ªñ¦~¾÷¾¹Â½Ä¶¡]machine translation¡^µo®i½´«k¡A¥«±¤W¥X²{¦UºØ¦Û°Ê½Ķ¨t²Î¡A¦pºô¤W½Ķ¨t²Î¤Î¤â¾÷½Ķµ{¦¡µ¥¡C¯S§O¬O¹L¥h¤@¨â¦~¶¡¡A°£¤F°ò©ó»y¨¥³W«h¤Î²Îp¤èªkªºÂ½Ä¶¨t²Î¥~¡A¶}µo°Ó³°Äò±À¥X¥H²`«×¾Ç²ß¡]deep learning¡^¤Î»¼Âk¯«¸gºôµ¸¡]recurrent neural network¡^¬°°ò¦ªº¯«¸g¾÷¾¹Â½Ä¶¨t²Î¡]neural machine translation systems¡^¡C
¦p¦¹¤@¨Ó¡A¥Î®a¿ï¾Ü§ó¦h¡A§Ú̦p¦óµûŲ¦U®a¨t²Î¡A¥[¥H¤ñ¸û¡A¾ÜÀu¦Ó¨ú¡A§ó§Î«n¡C¬°¦¹¥»¤å·§zY¤zµû»ù¤èªk¡A¨Ñ¦U¬É°Ñ¦Ò¡C
¡u¦^Ķ¡v¤ñ¸ûì¤å ¥¼¥²¥i¾a
³Ì²³æªº¤èªk¬O¡u¦^Ķ¡v¡]round-trip translation¡^¡A¥ý§Q¥Î¾÷¾¹Â½Ä¶¨t²Î§âì¤åĶ¥X¡AµM«á¦A¸g¹q¸£§âĶ¤åĶ¦^ì¨Óªº»y¨¥¡A¦p¦¹¤ñ¸û³Ìªìªºì¤å¤Î¸g¦^Ķ«áªº¡uì¤å¡v¡AY¨âªÌ¬Û¦ü¡A«h¨t²Îªí²{¸û¨Î¡A§_«h¥¼¦p²z·Q¡C¦¹ªk§Ö±¶Â²«K¡A¥Î®a§Y¨Ï¤£À´Ä¶»y¡A¤´¥i¾Ú¦¹¦ôp¹q¸£Ä¶¤å¬O§_ç´¡C
¤£¹L¡A¨t²Î§âĶ¤åĶ¦^ì»y®É¡AĶ¤å¤¤ªº¿ù»~©Î¦³¼W´î¡A¬G¨â½g¡uì¤å¡v¬Û¦ü»P§_¡A¥¼¥²¦b¥ô¦ó±¡ªp¤U§¡¯à¦p¹ê¤Ï¬M½Ķ¨t²Î¤§°ª¤U¡C
n§ó·Ç½T¦aµû»ù½Ķ¨t²Î¡A§ÚÌ¥i³]¥ßµû¤À·Ç«h¡AµM«áÁܽеû¼fû«ö¤§¬°¹q¸£Ä¶¤åµû¤À¡C¬ü°ê¤W¥@¬ö60¦~¥N³]¥ß»y¨¥¦Û°Ê³B²z¿Ô¸ß©eû·|¡]Automatic Language Processing Advisory Committee¡^¡A¬ã¨s«X^¦Û°Ê½Ķ¤ô¥¡A°µªk¥i¨ÑÉŲ¡C
·í§½´NĶ¤å·Ç½T¤Î¥iŪµ{«×¨â¤è±³]¥ßµû¤À¼Ð·Ç¡A¤À³]9Ó¦Ü10Óµ¥¯Å¡A´yz¦U¯ÅĶ¤åÀ³¦³¤ô¥¡C·í®ÉÁÙ¸u½Ð¤F¨â²Õµû¤À¤Hû¡A¤@²Õ¬°^»y¤Hû¡A¥t¤@²Õ«hݳq^«XÂù»y¡F©Ò¦³¤Hû§¡¶·´N¨â¤è±µû¤À¡A½T«O¤À¼Æ«ÈÆ[·Ç½T¡C
¤H¤âµû¤À»Ý®É¯Ó¶O¤j
µM¦Ó¡A¤H¤âµû¤À©¹©¹»Ýn¤£¤Ö®É¶¡¤Î¶}¤ä¡A¦Û°ÊªºµûŲ¤èªk¹E¦¨¼é¬y¡C¦Û°ÊµûŲ§Y¹q¸£¦Û¦æµû¦ô¨t²Î½è¯À¡A¤@¯ë¥H°Ñ¦ÒĶ¤å¡]¦p¤H¤u½ĶªºÄ¶¤å¡^¬°µû¤À¼Ð·Ç¡C¹q¸£¤ñ¸û¨t²Î¿é¥X¤Î°Ñ¦ÒĶ¤å¡A«ö¬Û¦ü«×¦Û°Êµû¤À¡A¨âªÌ¬Û¦ü«×°ª«hµ¹¤©¸û²z·Q¤À¼Æ¡C±`¨£pºâ¤èªk¦³«Ü¦h¡A¥]¬Aºë½T²v¡B¥l¦^²v¡BFÈ¡BBLEU¤ÎNISTµ¥¡Aºô¤W¬ÛÃö¸ê·½Â×´I¡AŪªÌ¥i¦Û¦æ°Ñ¦Ò¡C
ȱoª`·Nªº¬O¡A½Ķ¨t²Îªº¦UºØµûŲ¤èªk¡A°£¤FÀ°§U¥Î®a¬D¿ï¤Îµ½¥Î¨t²Î¥~¡A¹ï¾÷¾¹Â½Ä¶¥»¨µo®i¤]¦³ÃöÁä§@¥Î¡C§ó¨ãÅé¦Ó¨¥¡A·s¤@¥N½Ķ¨t²Î±j½Õ¦Û¦æ¾Ç²ß¤Î´£¤É½Ķ¯à¤O¡A¹q¸£«ç¼Ë·Ç½Tµû®Ö½Ķ¤ô¥¡A¦Û°Ê§ï¶i®zÂI¡A¹ê¦b«D±`«n¡Aȱo²`¤J±´°Q¡C¡½ùڥͺ޲z¾Ç°|°Ó°È½Ķ½Òµ{Á`ºÊ ¿½¥@©÷³Õ¤h
¹j¬P´Á¤T¨£³ø