添彩网导航,添彩网登陆

信雅達    1月5日下午9:05 北京時間
請選擇語言
[ 中文]

干貨!信雅達新一代AI+智能外呼機器人技術揭秘
發布者:信雅達推廣部 發布時間:2020-10-29 閱讀:1082次

JINQI,XINYADAXINYIDAIAI+ZHINENGWAIHUJIQIRENYIZAIZHESHANGYINXINGZHENGSHISHANGXIAN。

XIANGZHIDAOZUIXINYIDAIZHINENGWAIHUJIQIRENDAODIYUNYONGLESHENMEXINJISHUMA?XIANGXINZHEPIANJISHUXIANGJIEYIDINGHUIRANGNIDAHUGUOYIN~

 

 

01采用Sparse Self Attention構建模型

 

DUIYUWAIHUJIQIREN、ZHINENGKEFUZHELEIDUIXIANGYINGSHIJIANYOUZHUYANGEYAOQIUDEXITONGHECHANGJING,ZAISHIJISHENGCHANGUOCHENGZHONG,ZIRANSHAOBULEDUIGENGJIAKUAISU、WENDINGDEJISUAN,YIJIYONGSHAOLIANGDESUANLI、JIAOSHAODEZIYUANSUNHAOHUOQUGENGHAOYUNXINGXIAOGUODEJIXIANZHUIQIU。

 

ERRUJINNLPLINGYUZHONG,AttentionDAXINGQIDAO,QIHEXINZAIYUK、Q、VSANGEXIANGLIANGXULIEDEJIAOHURONGHE,QIZHONGQ、KDEJIAOHUGEICHULELIANGLIANGXIANGLIANGZHIJIANDEMOUZHONGXIANGGUANDU(QUANZHONG),ERZUIHOUDESHUCHUXULIEZESHIBAVANZHAOQUANZHONGQIUHEDEDAODE。

 

NEIMEWENTILAILE,JISHISHIMUQIANJIAOWEIZHULIUDESelf-AttentionDELILUNJISUANSHIJIANHEXIANCUNZHANYONGLIANGDOUSHIO(n2)JIBIEDE,ZHEYEYIWEIZHURUGUOXULIECHANGDUBIANCHENGYUANLAIDE2BEI,XIANCUNZHANYONGLIANGJIUSHIYUANLAIDE4BEI,JISUANSHIJIANZIRANYEJIUSHIYUANLAIDE4BEI!DANGRAN,JIASHEBINGXINGHEXINSHUZUGOUDUODEQINGKUANGXIA,JISUANSHIJIANWEIBIHUIZENGJIADAOYUANLAIDE4BEI,DANSHIXIANCUNDE4BEIQUESHISHISHIZAIZAIDE,WUKEBIMIAN。

 

WEILEJIEJUEYISHANGWENTI,XINYADAJINRONGDASHUJUYANJIUYUANZAIZHINENGWAIHUYITUSHIBIESUANFAMOKUAIGOUJIANGUOCHENGZHONG,FANGQICHUANTONGDESelf-Attention,CAIYONGSparse Self AttentionGOUJIANMOXING。

 

TONGGUOTop-kXUANZE,JIANGYUANYOUDEAttentionTUIHUAWEIXISHUAttetion,JIANDANCONGZHUYILIJUZHENSHANGKAN,JIUSHICHULEXIANGDUIJULIBUCHAOGUOkDE、XIANGDUIJULIWEIk,2k,3k,…DEZHUYILIDOUSHEWEI0,ZHEYANGYILAIAttentionJIUJUYOU“JUBUJINMIXIANGGUANHEYUANCHENGXISHUXIANGGUAN”DETEXING。

 

TONGGUOZHEZHONGFANGSHI,KEYIBAOLIUZUIYOUZHUYUYINQIZHUYIDEBUFEN,BINGSHANCHUQITAWUGUANDEXINXI,ZHEZAISHIJIDESHIYONGGUOCHENGZHONGYEQIDAOLEYIXIANGBUDAODEXIAOGUO。JINGGUOYANJIUYUANCEYAN,ZHEZHONGXUANZEXINGFANGFAZAIBAOCUNZHONGYAOXINXIHEXIAOCHUZAOSHENGFANGMIANSHIYOUXIAODE,KEYIQUEBAOZHUYILIKEYIGENGDUODIJIZHONGZAIZUIYOUGONGXIANDEJIAZHIYINSUSHANG。

 

TONGSHI,ZAISHIJISHENGCHANGUOCHENGZHONG,YOUYUWAIJIEZAOSHENGHEASRDEZHUANYIWENTI,JINGCHANGHUIZAIWENBENZHONGYINRUBUBIYAODEZAOSHENG。Sparse Self AttentionZAICHULIZHEFANGMIANWENTISHANG,HUODELELIANGHAODEXIAOGUO。

 

 

02 知識蒸餾

SUIZHUYUXUNLIANMOXINGJISHUDEFAZHAN,PIRUBert、XLNET、GPT-3DECHUXIAN,SHIDEAIJISHUDUIRENLEIRENZHIDETANSUOGENGJINYIBU。YOUQISHIZUIJINDAHUODEGPT-3,ZUOWEIDANXIANGYUYANMOXINGXUNLIANDEDAIBIAO,GPT-3DEMOXINGCHICUNDADAOLEKONGBUDE1750YI,ZAI45TBDEXUNLIANSHUJUJIACHIXIA,JIEJUELEYEJIECHANGYONGDEBertYUXUNLIANMOXINGDELINGYUNEIBIAOQIANSHUJUYILAIYIJILINGYUSHUJUFENBUGUONIHE,MUQIANGPT-3DEFew-shotZAIBUFENNLURENWUSHANGCHAOYUELEDANGQIANSOTA。

 

SUIRANYUXUNLIANMOXINGZAIHENDUOYEWUCHANGJINGBIAOXIANYOUYI,DANSHIQIMOXINGCANSHUGUODA、YUCESHIYANJIAOCHANG、YUNXINGYINGJIANCHENGBENJIAOGAODEWENTIYEDAOZHILEMOXINGHENNANLUODI。YINCI,RUHEDUIYUXUNLIANMOXINGJINXINGMOXINGYASUO,CHENGWEIGONGYEJIEHEXUESHUJIEGONGTONGGUANZHUDEWENTI。

 

WEILENENGGOUSHIYONGYEJIELINGXIANDEYUXUNLIANMOXING,XINYADAJINRONGDASHUJUYANJIUYUANDESHUJUKEXUEJIAMENCAIYONGLEZHISHIZHENGLIUMOXINGYASUOJISHU,SHIYUXUNLIANMOXINGZAIWAIHUXIANGMUZHONGDEDAOLEGAOXIAOLIYONG。

 

ZAIYUANSHIMOXINGXUNLIANJIEDUAN,YANJIUYUANDESHUJUKEXUEJIAMENJIYUYUXUNLIANMOXINGBert+SHENDUSHENJINGWANGLUOJINXINGGOUJIANWANGLUOJIEGOU,BINGJIANGJIYUYANGBENXUNLIANDEJIAOSHIMOXING(Net-Teacher)DEClass ProbabilitiesZUOWEIYASUOMOXINGXUNLIANDESUNSHIHANSHUSHURULAICANYUHOUXUZHISHIZHENGLIUXUNLIANGUOCHENG。

 

ZAIZHISHIZHENGLIUJIEDUAN,YANJIUYUANDESHUJUKEXUEJIAMENJIYUXIANGTONGDEYANGBENJINXINGNet-StudentMOXINGDEZHENGLIUXUNLIAN,DUIZHENGLIUWENDUJINXINGHESHIDEXUANQU。ZAIZHENGGEXUNLIANJIEDUAN,Net-TSHUCHUsoft-target,Net-STONGSHISHUCHUsoft-predictionsHEhard-predictions;JIANGNet-TDEsoft-targetYUNet-SDEsoft-predictionDUIYINGDEJIAOCHASHANGXIANGJIA,ZUOWEIZHENGGEMOXINGSUNSHIHANSHUDELsoftBUFEN,TONGSHIJIANGNet-SDEhard-predictionYUground-truthDEJIAOCHASHANGZUOWEIZHENGGEMOXINGSUNSHIHANSHUDELhardDEBUFEN,TONGGUOFANXIANGCHUANBODEXUNLIANFANGFAJINXINGZHENGGEMOXINGDEXUNLIAN,ZHIDAOXUNLIANTINGZHI,WANCHENGZHENGGEZHISHIZHENGLIUGUOCHENG。

 

ZHISHIZHENGLIUHOUDEMOXING,ZAIZHUNQUELVXIAJIANGBUDAO1%DEQINGKUANGXIA,YUCESHIJIANSUODUANLE8BEI,MOXINGYASUODAILAIDESHOUYISHIFENKEGUAN,KEYICHENGGONGYUNYONGDAOSHIYANYAOQIUGAODEYINGYONGCHANGJING。

 

 

03 自學習技術賦能模型自迭代

CHANGYANDAO:JUYIFANSANCAINENGSHIBANGONGBEI。JISHIJINNIANLAIZIXUEXIJISHUKUANGBIAOTUJINSHIFAZHAN,DANRENGONGZHINENGZAIMOFANGRENLEIZHETIAOLUSHANGYIJIUXUYAOFUZHONGQIANXING,BIJINGZHIYOUTONGGUOXUEXIYUJILEI,CAINENGSHIXIANCONGRUMENDAOZHUANYE、CONGQINGSEDAOCHENGSHUDEZHUANBIAN。

 

基于對“持續學習”的考慮,信雅達金融大數據研究院此次研發的外呼機器人系統支持學習迭代。在項目啟動后,生產語料不斷積累,伴隨而來的是不斷豐富的知識庫以及多種多樣的新增意圖。為了充分利用好這些寶貴數據,研究院的攻城獅們下了大功夫:將繁雜的訓練過程封裝為簡單的自訓練配置,通過一鍵式在線生成新模型,即可進行擴充語料的增量訓練,同時直接部署測試環境進行新模型測試,這樣大大縮短數據