Á¤º¸°úÇÐȸ³í¹®Áö (Journal of KIISE)
ÇѱÛÁ¦¸ñ(Korean Title) |
Ç¥Ãþ Á߸³È ±â¹ÝÀÇ ¾ð¾î ½ºÅ¸ÀÏ ÀüÀÌ |
¿µ¹®Á¦¸ñ(English Title) |
Language Style Transfer Based on Surface-Level Neutralization |
ÀúÀÚ(Author) |
ÃÖ¿ì¿ë
³ëÀ±¼®
¹Ú¼¼¿µ
Wooyong Choi
Yunseok Noh
Seyoung Park
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 47 NO. 02 PP. 0162 ~ 0169 (2020. 02) |
Çѱ۳»¿ë (Korean Abstract) |
°¨Á¤°ú °°Àº ÀÚ¿¬¾î ¹®ÀåÀÇ ½ºÅ¸ÀÏÀ» ÀüÀÌÇϱâ À§ÇØ ÇØ°áÇÒ ¹®Á¦´Â ¹®ÀåÀÇ ½ºÅ¸ÀÏÀ» ¾ø¾Ö´Â Á߸³È¿Í Á߸³ÈµÈ ¹®Àå¿¡ ½ºÅ¸ÀÏÀ» ÀÔÈ÷´Â ÀÛ¾÷ÀÌ´Ù. ±âÁ¸ ¿¬±¸¿¡¼´Â Àû´ëÀû ÇнÀÀ» ÅëÇؼ ÀáÀç °ø°£¿¡¼ Á߸³È¸¦ ¼öÇàÇß´Ù. ÇÏÁö¸¸ ÀÌ·± ¹æ½ÄÀº ¿ø·¡ÀÇ ³»¿ëÀ» À¯ÁöÇÏ¸é¼ ½ºÅ¸ÀÏÀ» ÀüÀÌÇÏ´Â °Í¿¡ ¾î·Á¿òÀ» °Þ´Â´Ù. º» ³í¹®¿¡¼´Â ÀáÀç °ø°£¿¡¼°¡ ¾Æ´Ñ Ç¥Ãþ ¼öÁØ¿¡¼ ½ºÅ¸ÀÏÀ» ¶ç´Â ´Ü¾î¸¦ Áö¿ì´Â °ÍÀ¸·Î Á߸³È¸¦ ¼öÇàÇÏ°í Áö¿î ´Ü¾îµéÀ» ÀûÀýÈ÷ ÀüÀÌµÈ ´Ü¾î·Î ¿¹ÃøÇØ º¹±¸ÇÏ´Â 2´Ü°è ¾ð¾î ½ºÅ¸ÀÏ ÀüÀÌ ¹æ¹ýÀ» Á¦¾ÈÇÑ´Ù. À̸¦ À§Çؼ ÀÚ±âÁÖÀÇ ±â¹Ý ºÐ·ù±âÀÇ È÷Æ®¸Ê°ú ´Ü¾î ¿¹Ãø±â¸¦ È°¿ëÇÑ´Ù. Á¦¾È ¸ðµ¨À» Æò°¡Çϱâ À§Çؼ Yelp¿Í Amazon ¸®ºä µ¥ÀÌÅͼÂ, ±×¸®°í Caption µ¥ÀÌÅͼÂÀ» È°¿ëÇØ ½ºÅ¸ÀÏ ÀüÀÌ ½ÇÇèÀ» ¼öÇàÇß´Ù. ÀÚµ¿ Æò°¡¿Í »ç¶÷ Æò°¡ °á°ú, Á¦¾È ¸ðµ¨ÀÌ ¿©·¯ Ãø¸é¿¡¼ ºñ±³ ¸ðµ¨º¸´Ù ³ôÀº ¼º´ÉÀ» º¸¿´´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
Two main concerns of language style transfer such as sentiment transfer are neutralization of a stylized sentence and re-stylization of the neutralized sentence with a target style. Generally, neutralization is accomplished by learning a neutralized latent space by adversarial learning. However, this neutralization method suffers from the difficulty of maintaining the original content after style transfer. In this paper, we propose a two-step language style transfer method comprised of a surface-level neutralization that removes style words and a target-style word prediction for the removed words. For this, a self-attentive style classifier and style-specific word predictors are used for the surface neutralization and style word generation, respectively. To evaluate the proposed method, several experiments of language style transfer were conducted with Yelp and Amazon review datasets and Caption dataset. As a result, the proposed method shows superior performance over baseline methods on various evaluation metrics including automatic and human evaluations.
|
Å°¿öµå(Keyword) |
¾ð¾î ½ºÅ¸ÀÏ ÀüÀÌ
Ç¥Ãþ Á߸³È
¹®¸Æ ±â¹Ý ´Ü¾î ¿¹Ãø
¸¶½ºÅ· ´Ü¾î ¿¹Ãø
language style transfer
surface-level neutralization
contextualized word
masked word prediction
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|