2017-01-01ãã1å¹´éã®è¨äºä¸è¦§
大æ¦æ¥ã§ããããæã¯è¨èªå¦çå¦ä¼å¹´æ¬¡å¤§ä¼ã¨åè«ã®å稿ãæ·»åããããããæã®æéããããã使ããªãã®ã§ãå¹´å ã«éã£ã¦ããã£ããã®ãè¿ãã¾ã§ã¯ãããããªãããæ¼ã«ãè¦èãã«è¡ã£ãããããå½ããåã§ããããå ¥é¢ãã¦ããã¨å¹´æ«å¹´å§ãé¢ä¿ãªã人ããâ¦
æ¨æ¥ã京é½åºå¼µã ã£ãã®ã§ãä»æ¥ããå¬ä¼ã¿ã§ãããã¨ã¯ãããæ¨æ¥ã®æ°å¹¹ç·ã®ä¸ã§ããã¯ãã ã£ãè¨èªå¦çå¦ä¼å¹´æ¬¡å¤§ä¼ã®å稿ï¼ç¬¬1稿ï¼ã®æ·»åãããã®ã§ãæããæ¼ã¾ã§æéãããã£ã¦5件è¿ããå稿ã ã¨ã大ä½1æ¬ããã1æéããªããä»å¹´ããã¯ãåºæ¬çã«å°çâ¦
æããæ°å¹¹ç·ã§äº¬é½ã¸ãå¹´æ«å¹´å§ã®å¸°çã©ãã·ã¥ã§å ¨ç¶åº§å¸ãåãããæå®å¸ã¯ãããã°ãªã¼ã³è»ã§ãæºå¸ãç¶ãã¦ãã¦ã³ã£ãããããããåãã座å¸ã ã£ãããã°ãªã¼ã³è»ã®é°å²æ°ãéã£ã¦ãã»ã¨ãã©æå®å¸ç¶æ ã§ãã£ããæéããªãã®ã§è»å ã§ãå¼å½ãé£ã¹ããâ¦
ååä¸ã¯ãªã³ã©ã¤ã³ã§ Google Hangout ã使ã£ã¦ B3 ã®ç 究室ã¤ã³ã¿ã¼ã³ã·ããã®ãã¼ãã£ã³ã°ãææ¥æ¥çã«ã¯ãããããªãã¦ããã®ã ããè¨èªå¦çå¦ä¼å¹´æ¬¡å¤§ä¼ã¾ã§ã®æ®ãæéãèããã¨ããã®æ¥ã«ã¿ããªã§éã¾ã£ã¦ç¸è«ããå¿ è¦ããã£ãã®ã§ããã10,000件ã®ã¢â¦
ä»æ¥ã¯é¦é½å¤§ç§èåãã£ã³ãã¹ã«ã¦å ±åç 究ã®ãã¼ãã£ã³ã°ãå°æ¹ããæ¥ãæ¹ã ã«ã好è©ã®ç«å°ã§ãããã£ãï¼æ±äº¬é§ ãã1é§ ã ããé§ åã ãï¼ãããããæç³»ã®äººãåå¤å±ã大éªãä»å°ãããã®ç«å°ã ã¨ãæ±äº¬ã«ä½ãã§ãã人ãçãããªãï¼ææ¥ãä¼è°ã®ããé±3æ¥â¦
ååä¸ã¯å¹´å æå¾ã®ä¿®è«ã»åè«ã®é²æå ±åããã¤ãè¨èªå¦çå¦ä¼å¹´æ¬¡å¤§ä¼ã®é²æå ±åã¨åæé²è¡ãªã®ã§æ ãã ãããã¾ããæè¿ã¯ä¿®è«ã®è³ªãå®å®ãã¦ãã¦ãé·ãã大ä½åãã«ãªã£ãããã¯ãªãªãã£ãï¼å°ãªãã¨ã1æ¬ã¯æ»èªä»ãã®å½éä¼è°ã«éãã¦ããå¦çãã¡ã«ãªã£â¦
ååä¸ãããè¨èªå¦çå¦ä¼å¹´æ¬¡å¤§ä¼ã«åããä»å¹´æå¾ã®é²æå ±åãèããä»å¹´ã¯10件ç¨åº¦ã®äºå®ï¼å°çºã®è² è·è»½æ¸ã®ããããã§ã«2件æ念ãã¦ããã£ããï¼ãç§æ¸ããã®åºå¤æ¥ã ã£ãã®ã§ããæ¼ã¯ã¹ã¼ãã¼äºåå¦çã¿ã¤ã ãæ¯å¹´ç¥è¦ãå¾ããã¦å¹çåããã¦ããã¯ããªâ¦
ååä¸ãã¡ãã£ã¨ãä»äºã§é§ åã®å«è¶åºã¾ã§è¡ããã¦ããããåæ¥ãç¥æ¥ã¯ä»äºããã«ãã¦ãã移åæéããã£ãããªãã®ã§ã大å¦ã«ã¯è¡ããããªãã®ã§ããã1ã¶æåã®ãã¨ã«ãªãã¨ãä½ãããã®ããã£ã±ãæãåºããªããå¤æ¹ã¯ãè¦èãã«è¡ããã¯ãªã¹ãã¹ãã¬ã¼â¦
ä»æ¥ã¯å¨ï¼3æ³7ã¶æï¼ã®éãä¿è²åã®ã楽ãã¿ä¼ï¼ãèå± ãªã©ã®ãæ«é²ç®ä¼ï¼ãå¨ãåå ããã®ã¯ãããã3åç®ã§ã0æ³ã1æ³ããã³2æ³ã®ã¨ãã¯ã»ãªãããªãæã«åããã¦è¸ãã ãï¼0æ³ã®ã¨ãã¯æ¬äººãè¸ãã®ã§ããªããä¿è²è ããã«æ±ããã¦ãã ãï¼ã ã£ãããä»å¹´â¦
ååä¸ã¯ãã¿ã¼ã³èªèã¨æ©æ¢°å¦ç¿ã®ææ¥ãã©ããå»å¹´ã¨æ¯ã¹ã¦è©±ãããã¨ãå°ãªããªãã¨æã£ã¦ããããã©ãããä»å¹´åº¦ã¯æ¨å¹´åº¦ã¨æ¯ã¹ã¦1ã³ãåå°ãªãï¼å»å¹´ã¯8ã³ããã£ãã®ã7ã³ããããªãï¼ããã§ãããããã話ãããã¨ãå°ãªãããã â¦â¦ãåå¾ã¯æ©æ¢°ç¿»è¨³åå¼·â¦
ä»æ¥ã¯ NL ç ã«ã¯è¡ããã«æ¥éãã£ã³ãã¹ã«åºå¤ã2é㯠B3 ã®ç 究室ã¤ã³ã¿ã¼ã³ã·ãããè¨èªå¦çå¦ä¼å¹´æ¬¡å¤§ä¼ã«æ稿ãããã¨ã«ãªã£ãã®ã§ããã¼ã¹é åã決ãã¦æ·¡ã ã¨ã¢ããã¼ã·ã§ã³ã»è°è«ã»åæããã¤ã¤ãå稿ãé²ãã¦ããããæ¼ä¼ã¿ã¯ B3 ã«åããç 究室é å±â¦
ä»æ¥ã¯ç¬¬4åèªç¶è¨èªå¦çã·ã³ãã¸ã¦ã ï¼æ å ±å¦çå¦ä¼ç¬¬234åèªç¶è¨èªå¦çç 究ä¼=NLç ï¼åå ã®ããã«æ©ç¨²ç°å¤§å¦ã¸ãããããä»»æ2å¹´ã2åãåè¨4å¹´ã®ç 究ä¼ã®å¹¹äºæ¥åã«å¾äºãã¦ããããä»åãæå¾ã§ã役御å ã§ãããæã«ä¼å ´ã«è¡ããä¼å ´ä¿ã®æ©å¤§ã®èæ± ããâ¦
ååä¸ã¯è«æç´¹ä»ãæ¬ å¸ããã¦ããããã¡ã¼ã«å¦çãä»é±ã¯ NL ç ãªã©è²ã ãã£ã¦ä»äºã®ã§ããæéããªããæ¼ããé²æå ±åãèããã1æéã§ã¯å ¨ç¶çµãããªããå士ã®å¦ä½è«æ審æ»ã®ç·´ç¿ãèãããå°ãããã³ã¡ã³ãã§ãããç³ã訳ãªããåå¾ã¯å ±åç 究ã®æã¡åâ¦
ååä¸ã¯è«æç´¹ä»ãèããè«æç´¹ä»ã®ã¹ã©ã¤ãã¯ååå ¬éãã¦ã»ããã¨ä¼ãã¦ããã®ã ããå ¬éãã¦ãããªã人ãå¤ããªããä»å¾å°±è·æ´»åãã人ã«ã¯ãå ¬éãããã¡ãªããããã¡ãªããã®æ¹ã大ããã¨æãã®ã ãã©ãæ¥å¹´åº¦ããã修士ã®å¦çã®ã¹ã©ã¤ãå ¬éã«ã¤ãã¦â¦
å¤æ¹ã¯å¨ã¨2人ã§ãè¦èãã«è¡ããä»é±ãããªãããªãã¼ã·ã§ã³å°éã®ç é¢ã«è»¢é¢ããããããåé£ãã§ç æ£ã«ä¸ãããªãã£ãã®ã§ç¦ã£ããããªãã¨ãé»è©±ã§é£çµ¡ãåã£ã¦ä¼ãã¦ããã£ãã夫婦ã¨ãã«ä»äºãããã¨ããããã¾ã§ã¯å®å®¶ã«ãé¡ããã¦ããããããã§ããªâ¦
ååä¸ã¯å¨ï¼3æ³7ã¶æï¼ã温æ³ã«è¡ãããã¨ããã®ã ããè¿æã®ã¹ã¼ãã¼é湯ã¯æ¹ä¿®å·¥äºä¸ãªã®ã§ãæ±ä¹ çç±³ã®ã¹ã¼ãã¼é湯ã¾ã§è¶³ã伸ã°ãï¼2åç®ï¼ãããæ°·ãé£ã¹ããã¨ããã®ã§é£ã¹ãããã©ããå¬ã«ããæ°·ã¯ãããªãã®ã§ã¯ãªããï¼ã¨ï¼ä»ããï¼æã£ããããï¼â¦
ååä¸ã¯ãã¿ã¼ã³èªèã¨æ©æ¢°å¦ç¿ã®ææ¥ã®2åç®ã2å¹´ç®ãªã®ã§ä½ãã©ãåãã°ãããã¯åãã£ã¦ãã¦ããã¤ããã ãããªããå»å¹´ã¨æ¯è¼ãã¦è¥å¹²é²åº¦ãé ããå»å¹´ã®æ¹ãããããè±ç·ãã¦ããæ°ãããã®ã ããä½ãéããã ããããæ¼ä¼ã¿ã®ååã¯å¦ä¼ã®ä»äºã§ Skypâ¦
ååä¸ã¯ B3 ã®ç 究室ã¤ã³ã¿ã¼ã³ã·ãããå ã®ãã¼ã¿ã¨å¦çå¾ã®ãã¼ã¿ãè¦ã¦ãããçæããããã°ã©ã ãä½æããï¼çæç¨ã®ããã°ã©ã ã¯ããå ¥æä¸è½ï¼ã¨ããéã¢ã»ã³ãã«ã¿ãããªãã¨ãããªãã¨ãããªãã®ã ããããããããã°ã©ã ãæ¸ãã®ãéçºããã¦ããâ¦
ååä¸ã¯å ±åç 究ã®ãã¼ãã£ã³ã°ãä»å¹´åº¦ã®ååã¯ããã£ãããå¾åã¯æã£ããããªå®é¨çµæãå¾ãããªãã£ãã®ãæ°ããããããããã¡ããã¨ä¼æ¥å ã®ãã¼ã¿ã«ç¹åããå 容ã®ãã¨ããããªããæ¬å½ã«ãå ±åãã§ï¼ã¤ã¾ãå æ¹ã«ãèªç¶è¨èªå¦çã«ãããã詳ãã人â¦
ååä¸ã¯è«æç´¹ä»ã«åºãã Adithya Renduchintala, Rebecca Knowles, Philipp Koehn, Jason Eisner. User Modeling in Language Learning with Macaronic Texts (ACL 2016). ãããã¿ã¹ã¯è¨å®ãåããã«ããã®ã ããä¸é¨ã ãæ¯èªã§æ¸ããã¦ãããããªããã¹â¦
ååä¸ãè«æç´¹ä»ãèããæ°å ¥çã®è«æèªè§£ã¹ãã«ã®ãã¬ã¼ãã³ã°ã«èª²é¡ãããã®ã ããè«æ100æ¬èªãã§ãããã®ã大å¦é¢ã«å ¥ã£ã¦ãããªã®ã«ãå¦é¨çã§ãå¾æã¯è«æç´¹ä»ããã¦ããã£ã¦ããã®ã§ãé çªãéã«ããæ¹ãããã®ãããã¨æã£ããããããã ãç´¹ä»ããâ¦
ä»é±ã¯æ¼ãããè¦èãã«è¡ããæ¼ã ã¨äººãå¤ããã¨æã£ã¦é»è»ã¨ã¿ã¯ã·ã¼ã§æ¥ãã®ã ããå®å ¨ã«è»ã§æ¥ãæ¹ãããã£ãæ¡ä»¶ããè¦èãã«ã¯å 夫婦ãæ¥ã¦ãã¦ããããªã«å®¶æãéã¾ãã®ãçãããã¨ããå æ¯ãªã®ã§ãè¨å¿µã«åçãæ®ããå¨ã¯å®å ¨ã«é¡ãé ãã¦ãããâ¦â¦â¦
ä»æ¥ã¯æãããã天æ°ã ã£ãã®ã§ãã¹ã§å祥寺ã¸ãå¨ï¼3æ³7ã¶æï¼ã¯é§ åã®ãã¯ãªååããæ°ã«ãªãããããä½åº¦ãè¦ã«ãããé åã«ãªã£ã¦ããã¨ãæ»ãã§ãããã¨å¤æããããã§ãããªãã§è±¡ããæ»ãããã£ãã®ï¼ãã¨èãããããããã¯ãªåã¯2016å¹´ã«æ»ãã ãâ¦
éææ¥ã®ååä¸ã¯ä»æ¥ããæ¯é±ãã¿ã¼ã³èªèã¨æ©æ¢°å¦ç¿ã®ææ¥ï¼ã®æ©æ¢°å¦ç¿é¨åï¼ã1åç®ã¯å®é¨ã«é¢ããè²ã ãªãã¨ã話ããã¨ã«ãã¦ãããã¨ããã®ãããã®ææ¥ã§è©±ããªãã¨ãè¨ç·´ãã¼ã¿ãéçºãã¼ã¿ããã¹ããã¼ã¿ãã¨ãããããªæ¦å¿µããç¥ããªãã¾ã¾ä¿®å£«ãåºâ¦
ååä¸ã¯ç 究室ã¤ã³ã¿ã¼ã³ã·ãããè¨èªå¦çå¦ä¼å¹´æ¬¡å¤§ä¼ã«åºããã¨ã«ãªã£ãããããã°ã©ã ãæ¸ãå¿ è¦ãããã®ã§ãä½ãæ¸ãã°ããããã¨ãããã¨ãæ´çããããããã¾ããèªç¶è¨èªå¦çã¯åå¦çå¾å¦çã ãããªã®ã§ããããç 究ã®ãã¡ããªã¨â¦â¦ããµã¼ã室ã«ç½®ãâ¦
ååä¸ã¯ EMNLP 2017 èªã¿ä¼ï¼ã¨ ACL 2017 èªã¿ä¼ï¼ãä»å¦æï¼ã¨ãããå¬å¦æï¼ã®ååã¯ææ¥ãããã®ã§å ¨ç¶åå ã§ããªãã£ãããä»åã¯åå ã§ããï¼æ¥é±ã¯å ±åç 究ã®ãã¼ãã£ã³ã°ã§åå ã§ããªããï¼ãèªãè«æããããã¾ã§ã¯ short paper ãï¼åéçãªæå³ã§â¦
æ¬æ¥ã¯ [twitter:@keiskS] ããã«æ¥éãã£ã³ãã¹ã«æ¥ã¦ããã£ã¦ãã¼ã¯ããã¦ããããã¡ããã©èªåãé¦é½å¤§ã«æ¥ãã¿ã¤ãã³ã°ã§å½¼ã Johns Hopkins University ã«é²å¦ãããã«ã¬ãªã¢éå¬ã® ACL 2013 ã§å½¼ã® Microsoft Research Asia ã§ã®ç 究ã®çºè¡¨ã§åæµããâ¦
ååä¸ãä¹ ãã¶ãã«è«æç´¹ä»ã«åºãã Sakaguchi et al. Grammatical Error Correction with Neural Reinforcement Learning. IJCNLP 2017. ããã深層強åå¦ç¿ãç¨ãã¦ææ³èª¤ãè¨æ£ãããç 究ã§ãGLEU ã¨ããæã«å¯¾ããè©ä¾¡å°ºåº¦ã§èª¤ãè¨æ£ãæé©åããããã«â¦
ååä¸ã諸äºæ ã§æ¨æ¥ã«ã§ããªãã£ãä»äºãããã¦ããããæ¯ã«å¨ãè¦ã¦ãããã¤ããã ã£ãããæ¥ã«ç¡çã«ãªã£ãã®ã§ã妻ã«å¨ãè¦ã¦ããããæ¥ææ¥ã«æéã確ä¿ããã¦ããã£ãã®ã§ãããä»åã®ä»äºèªä½ã¯ç¡äºçµäºããããåºã¾ã£ãæéããªãã¨ã§ããªãä»äºãâ¦
å¾æã¯å¦»ãåææ¥ã®åå¾ã«ææ¥ãããããã°ããï¼æ°ã¶æåä½ï¼ï¼æ¯ã«ãå¨ããé¡ãã§ããªãè¦éããªã®ã§ãæ¯é±ã¯ã³ãªãè²å ã§ãããååä¸ã¯å°éäºå ¬åã«è¡ããå¨ï¼3æ³åï¼ã¯ããæ°·ãé£ã¹ããã£ãããã ãããããã«å£²åºã§ã売ããªããªã£ã¦ãã¾ã£ãããã§ãå¨â¦