span class="string">'找到工作之后还要继续找吗。5k 加班严重 春招还想继续找 大家有什么好的建议 #我的求职思考# ...双非应届本科 拿了一个广州嵌入式offer 待遇9.'</span></span><br><span class="line">]</span><br><span class="line"></span><br><span class="line"><span class="built_in">print</span>(model_predict(ss))</span><br><span class="line"></span><br><span class="line"><span class="comment"># output: ['经验贴', '求助贴']</span></span><br><span class="line"></span><br></pre></td></tr></table></figure>
<p>项目guthub地址：<a target="_blank" rel="noopener" href="https://github.com/chadqiu/newcoder-crawler">https://github.com/chadqiu/newcoder-crawler</a></p>
</div>
<div class="article-footer">
<blockquote class="mt-2x">
<ul class="post-copyright list-unstyled">
<li class="post-copyright-link hidden-xs">
<strong>本文链接：</strong>
<a href="http://chadqiu.github.io/9244fc05ea6b.html" title="如何从零开始构建一个网络讨论帖分类模型？" target="_blank" rel="external">http://chadqiu.github.io/9244fc05ea6b.html</a>
</li>
<li class="post-copyright-license">
<strong>版权声明： </strong> 本博客所有文章除特别声明外，均采用 <a href="http://creativecommons.org/licenses/by/4.0/deed.zh" target="_blank" rel="external">CC BY 4.0 CN协议</a> 许可协议。转载请注明出处！
</li>
</ul>
</blockquote>
<div class="panel panel-default panel-badger">
<div class="panel-body">
<figure class="media">
<div class="media-left">
<a href="https://github.com/chadqiu" target="_blank" class="img-burn thumb-sm visible-lg">
<img src="/images/logo.jpg" class="img-rounded w-full" alt="">
</a>
</div>
<div class="media-body">
<h3 class="media-heading"><a href="https://github.com/chadqiu" target="_blank"><span class="text-dark">chadqiu</span><small class="ml-1x">Developer & Researcher</small></a></h3>
<div>在校程序猿</div>
</div>
</figure>
</div>
</div>
</div>
</article>
<section id="comments">
<div id="lv-container" data-id="city" data-uid="MTAyMC81Nzk4OC8zNDQ1MQ==">
<noscript> 为正常使用来必力评论功能请激活JavaScript</noscript>
</div>
</section>
</div>
<nav class="bar bar-footer clearfix" data-stick-bottom>
<div class="bar-inner">
<ul class="pager pull-left">
<li class="prev">
<a href="/6fcf23854660.html" title="如何构建一个自定义huggingface dataset数据集？"><i class="icon icon-angle-left" aria-hidden="true"></i><span> 上一篇</span></a>
</li>
<li class="next">
<a href="/79bdc4d1f745.html" title="二分搜索"><span>下一篇 </span><i class="icon icon-angle-right" aria-hidden="true"></i></a>
</li>
<li class="toggle-toc">
<a class="toggle-btn " data-toggle="collapse" href="#collapseToc" aria-expanded="false" title="文章目录" role="button"> <span>[ </span><span>文章目录</span>
<i class="text-collapsed icon icon-anchor"></i>
<i class="text-in icon icon-close"></i>
<span>]</span>
</a>
</li>
</ul>
<div class="bar-right">
</div>
</div>
</nav>
</main>
<footer class="footer" itemscope itemtype="http://schema.org/WPFooter">
<ul class="social-links">
<li><a href="https://github.com/chadqiu" target="_blank" title="Github" data-toggle=tooltip data-placement=top><i class="icon icon-github"></i></a></li>
<li><a href="/atom.xml" target="_blank" title="Rss" data-toggle=tooltip data-placement=top><i class="icon icon-rss"></i></a></li>
</ul>
<div class="copyright">
<div class="publishby">
Theme by <a href="https://github.com/cofess" target="_blank"> cofess </a>base on <a href="https://github.com/cofess/hexo-theme-pure" target="_blank">pure</a>.
</div>
</div>
</footer>
<script src="//cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js"></script>
<script>
window.jQuery || document.write('<script src="js/jquery.min.js"><\/script>')
</script>
<script src="/js/plugin.min.js"></script>
<script src="/js/application.js"></script>
<script>
(function (window) {
var INSIGHT_CONFIG = {
TRANSLATION: {
POSTS: '文章',
PAGES: '页面',
CATEGORIES: '分类',
TAGS: '标签',
UNTITLED: '(未命名)',
},
ROOT_URL: '/',
CONTENT_URL: '/content.json',
};
window.INSIGHT_CONFIG = INSIGHT_CONFIG;
})(window);
</script>
<script src="/js/insight.js"></script>
<script defer type="text/javascript">
(function(d, s) {
var j, e = d.getElementsByTagName(s)[0];
if (typeof LivereTower === 'function') { return; }
j = d.createElement(s);
j.src = 'https://cdn-city.livere.com/js/embed.dist.js';
j.async = true;
e.parentNode.insertBefore(j, e);
})(document, 'script');
</script>
</body>
</html>