# bert_mulit_label_chinese
中文多標籤主題意圖分類

原始碼連結
基本上就是透過內容分析主題。

為求方便，訓練原料這次我就改成使用 [Kashgari]的SMP2018ECDTCorpus資料集，安裝方式請看官網
或是透過pip

pip install kashgari

快速啟動

python3 mulit_label.py

訓練部分

訓練參數

train_batch_size = 16
eval_batch_size = 16
predict_batch_size = 16
learning_rate = 5e-5
num_train_epochs = 5

輸入資料

範例檔案（使用SMP2018ECDTCorpus資料集製作,31種意圖種類）

  data/
      - dev-kashgari.tsv
      - test-kashgari.tsv
      - train-kashgari.tsv

範例如下（標籤可能為複數）

標籤	內容
[weather, news]	查看汕头天气讯飞语音我想知道山东省东营市今天的新闻。
[cookbook]	土豆饼怎么做啊？
[map, chat]	到郑州火车站唠什么
[app, message]	打开熊猫看书发信息给董昊老婆
[website, chat]	打开索尼官方网站怎么不懂讲晚安呢
[chat, video]	能想起我吗换到米老鼠

### 評估結果

總項

class_intent_f1 = 0.9022081
class_precision = 0.9178434
class_recall = 0.88709676
class_threshold_ = 0.34673366
loss = 1.3868933
global_step = 587

評估結果

總項

class_intent_f1 = 0.9022081
class_precision = 0.9178434
class_recall = 0.88709676
class_threshold_ = 0.34673366
loss = 1.3868933
global_step = 587

分項對應的標籤可參考output/label_map.txt

class00_f1 = 0.7727272
class00_precision = 0.68
class00_recall = 0.8947368
class00_threshold = 0.03517588

class01_f1 = 1.0
class01_precision = 1.0
class01_recall = 1.0
class01_threshold = 0.7085427

class02_f1 = 1.0
class02_precision = 1.0
class02_recall = 1.0
class02_threshold = 0.30653265

class03_f1 = 0.9253731
class03_precision = 0.92537314
class03_recall = 0.92537314
class03_threshold = 0.31658292

class04_f1 = 0.96551716
class04_precision = 1.0
class04_recall = 0.93333334
class04_threshold = 0.61809045

class05_f1 = 0.88888884
class05_precision = 0.8888889
class05_recall = 0.8888889
class05_threshold = 0.55778897

class06_f1 = 0.9906541
class06_precision = 0.9814815
class06_recall = 1.0
class06_threshold = 0.07537688

class07_f1 = 0.10526314
class07_precision = 0.05882353
class07_recall = 0.5
class07_threshold = 0.03517588

class08_f1 = 0.9333333
class08_precision = 1.0
class08_recall = 0.875
class08_threshold = 0.06030151

class09_f1 = 0.84782594
class09_precision = 0.90697676
class09_recall = 0.79591835
class09_threshold = 0.3115578

class10_f1 = 0.9803921
class10_precision = 1.0
class10_recall = 0.96153843
class10_threshold = 0.14572865

class11_f1 = 0.98305076
class11_precision = 0.96666664
class11_recall = 1.0
class11_threshold = 0.055276383

class12_f1 = 0.9
class12_precision = 0.8181818
class12_recall = 1.0
class12_threshold = 0.030150754

class13_f1 = 0.9166666
class13_precision = 0.9166667
class13_recall = 0.9166667
class13_threshold = 0.16080402

class14_f1 = 0.8421052
class14_precision = 0.8
class14_recall = 0.8888889
class14_threshold = 0.08040201

class15_f1 = 1.0
class15_precision = 1.0
class15_recall = 1.0
class15_threshold = 0.6381909

class16_f1 = 0.9285714
class16_precision = 1.0
class16_recall = 0.8666667
class16_threshold = 0.6934673

class17_f1 = 0.98181814
class17_precision = 0.96428573
class17_recall = 1.0
class17_threshold = 0.8190955

class18_f1 = 0.74999994
class18_precision = 1.0
class18_recall = 0.6
class18_threshold = 0.040201005

class19_f1 = 0.98305076
class19_precision = 0.96666664
class19_recall = 1.0
class19_threshold = 0.16080402

class20_f1 = 0.96296287
class20_precision = 1.0
class20_recall = 0.9285714
class20_threshold = 0.085427135

class21_f1 = 1.0
class21_precision = 1.0
class21_recall = 1.0
class21_threshold = 0.04522613

class22_f1 = 1.0
class22_precision = 1.0
class22_recall = 1.0
class22_threshold = 0.06532663

class23_f1 = 0.98113203
class23_precision = 1.0
class23_recall = 0.962963
class23_threshold = 0.44221106

class24_f1 = 0.98181814
class24_precision = 0.96428573
class24_recall = 1.0
class24_threshold = 0.14070351

class25_f1 = 0.982456
class25_precision = 1.0
class25_recall = 0.9655172
class25_threshold = 0.06030151

class26_f1 = 1.0
class26_precision = 1.0
class26_recall = 1.0
class26_threshold = 0.63316584

class27_f1 = 0.70270264
class27_precision = 0.8125
class27_recall = 0.61904764
class27_threshold = 0.41708544

class28_f1 = 0.7368421
class28_precision = 0.7291667
class28_recall = 0.7446808
class28_threshold = 0.24120604

class29_f1 = 0.87499994
class29_precision = 0.9130435
class29_recall = 0.84
class29_threshold = 0.06532663

class30_f1 = 0.8999999
class30_precision = 0.9
class30_recall = 0.9
class30_threshold = 0.6683417

預測

有輸出一份結果在 bert_mulit_label_chinese/output/test_results.tsv裡面

		答案標籤	預測標籤
O	O	message radio	message radio	[CLS]帮我调频92.8桐庐人民广播电台行不行？发条短信给杨桑[SEP]
O	O	cookbook music	cookbook music	[CLS]我想听星月神话黑芝麻怎么炒？[SEP]
O	O	novel weather	novel weather	[CLS]帮我查一下明日天气我要看唯美童话小说[SEP]
X	X	chat epg	datetime epg train	[CLS]我想去上海坐火车得几天中央六台的节目单[SEP]
O	O	chat	chat	[CLS]什么一年多了你怎么认识我的[SEP]
X	X	tvchannel website	app website	[CLS]新闻解码访问京东商城的网页[SEP]
O	O	news	news	[CLS]等，请播报新闻钓鱼岛新闻[SEP]
O	O	cinemas	cinemas	[CLS]搜索合肥左岸电影城。给我放一部最新的电影。[SEP]
O	O	cookbook map	cookbook map	[CLS]打了怎么做红烧肉的做法？从武汉到呼和浩特[SEP]
O	O	app cookbook	app cookbook	[CLS]糖醋排骨怎么做？好了额，你。打开qq同步助手[SEP]
X	X	chat tvchannel	chat match tvchannel	[CLS]cctv农业欧洲杯谁能夺冠[SEP]
O	O	cookbook message	cookbook message	[CLS]五梆瓜鸡蛋汤怎么做？我要发信息给高鹤航[SEP]
O	O	app video	app video	[CLS]惊情四百年打开相机这[SEP]
O	O	cookbook video	cookbook video	[CLS]萝卜炖排骨汤怎么做？尼古拉斯凯奇[SEP]
O	O	chat	chat	[CLS]你怎么知道～怎么写[SEP]
O	O	chat	chat	[CLS]顾超是谁你们你过生日吃蛋糕吗[SEP]
O	O	bus map	bus map	[CLS]从无锡市到西安市的汽车。帮我查一下到狼牙山怎么走[SEP]
X	X	translation video	translation	[CLS]抽身游戏翻译秋天[SEP]
O	O	chat poetry	chat poetry	[CLS]么么么哒朗诵一首李白的诗词。[SEP]
O	O	chat	chat	[CLS]怎么走爽了呢喝啥子哩[SEP]
X	X	email video	email	[CLS]我想打开新邮件请帮我播放樱桃[SEP]
X	X	chat weather	map weather	[CLS]笨笨家在哪今日天气怎么样啊[SEP]
O	O	chat music	chat music	[CLS]我要听张雨生的歌。我能把你教坏吗[SEP]
X	X	app poetry	poetry video	[CLS]帮我找到微信读首唐诗[SEP]
O	O	chat website	chat website	[CLS]我就不是好孩子了怎么了打开网页猫扑[SEP]
O	O	music translation	music translation	[CLS]翻译今天出去玩毛阿敏的歌[SEP]
O	O	app chat	app chat	[CLS]没听清吗请帮我打开uc[SEP]
O	O	calc telephone	calc telephone	[CLS]呼叫浦蒲54乘上45[SEP]
X	X	flight weather	chat flight weather	[CLS]你知道今天南京的天气吗明天去纽约的航班。[SEP]
X	X	message website	bus map message	[CLS]上广州坐车网查一下从这里到大咋地怎么走传简讯给家里[SEP]
X	X	stock website	website	[CLS]创元帮我链接到新浪网[SEP]
X	X	cookbook website	bus cookbook map	[CLS]面包怎么做上广州坐车网查一下从这里到大咋地怎么走[SEP]
O	O	epg website	epg website	[CLS]年代秀什么时候播帮我上下新浪网[SEP]
O	O	cookbook news	cookbook news	[CLS]是否还有其他关于钓鱼岛的新闻好吧，我想知道稻香排骨是怎么做的？[SEP]
X	X	contacts website	website	[CLS]查看之丽丽的号码打开江苏移动网上营业厅网页[SEP]
O	O	video	video	[CLS]青瓷电视剧蝎子王蝎子王[SEP]
O	O	chat cookbook	chat cookbook	[CLS]你是妹妹还是哥哥那好，那你告诉我西红柿炒鸡蛋，怎么去做？[SEP]
O	O	cookbook riddle	cookbook riddle	[CLS]用菜籽油做面包出一个谜语给我猜猜吧！[SEP]
O	O	cinemas train	cinemas train	[CLS]查询湛江到广州的火车票搜索合肥左岸电影城。[SEP]
O	O	news video	news video	[CLS]泰州新闻。要看特种兵[SEP]
X	X	chat translation	map translation	[CLS]你还没有结婚用英语怎么讲笨笨家在哪[SEP]
O	O	cookbook health	cookbook health	[CLS]肠胃炎怎么治？臭豆腐怎么做啊？[SEP]
O	O	chat website	chat website	[CLS]打开聚划算网站推荐几部呗[SEP]
O	O	epg message	epg message	[CLS]cctv9明晚有转播仁心解码2呢给李慧发个短信说我要回家[SEP]
O	O	cookbook	cookbook	[CLS]炸里鸡的做法。鸡蛋的做法[SEP]
O	O	news tvchannel	news tvchannel	[CLS]搜索最新新闻。找深圳卫视[SEP]
O	O	riddle	riddle	[CLS]猜脑筋急转弯二。给我打个谜语，啊！[SEP]
O	O	map news	map news	[CLS]钓鱼岛新闻查查科大讯飞在哪里[SEP]
O	O	cookbook train	cookbook train	[CLS]糖醋鲤鱼怎么做啊？你只负责吃，c则c。查询广州到北京的火车[SEP]
O	O	chat health	chat health	[CLS]更年期综合症。你知道怎么拍照么[SEP]
O	O	bus news	bus news	[CLS]盘锦新闻。去深圳怎么坐车？[SEP]
O	O	message news	message news	[CLS]发信息给[UNK][UNK]k新闻想听新闻。[SEP]
X	X	app poetry	poetry	[CLS]我背一首生查子。帮我找到微信[SEP]
O	O	cookbook video	cookbook video	[CLS]拔丝芋头怎么做查雷神电影[SEP]
O	O	schedule train	schedule train	[CLS]早晨8:10叫我。查询广州到贵港的火车[SEP]
X	X	chat video	chat epg	[CLS]可以告诉我你的性别么最近好评的晚会[SEP]
O	O	map music	map music	[CLS]上海到北京怎么走我想听星月神话[SEP]
O	O	video weather	video weather	[CLS]帮我看看今天的天气恶熊出没[SEP]
X	X	poetry tvchannel	epg poetry tvchannel	[CLS]背诵唐诗三百首静夜思。切换上海频道[SEP]
O	O	poetry video	poetry video	[CLS]来一首李清照的词小c帮我搜一下断剑视频[SEP]
O	O	health video	health video	[CLS]爱情工寓内分泌失调了怎么调理？[SEP]
O	O	cookbook poetry	cookbook poetry	[CLS]酸菜粉条肉的做法。我给我念一首李白的诗吧？[SEP]
O	O	chat weather	chat weather	[CLS]你是哪里的人啊今天广州的天气怎样[SEP]
X	X	music train	app music train	[CLS]开封到信阳的火车我叫小沈阳这首歌[SEP]
O	O	cookbook lottery	cookbook lottery	[CLS]红烧肉怎么煮啊任意购买三张大乐透[SEP]
O	O	cookbook news	cookbook news	[CLS]讯飞语点请你搜索一下番茄炒蛋怎么做？帮我查一查今天的新闻。[SEP]
X	X	map music	map poetry	[CLS]新东站怎么走给我来一首，相见恨晚的音乐。[SEP]
O	O	chat map	chat map	[CLS]怎么去农业银行两点半夜鸡叫什么了不少钱包[SEP]
O	O	chat	chat	[CLS]你不是本人吗嗯你叫啥呀[SEP]
O	O	message video	message video	[CLS]魔幻电影发消息给周磊[SEP]
O	O	epg flight	epg flight	[CLS]回放昨晚的焦点访谈查询今天上海到成都的航班[SEP]
O	O	schedule translation	schedule translation	[CLS]苹果翻译成英语明天叫我8:10起床。[SEP]
O	O	cinemas train	cinemas train	[CLS]星期四去上海的火车离我最近的电影院。[SEP]
O	O	chat music	chat music	[CLS]she的歌几八痒[SEP]
O	O	chat weather	chat weather	[CLS]你是穆斯林吗我想看一下温州的天气[SEP]
O	O	chat website	chat website	[CLS]这么笨啊怪不得叫笨笨易车网看看[SEP]
O	O	chat flight	chat flight	[CLS]哪里牛逼了去桂林的航班怎样[SEP]
O	O	message weather	message weather	[CLS]发短信给王勋连说我很好明天深圳天气怎么样[SEP]
X	X	train website	bus map	[CLS]打开新浪网我肏查一下从杭州到上海高铁时刻表[SEP]
O	O	epg news	epg news	[CLS]湖南卫视现在演什么节目我想听重庆新闻。[SEP]
X	X	chat tvchannel	chat epg	[CLS]cctv-12套的节目单是我说话太直接了么[SEP]
O	O	video weather	video weather	[CLS]帮我看下最近有什么好看的电影？天气预报广州[SEP]
O	O	calc train	calc train	[CLS]北京到上海的火车票54乘上45[SEP]
O	O	riddle website	riddle website	[CLS]开网页给我打个谜语，啊！[SEP]
X	X	email stock	email stock weather	[CLS]帮我发邮件给婷婷我想了解地产板块股票现在怎么样[SEP]
X	X	tvchannel video	tvchannel	[CLS]北京二三女休夫[SEP]
O	O	chat train	chat train	[CLS]这是哪年的事帮我查下上海到合肥的火车[SEP]
O	O	chat music	chat music	[CLS]凭啥来点歌听听[SEP]
O	O	poetry train	poetry train	[CLS]若相惜不弃下一句是什么？从广州到南昌的火车票时刻表[SEP]
X	X	chat cinemas	cinemas	[CLS]我要查找雷迪波尔电影院最新上映的大片你长什么样子[SEP]
O	O	contacts telephone	contacts telephone	[CLS]打电话给幺三七幺零五五二幺九九新建联系人李四号码是一二三四五六七八[SEP]
O	O	cookbook message	cookbook message	[CLS]发短信给王二白叫她明天过来玩不懂饭的做法。[SEP]
O	O	chat tvchannel	chat tvchannel	[CLS]央视9可以教教吗?[SEP]
O	O	app chat	app chat	[CLS]刘挺是谁？搜索手机办公软件[SEP]
O	O	news video	news video	[CLS]么，哥来一下最新的新闻。最近最热的电视剧[SEP]
O	O	news	news	[CLS]我想听重庆新闻。搜索最新新闻。[SEP]
O	O	chat train	chat train	[CLS]你什么梦想重庆到成都的火车票多少钱[SEP]
O	O	cookbook map	cookbook map	[CLS]板栗怎么煮？去红心中路怎么走[SEP]
X	X	chat novel	novel	[CLS]下面有什么找找三毛的爱情小说[SEP]
O	O	map weather	map weather	[CLS]南山医院到北大医院路线我想知道北京的天气[SEP]

發佈留言取消回覆

Related Stories

單一模型無法解決一切？代理系統將是未來

從圖形資料庫到雲端服務：理解數據世界中的關係魔法

從 Presto 到 SuperSQL-Spark：兩大查詢引擎的比較與未來展望

You may have missed

單一模型無法解決一切？代理系統將是未來

從圖形資料庫到雲端服務：理解數據世界中的關係魔法

從 Presto 到 SuperSQL-Spark：兩大查詢引擎的比較與未來展望

探索人工智慧的抽象推理之旅：ARC-AGI 測試與最新進展

快速啟動

訓練部分

訓練參數

輸入資料

評估結果

預測

相關

發佈留言 取消回覆

Related Stories

單一模型無法解決一切？代理系統將是未來

從圖形資料庫到雲端服務：理解數據世界中的關係魔法

從 Presto 到 SuperSQL-Spark：兩大查詢引擎的比較與未來展望

You may have missed

單一模型無法解決一切？代理系統將是未來

從圖形資料庫到雲端服務：理解數據世界中的關係魔法

從 Presto 到 SuperSQL-Spark：兩大查詢引擎的比較與未來展望

探索人工智慧的抽象推理之旅：ARC-AGI 測試與最新進展

發佈留言取消回覆