çŸãããã®ãäœãããšãä»äºã®å Žåãçµæã¯ã¿ããªã®ç®ã®åã«ããã®ã§ãããã«ã€ããŠããŸã話ãå¿
èŠã¯ãããŸããã ãããããã§ã³ã¹ã®ç¢æãæ¶ããŠãããã§ã³ã¹ããã¡ããšããŠããéãããŸãã¯äœãééã£ããã®ãæ¶å»ãããŸã§ã¯ã誰ãããªãã®ä»äºã«æ°ã¥ããŸããã
ã³ã¡ã³ããæ®ããããã¬ãã¥ãŒãæ®ããããã¡ãã»ãŒãžãéä¿¡ããããåçãã¢ããããŒããããã§ãããµãŒãã¹ã¯ãé ããæ©ããã¹ãã ãè©æ¬ºããããã€ã®åé¡ã«çŽé¢ããŸãã ããã¯é¿ããããšã¯ã§ããŸãããã察åŠããå¿ èŠããããŸãã
ç§ã®ååã¯ããã€ã«ã§ããYandex ãµãŒãã¹ã®ãŠãŒã¶ãŒããã®ãããªåé¡ããä¿è·ããã¹ãã 察çããŒã ã§åããŠããŸãã ç§ãã¡ã®ä»äºã¯ããŸã泚ç®ãããªãã®ã§ (ããã¯è¯ãããšãªã®ã§ã!)ãä»æ¥ã¯ããã«ã€ããŠè©³ããã話ããŸãã ç¯åºŠã圹ã«ç«ããªãå Žåãšãç¯åºŠã®æå¹æ§ã瀺ãå¯äžã®ææšã粟床ã ãã§ã¯ãªãçç±ãåŠã³ãŸãã ãŸããç¬ãšç«ã®äŸã䜿ã£ãŠæªå£ã«ã€ããŠèª¬æãããæªå£ãèšã人ã®ããã«èãããããšãæã
圹ç«ã€çç±ã«ã€ããŠã説æããŸãã
Yandex ã«ã¯ããŠãŒã¶ãŒãã³ã³ãã³ããå
¬éãããµãŒãã¹ããŸããŸãå¢ããŠããŸãã Yandex.Q ã§è³ªåãããåçãæžããããYandex.District ã§ã€ãŒã ãã¥ãŒã¹ã«ã€ããŠè©±ãåã£ãããYandex.Maps ã§ã®äŒè©±ã§äº€éç¶æ³ãå
±æãããã§ããŸãã ãããããµãŒãã¹ã®å©çšè
ãå¢ãããšãè©æ¬ºåž«ãã¹ãã éä¿¡è
ã«ãšã£ãŠé
åçã«ãªããŸãã 圌ãã¯ãã£ãŠæ¥ãŠã³ã¡ã³ããèšå
¥ããŸãã圌ãã¯ç°¡åã«ãéãæäŸããå¥è·¡ã®æ²»çæ³ã宣äŒãã瀟äŒçå©çãçŽæããŸãã ã¹ãããŒã®ããã§ããéã倱ããŠãŒã¶ãŒãããŸãããã¹ãã ãããµãããã£ãç²æ«ãªãµãŒãã¹ã«æéãè²»ããæ欲ã倱ããŠãŒã¶ãŒãããŸãã
åé¡ã¯ããã ãã§ã¯ãããŸããã ç§ãã¡ã¯ãŠãŒã¶ãŒãè©æ¬ºåž«ããå®ãã ãã§ãªããã³ãã¥ãã±ãŒã·ã§ã³ãããããé°å²æ°ãäœãåºãããåªããŠããŸãã ã³ã¡ã³ãã§æªå£ã䟮蟱ã«çŽé¢ããå Žåããã®äººã¯å»ã£ãŠäºåºŠãšæ»ããªãå¯èœæ§ããããŸãã ã€ãŸããããªããããã«å¯ŸåŠã§ããå¿ èŠããããŸãã
ã¯ãªãŒã³ãŠã§ã
ç§ãã¡ã«ã¯ããããããšã§ãããæåã®éçºã¯æ€çŽ¢ãã€ãŸãæ€çŽ¢çµæå ã®ã¹ãã ãšæŠãéšåã§çãŸããŸããã çŽ 18 幎åããã¡ããªãŒæ€çŽ¢ã XNUMX æ³ä»¥äžã®ã«ããŽãªããã®åçãå¿ èŠãšããªãã¯ãšãªã«å¯ŸããŠã¢ãã«ã ã³ã³ãã³ãããã£ã«ã¿ãªã³ã°ããã¿ã¹ã¯ãç»å ŽããŸããã ãããããã«ããšæªå£ã«é¢ããæåã®æå ¥åèŸæžãç»å Žããæ¹æ³ã§ãããåæè ã«ãã£ãŠè£å ãããŸããã äž»ãªã¿ã¹ã¯ã¯ããªã¯ãšã¹ããã¢ãã«ã ã³ã³ãã³ãã®è¡šç€ºã蚱容ããããªã¯ãšã¹ããšããã§ãªããªã¯ãšã¹ãã«åé¡ããããšã§ããã ãã®ã¿ã¹ã¯ã§ã¯ãããŒã¯ã¢ãããåéããããã¥ãŒãªã¹ãã£ãã¯ãæ§ç¯ãããã¢ãã«ããã¬ãŒãã³ã°ãããŸããã ããããäžèŠãªã³ã³ãã³ãããã£ã«ã¿ãªã³ã°ããããã®æåã®éçºãç»å Žããæ¹æ³ã§ãã
æéãçµã€ã«ã€ããŠãUGC (ãŠãŒã¶ãŒçæã³ã³ãã³ã) ã Yandex ã«è¡šç€ºãããããã«ãªããŸãããããã¯ããŠãŒã¶ãŒèªèº«ãäœæããã¡ãã»ãŒãžã§ãããYandex ãå ¬éããã ãã§ãã äžèšã®çç±ã«ãããå€ãã®ã¡ãã»ãŒãžã¯ç¢ºèªããã«å ¬éã§ããã管çãå¿ èŠã§ããã 次ã«ããã¹ãŠã® Yandex UGC 補åãã¹ãã ãšæ»æè ããä¿è·ããæ€çŽ¢ã§äžèŠãªã³ã³ãã³ãããã£ã«ã¿ãªã³ã°ããéçºãå©çšãããµãŒãã¹ãäœæããããšã«ããŸããã ãã®ãµãŒãã¹ã¯ãClean WebããšåŒã°ããŠããŸããã
æ°ããã¿ã¹ã¯ãšæšé²è ããã®æ¯æŽ
æåã¯åçŽãªèªååã ãã圹ã«ç«ã¡ãŸããããµãŒãã¹ã¯ããã¹ããéä¿¡ãããããã€èŸæžããã«ãèŸæžãæ£èŠè¡šçŸããããã«å¯ŸããŠå®è¡ããŸãããã¢ããªã¹ãããã¹ãŠãæåã§ç·šéããŸããã ããããæéãçµã€ã«ã€ããŠããã®ãµãŒãã¹ã¯ãŸããŸãå€ãã® Yandex 補åã§äœ¿çšãããããã«ãªããç§ãã¡ã¯æ°ããªåé¡ã«å¯ŸåŠããæ¹æ³ãåŠã°ãªããã°ãªããŸããã§ããã
å€ãã®å ŽåããŠãŒã¶ãŒã¯ã¬ãã¥ãŒã®ä»£ããã«ãèªåã®æ¥çžŸãå¢ããããšæå³ã®ãªãäžé£ã®æçŽãå ¬éããæã«ã¯ç«¶åä»ç€Ÿã®äŒç€Ÿã®ã¬ãã¥ãŒã§èªç€Ÿã®å®£äŒããããŸãæã«ã¯åã«çµç¹ãæ··ä¹±ãããŠããã ã·ã§ããã«ã€ããŠã®ã¬ãã¥ãŒãæžã蟌ãããšããããŸããå®ç§ã«èª¿çãããéïŒã ãããããã€ãã人工ç¥èœã¯ããããããã¹ãã®æå³ãå®å šã«ææ¡ã§ããããã«ãªãã§ãããããããçŸåšãèªååã¯äººéããã察åŠãäžæãªå ŽåããããŸãã
æäœæ¥ã§ããŒãã³ã°ãè¡ããªããšãããå®è¡ã§ããªãããšãæããã«ãªã£ãã®ã§ãåè·¯ã«ç¬¬ XNUMX 段éãè¿œå ãã人ã«ããæåæ€æ»ã«éããŸããã åé¡åšã§åé¡ãæ€åºãããªãã£ãåºçæžã¿ã®ããã¹ããããã«å«ãŸããŠããŸããã ãã®ãããªäœæ¥ã®èŠæš¡ã¯å®¹æã«æ³åã§ãããããç§ãã¡ã¯è©äŸ¡è ã«é Œãã ãã§ãªããã矀è¡ã®ç¥æµããå©çšããŸãããã€ãŸããè©äŸ¡è ã«å©ããæ±ããŸããã 圌ãã¯ããã·ã³ãèŠéãããã®ãç¹å®ããããã«ãã£ãŠãã·ã³ã«æããã®ãå©ããŠãããŸãã
ã¹ããŒã ãã£ãã·ã³ã°ãš LSH ããã·ã¥
ã³ã¡ã³ããæ±ãéã«ç§ãã¡ãééãããã 5 ã€ã®åé¡ã¯ãã¹ãã ãããæ£ç¢ºã«ã¯ã¹ãã ã®éãšæ¡æ£é床ã§ããã Yandex.Region ã®å©çšè ãæ¥éã«å¢å ãå§ãããšãã¹ãããŒãããã«ãã£ãŠæ¥ãŸããã 圌ãã¯ãããã¹ãããããã«å€æŽããããšã§æ£èŠè¡šçŸãåé¿ããæ¹æ³ãåŠã³ãŸããã ãã¡ãããã¹ãã ã¯åŒãç¶ãçºèŠãããåé€ãããŸããããYandex ã®èŠæš¡ã§ã¯ãããšã XNUMX åéæçš¿ãããåãå ¥ããããã¡ãã»ãŒãžãæ°çŸäººãé²èŠ§ããå¯èœæ§ããããŸããã
ãã¡ãããããã¯ç§ãã¡ã«ã¯åããªãã£ããããLSH ã«åºã¥ããŠã¹ããŒã ããã¹ã ãã£ãã·ã¥ãäœæããŸãã (
ãã®ãœãªã¥ãŒã·ã§ã³ã«ãããé¡äŒŒããããã¹ãã«å¯Ÿããåé¡åãšããŒã¯è ã®å€å®ãåå©çšã§ããããã«ãªããŸããã ã¹ãã æ»æäžã¯ãæåã®ã¡ãã»ãŒãžãã¹ãã£ã³ã«åæ Œãããã¹ãã ãå€å®ã§ãã£ãã·ã¥ã«å ¥ããšããã«ãå€æŽããããã®ã§ãã£ãŠãããã¹ãŠã®æ°ããåæ§ã®ã¡ãã»ãŒãžãåãå€å®ãåããèªåçã«åé€ãããŸããã ãã®åŸãã¹ãã åé¡åšããã¬ãŒãã³ã°ããèªåçã«åãã¬ãŒãã³ã°ããæ¹æ³ãåŠã³ãŸãããããã®ãã¹ããŒã ãã£ãã·ã¥ãã¯ç§ãã¡ã®å¿ã«æ®ããä»ã§ãé »ç¹ã«åœ¹ã«ç«ã¡ãŸãã
åªããããã¹ãåé¡å
ã¹ãã 察çããäŒãæéããªãã£ãã®ã§ãã³ã³ãã³ãã® 95% ãæåã§ç®¡çãããŠããããšã«æ°ä»ããŸãããåé¡åã¯éåã«ã®ã¿åå¿ããããã¹ãã®ã»ãšãã©ã¯è¯å¥œã§ãã åœç€Ÿã§ã¯ã95 ä»¶äž 100 件ã§ããã¹ãŠåé¡ãããŸããããšããè©äŸ¡ãäžããæž ææ¥è ã掟é£ããŠããŸãã ç§ã¯ãè¯ãã³ã³ãã³ãã®åé¡åšãäœæãããšããçããä»äºãããªããã°ãªããŸããã§ãããã幞éãªããšã«ããã®éã«ååãªããŒã¯ã¢ãããèç©ãããŸããã
æåã®åé¡åã¯æ¬¡ã®ããã«ãªããŸããããã¹ããèŠåºãèªåã (åèªãåæ圢åŒã«æ»ã)ãè£å©åè©ããã¹ãŠæšãŠãŠãäºåã«æºåããããé©åãªèŠåºãèªã®èŸæžãã䜿çšããŸãã ããã¹ãå ã®ãã¹ãŠã®åèªããè¯å¥œãã§ããå Žåãããã¹ãå šäœã«ã¯éåãå«ãŸããŠããŸããã ããŸããŸãªãµãŒãã¹ã§ããã®ã¢ãããŒãã«ãããæåããŒã¯ã¢ããã® 25 ïœ 35% ãããã«èªååãããŸããã ãã¡ããããã®ã¢ãããŒãã¯çæ³çã§ã¯ãããŸãããããã€ãã®ç¡å®³ãªåèªãçµã¿åããããšãéåžžã«æ»æçãªçºèšãåŸãããã®ã¯ç°¡åã§ãããããã«ãããé©åãªã¬ãã«ã®èªååã«ããã«å°éããããšãã§ããããè€éãªã¢ãã«ããã¬ãŒãã³ã°ããæéãåŸãããšãã§ããŸããã
åªããããã¹ãåé¡åšã®æ¬¡ã®ããŒãžã§ã³ã«ã¯ãç·åœ¢ã¢ãã«ããã·ãžã§ã³ ããªãŒãããã³ãããã®çµã¿åããããã§ã«å«ãŸããŠããŸãã ããšãã°ãç¡ç€Œã䟮蟱ãããŒã¯ããã«ã¯ãBERT ãã¥ãŒã©ã« ãããã¯ãŒã¯ãè©ŠããŸãã æèå
ã®åèªã®æå³ãšãç°ãªãæã®åèªéã®ã€ãªãããææ¡ããããšãéèŠã§ãããBERT ã¯ãããããŸãæ©èœããŸãã ïŒãšããã§ãæè¿ãã¥ãŒã¹ã®ååã
æ£ç¢ºããå®å šæ§ãã¹ããŒã
éçºããã«ã¯ãç¹å®ã®èªååé¡åãã©ã®ãããªå©ç¹ããããããããã®å€æŽç¹ãããã³æåãã§ãã¯ã®å質ãäœäžããŠãããã©ãããç解ããå¿ èŠããããŸãã ãããè¡ãããã«ã粟床ãšåçŸçã®ã¡ããªã¯ã¹ã䜿çšããŸãã
粟床ã¯ãæªãã³ã³ãã³ãã«é¢ãããã¹ãŠã®å€å®ã®ãã¡ãæ£ããå€å®ã®å²åã§ãã 粟床ãé«ããªãã»ã©ã誀æ€ç¥ãå°ãªããªããŸãã æ£ç¢ºãã«æ³šæãæããªããã°ãçè«äžã¯ã¹ãã ããããã€ãªã¡ãã»ãŒãžããã¹ãŠåé€ãããããã«å ããŠé©åãªã¡ãã»ãŒãžã®ååãåé€ã§ããŸãã äžæ¹ã§ã粟床ã®ã¿ã«äŸåããå Žåãæé«ã®ãã¯ãããžãŒã¯èª°ã«ããŸã£ããåŒã£ããããªããã®ã«ãªããŸãã ãããã£ãŠãå®å šæ§ã®ææšãšããŠãäžè¯ã³ã³ãã³ãã®ç·éã«å ããç¹å®ãããäžè¯ã³ã³ãã³ãã®å²åãšããææšããããŸãã ããã XNUMX ã€ã®ææšã¯çžäºã«ãã©ã³ã¹ããšããŸãã
枬å®ããããã«ãåãµãŒãã¹ã®åä¿¡ã¹ããªãŒã å šäœããµã³ããªã³ã°ããå°é家ã«ããè©äŸ¡ãšãã·ã³ ãœãªã¥ãŒã·ã§ã³ãšã®æ¯èŒã®ããã«è©äŸ¡è ã«ã³ã³ãã³ã ãµã³ãã«ãæäŸããŸãã
ããããããäžã€éèŠãªææšããããŸãã
äžã§ãåãå ¥ããããã¡ãã»ãŒãžã¯ 5 å以å ã§ãäœçŸäººãã®äººã«èªãŸããå¯èœæ§ããããšæžããŸããã ããã§ç§ãã¡ã¯ãæªè³ªãªã³ã³ãã³ããé衚瀺ã«ããåã«ãäœå人ã ã«èŠããããæ°ããŸãã å¹ççã«äœæ¥ããã ãã§ã¯ååã§ã¯ãªããè¿ éã«äœæ¥ããå¿ èŠããããããããã¯éèŠã§ãã ãããŠãæªå£ã«å¯Ÿããé²åŸ¡çãæ§ç¯ãããšããç§ãã¡ã¯ãããæ倧éã«æããŸããã
ç¬ãšç«ãäŸã«æããåããã£ãºã
ã¡ãã£ãšããåæ çãªäœè«ã ãããã€ã䟮蟱ã¯æªæã®ãããªã³ã¯ã»ã©å±éºã§ã¯ãªããã¹ãã ã»ã©è¿·æã§ã¯ãªããšèšã人ããããããããŸããã ããããç§ãã¡ã¯äœçŸäžãã®ãŠãŒã¶ãŒã«ãšã£ãŠå¿«é©ãªã³ãã¥ãã±ãŒã·ã§ã³ç°å¢ãç¶æããããåªããŠããã人ã ã¯äŸ®èŸ±ãããå Žæã«ã¯æ»ãããããŸããã ããã¬ãå«ãå€ãã®ã³ãã¥ããã£ã®èŠåã«ã眵ãã䟮蟱ã®çŠæ¢ãæèšãããŠããã®ã¯åœç¶ã®ããšã§ãã ããã話ã¯éžããŸãã
æªå£ãèšãèŸæžã¯ãã·ã¢èªã®è±ããããã¹ãŠåŠçããããšã¯ã§ããŸããã äž»èŠãªå®£èªã«ãŒãã XNUMX ã€ãããªããšããäºå®ã«ããããããããããããéåžžã®ãšã³ãžã³ã§ã¯ææã§ããªãç¡æ°ã®åèªãäœæã§ããŸãã ããã«ãåèªã®äžéšãé³èš³ããããé¡äŒŒããçµã¿åããã§æåã眮æããããæåã䞊ã¹æ¿ããããã¢ã¹ã¿ãªã¹ã¯ãè¿œå ãããããããšãã§ããŸããå Žåã«ãã£ãŠã¯ãæèããªããã°ããŠãŒã¶ãŒãæ±ãèšèãæå³ãããã®ã§ãããšå€æããããšãåºæ¬çã«äžå¯èœã§ãã ç§ãã¡ã¯ããã«ã®ã«ãŒã«ãå°éããŠããã®ã§ãå®éã®äŸã§ã¯ãªããç«ãç¬ã䜿ã£ãŠãããå®èšŒããŸãã
ãåŸæ³ããšç«ã¯èšããŸããã ããããç«ãå¥ã®èšèãèšã£ãããšã¯ããããŸã...
ç§ãã¡ã¯èŸæžã®ããããŸãäžèŽãã¢ã«ãŽãªãºã ãšãããã¹ããŒããªååŠçã«ã€ããŠèãå§ããŸãããé³èš³ãæäŸããã¹ããŒã¹ãšå¥èªç¹ã貌ãåããããã¿ãŒã³ãæ¢ãããããã«åå¥ã®æ£èŠè¡šçŸãæžããŸããã ãã®ã¢ãããŒãã¯çµæããããããŸããããå€ãã®å Žåã粟床ãäœäžããæãŸããå®å šæ§ãåŸãããŸããã§ããã
ããã§ç§ãã¡ã¯ãæªå£ãèšã人ã®æ°æã¡ã«ãªã£ãŠèãããããšã«ããŸããã ç§ãã¡ã¯èªåãã¡ã§ããŒã¿ã«ãã€ãºãå°å ¥ãå§ããŸãããæåã䞊ã¹æ¿ããããã¿ã€ããã¹ãçæããããæåã䌌ãã¹ãã«ã«çœ®ãæãããããŸããã ãã®ããã®æåã®ããŒã¯ã¢ããã¯ããããèŸæžã倧éã®ããã¹ã ã³ãŒãã¹ã«é©çšããããšã«ãã£ãŠååŸãããŸããã XNUMX ã€ã®æãããã€ãã®æ¹æ³ã§ã²ãããšãå€ãã®æãã§ããŸãã ãã®ããã«ããŠããã¬ãŒãã³ã° ãµã³ãã«ãæ°ååã«å¢ããããšãã§ããŸãã æ®ã£ãŠããã®ã¯ãã³ã³ããã¹ããèæ ®ããå€ããå°ãªããã¹ããŒããªã¢ãã«ããçµæãšããŠåŸãããããŒã«ã§ãã¬ãŒãã³ã°ããããšã ãã§ãã
æçµçãªæ±ºå®ã«ã€ããŠè©±ãã®ã¯ææå°æ©ã§ãã ç§ãã¡ã¯ãŸã ãã®åé¡ãžã®ã¢ãããŒããå®éšäžã§ãããããã€ãã®å±€ãããªãåçŽãªã·ã³ããªãã¯ç³ã¿èŸŒã¿ãããã¯ãŒã¯ãèŸæžãéåžžã®ãšã³ãžã³ããã倧å¹
ã«åªããŠããããšããã§ã«ããããŸããã€ãŸãã粟床ãšåçŸçã®äž¡æ¹ãåäžãããããšãå¯èœã§ãã
ãã¡ãããç¹ã«åé¡ãéåžžã«å±éºãªå Žåã«ã¯ãæãé«åºŠãªèªååã§ãããåé¿ããæ¹æ³ãåžžã«ååšããããšãç§ãã¡ã¯ç解ããŠããŸããæããªæ©æ¢°ãç解ã§ããªããããªæ¹æ³ã§èšè¿°ããŠãã ããã ããã§ããã¹ãã ãšã®æŠããšåæ§ãç§ãã¡ã®ç®æšã¯ããããã€ãªãã®ãæžãããå¯èœæ§ãã®ãã®ãæ ¹çµ¶ããããšã§ã¯ãªããç§ãã¡ã®ä»»åã¯ãã²ãŒã ãããããã«å€ããªãããšã確èªããããšã§ãã
èªåã®æèŠãå
±æããã³ãã¥ãã±ãŒã·ã§ã³ããã³ã¡ã³ãããæ©äŒãéãããšã¯é£ãããããŸããã å®å
šã§å¿«é©ãªç°å¢ãšäººã
ãžã®æ¬æãæã£ãæ±ããéæããããšã¯ã¯ããã«å°é£ã§ãã ãããŠããããªããã°ã³ãã¥ããã£ã®çºå±ã¯ãããŸããã
åºæïŒ habr.com