• Àüü
  • ÀüÀÚ/Àü±â
  • Åë½Å
  • ÄÄÇ»ÅÍ
´Ý±â

»çÀÌÆ®¸Ê

Loading..

Please wait....

±¹³» ³í¹®Áö

Ȩ Ȩ > ¿¬±¸¹®Çå > ±¹³» ³í¹®Áö > Çѱ¹Á¤º¸°úÇÐȸ ³í¹®Áö > Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)

Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)

Current Result Document :

ÇѱÛÁ¦¸ñ(Korean Title) ´Ü¾î ÀÓº£µù ºÐ¼®À» ÅëÇÑ ½Å°æ¸Á ±â°è ¹ø¿ª ½Ã½ºÅÛÀÇ ¼º´É ºñ±³: Çѱ¹¾î-ÀϺ»¾î, Çѱ¹¾î-¿µ¾î¸¦ Áß½ÉÀ¸·Î
¿µ¹®Á¦¸ñ(English Title) A Performance Comparison of Neural Machine Translation Systems through Vocabulary sets and Word Embeddings: Focus on Korean-English and Korean-Japanese
ÀúÀÚ(Author) ÃÖ¿ë¼®   ¹Ú¿äÇÑ   À±½Â   ±è»óÈÆ   ÀÌ°øÁÖ   Seung Yun   Sanghun Kim   Kong Joo Lee   Yong-Seok Choi   Yo-Han Park  
¿ø¹®¼ö·Ïó(Citation) VOL 28 NO. 02 PP. 0081 ~ 0088 (2022. 02)
Çѱ۳»¿ë
(Korean Abstract)
º» ¿¬±¸¿¡¼­´Â MASS¸¦ ÀÌ¿ëÇØ »çÀü ÇнÀ ¸ðµ¨À» ±¸ÃàÇÏ°í º´·Ä µ¥ÀÌÅÍ·Î ÆÄÀÎ Æ©´×ÇÏ¿© Çѱ¹¾î-¿µ¾î¿Í Çѱ¹¾î-ÀϺ»¾î ±â°è ¹ø¿ª ¸ðµ¨À» ±¸ÃàÇÑ´Ù. Çѱ¹¾î, ÀϺ»¾î, ¿µ¾î´Â ¸ðµÎ ´Ù¸¥ ¹®ÀÚ Ã¼°è¸¦ »ç¿ëÇÑ´Ù. Çѱ¹¾î¿Í ÀϺ»¾î´Â ÁÖ¾î-¸ñÀû¾î-µ¿»çÀÇ ¾î¼øÀ» °®´Â ¹Ý¸é ¿µ¾î´Â ÁÖ¾î-µ¿»ç-¸ñÀû¾îÀÇ ¾î¼øÀ» °®´Â´Ù. º» ¿¬±¸¿¡¼­´Â ½Å°æ¸Á ±â¹ÝÀÇ ±â°è ¹ø¿ªÀ» ±¸ÃàÇÒ ¶§ µÎ ¾ð¾î »çÀÌÀÇ ¹®ÀÚ Ã¼°è¸¦ °øÀ¯ÇÏ´Â ¿©ºÎ¿Í ¹®Àå ¾î¼øÀÇ À¯»ç¼º¿¡ µû¸¥ ±â°è¹ø¿ªÀÇ ¼º´ÉÀ» Æò°¡ÇØ º¸¾Ò´Ù. ¸ðµ¨ÀÇ ¼º´É Â÷À̸¦ ´Ü¾î ÀÓº£µùÀ» ÅëÇØ ºÐ¼®ÇØ º¸±â À§ÇØ ¾îÈÖ ¹ø¿ª ½ÇÇè°ú ¹®Àå °Ë»ö ±â°è ¹ø¿ª ½ÇÇèÀ» ¼öÇàÇÏ¿´´Ù. ½ÇÇè °á°ú ÀÎÄÚ´õÀÇ ´Ü¾î ÀÓº£µùÀÌ µðÄÚ´õ¿¡ ºñÇØ ÈξÀ Áß¿äÇÏ°í Çѱ¹¾î-¿µ¾îº¸´Ù´Â Çѱ¹¾î-ÀϺ»¾îÀÇ °æ¿ì ´õ ÁÁÀº ¼º´ÉÀ» ¹ßÈÖÇÔÀ» ¾Ë ¼ö ÀÖ¾ú´Ù. ¹®Àå °Ë»ö ±â°è ¹ø¿ª ½ÇÇè¿¡¼­ Çѱ¹¾î-¿µ¾îÀÇ °æ¿ì¿¡´Â ¼Ò·®ÀÇ º´·Ä µ¥ÀÌÅ͸¸À¸·Îµµ Å« ÆøÀÇ ¼º´É Çâ»óÀÌ °üÂûµÇ¾ú´Ù.
¿µ¹®³»¿ë
(English Abstract)
In this study, we have pre-trained MASS models and built neural machine translation (NMT) systems for Korean-English and Korean-Japanese based on them. Korean, Japanese, and English use different writing scripts. Korean and Japanese are Subject-Object-Verb languages, while English is a Subject-Verb-Object language. In this study, we have evaluated the performances of NMT systems according to the similarity between languages, such as word order and writing scripts. To compare the performances of NMT models from the perspective of word embeddings, we have conducted the following two experiments: word translation and sentence translation retrieval using word embedding learned by NMT models. The accuracies of word translation and sentence translation for word embeddings of a Korean-Japanese NMT model were higher than those of a Korean-English pair. Moreover, the word embeddings learned by an encoder were more important than those learned by a decoder when used in NMT. Based on the result of sentence translation retrieval of a Korean-English NMT model, we found that a Korean-English unsupervise NMT model could be significantly improved when trained even with a small amount of parallel data.
Å°¿öµå(Keyword) MASS   ±â°è ¹ø¿ª   ´Ü¾î ÀÓº£µù   ¹®ÀÚ Ã¼°è   SVO ¼ø¼­   SOV ¼ø¼­   MASS   machine translation   word embedding   scripts   SVO order   SOV order  
ÆÄÀÏ÷ºÎ PDF ´Ù¿î·Îµå