Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
´ÙÁß ÀÎÄÚ´õ Transformer ±â¹Ý ¹ø¿ª¹® ÀÚµ¿ »çÈÄ ±³Á¤ ¸ðµ¨ÀÇ µðÄÚ´õ ÁÖÀÇ ±¸Á¶ ¿¬±¸ |
¿µ¹®Á¦¸ñ(English Title) |
Research on the Decoder Attention Structure of Multi-encoder Transformer-based Automatic Post-Editing Model |
ÀúÀÚ(Author) |
½ÅÀçÈÆ
ÀÌ¿ø±â
±è¿µ±æ
ÇÑÈ¿Á¤
ÀÌÁ¾Çõ
Jaehun Shin
WonKee Lee
Youngkil Kim
H.Jeung Han
Jong-Hyeok Lee
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 26 NO. 08 PP. 0367 ~ 0372 (2020. 08) |
Çѱ۳»¿ë (Korean Abstract) |
¹ø¿ª¹® ÀÚµ¿ »çÈÄ ±³Á¤Àº ±â°è ¹ø¿ª ½Ã½ºÅÛÀÇ °á°ú¹°À» ÀÚµ¿À¸·Î ±³Á¤ÇÏ¿© ´õ ³ªÀº ¹ø¿ª¹®À» ¸¸µé¾î³»´Â °úÁ¤À¸·Î, ±â°è ¹ø¿ª ½Ã½ºÅÛ ¿ÜÀûÀ¸·Î ±â°è ¹ø¿ªÀÇ Ç°Áú Çâ»óÀ» ¼öÇàÇϱâ À§ÇØ Á¦¾ÈµÈ ¿¬±¸ ºÐ¾ßÀÌ´Ù. º» ³í¹®¿¡¼´Â ¹ø¿ª¹® ÀÚµ¿ »çÈÄ ±³Á¤ ¹®Á¦¿¡ »ç¿ëµÇ´Â ´ÙÁß ÀÎÄÚ´õ Transformer ±â¹Ý ±³Á¤ ¸ðµ¨ÀÇ ±âº» ±¸Á¶¸¦ »ìÆ캻 µÚ, µðÄÚ´õ¿¡¼ ÀÎÄÚ´õ Ãâ·Â°úÀÇ »óÈ£ ÀÇÁ¸¼ºÀ» ´ã´çÇÏ´Â ÁÖÀÇ ±¸Á¶¸¦ ´Ù¾çÇÏ°Ô ±¸¼ºÇÏ°í Àû¿ëÇØ º¸¾Ò´Ù. WMT18 »çÈÄ ±³Á¤ ¸»¹¶Ä¡¸¦ ÀÌ¿ëÇÑ ½ÇÇè¿¡¼´Â ´ÙÁß ÀÎÄÚ´õ Transformer¸¦ ÀÌ¿ëÇÑ ¸ðµ¨ ÀüºÎ°¡ ±â°è ¹ø¿ª ½Ã½ºÅÛÀÇ °á°ú¹°¿¡ ºñÇØ ´õ ³ªÀº Ç°ÁúÀÇ ¹®ÀåÀ» »ý¼ºÇÏ¿´À¸¸ç, µðÄÚ´õ ³»¿¡ ¹ø¿ª¹®¿¡ ¿ø¹®ÀÇ ¹®¸Æ Á¤º¸¸¦ ¹Ý¿µÇÏ´Â ±¸Á¶¸¦ Àû¿ëÇÏ´Â °ÍÀÌ »çÈÄ ±³Á¤ÀÇ ¼º´É Çâ»ó¿¡ Å©°Ô ±â¿©ÇÔÀ» È®ÀÎÇÒ ¼ö ÀÖ¾ú´Ù
|
¿µ¹®³»¿ë (English Abstract) |
Automatic Post-Editing (APE) is a study on the correction for the output of machine translation (MT) system for the purpose of improving its translation quality independent of the MT system itself. In this paper, we examine the basic architecture of a multi-encoder transformer-based APE model, and implement several variants of the system¡¯s encoder-decoder attention layer which takes the outputs of the multi-encoder as its inputs. In experiments with the WMT18 APE data, all variations on our model successfully improve the translation quality of the original MT outputs. In particular, we find modeling the attention to incorporate source sentence context into the translated sentence improves post-editing performance.
|
Å°¿öµå(Keyword) |
±â°è ¹ø¿ª
¹ø¿ª¹® ÀÚµ¿ »çÈÄ ±³Á¤
Æ®·£½ºÆ÷¸Ó
´ÙÁß ÀÎÄÚ´õ ±¸Á¶
ÁÖÀÇ ÁýÁß ±â¹ý
machine translation
automatic post-editing
transformer
attention mechanism
multi-encoder architecture
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|