• 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 ¼Õ½ÇÇÔ¼ö¿Í ¹Ýº¹ Æä³ÎƼ¸¦ Ãß°¡ÇÑ Æ®·£½ºÆ÷¸Ó ÀÎÄÚ´õ-µðÄÚ´õ Á¦¸ñ »ý¼º ¸ðµ¨
¿µ¹®Á¦¸ñ(English Title) Transformer Encoder-Decoder based Title Generation Model with Word Loss and Repetition Penalty
ÀúÀÚ(Author) ¼º¼öÁø   Â÷Á¤¿ø   Su-Jin Seong   Jeong-Won Cha                             
¿ø¹®¼ö·Ïó(Citation) VOL 27 NO. 04 PP. 0210 ~ 0215 (2021. 04)
Çѱ۳»¿ë
(Korean Abstract)
Á¦¸ñÀº ¹®¼­¸¦ ´ëÇ¥ÇÏ´Â ¾î±¸ ȤÀº ¹®ÀåÀ̶ó Á¤ÀÇÇÒ ¼ö ÀÖ´Ù. ¿ì¸®´Â ¹®¼­ÀÇ Á¦¸ñÀ» »ý¼ºÇϱâ À§ÇØ Æ®·£½ºÆ÷¸Ó ±â¹Ý ÀÎÄÚ´õ-µðÄÚ´õ ±¸Á¶¸¦ Á¦¾ÈÇÑ´Ù. ´ë¿ë·® ¹®¼­¸¦ ÀÌ¿ëÇÏ¿© Æ®·£½ºÆ÷¸Ó ÀÎÄÚ´õ-µðÄÚ´õ ±¸Á¶ÀÇ »çÀüÇнÀ(pre-training)À» ÁøÇàÇÏ°í º»¹®°ú Á¦¸ñ ½ÖÀ¸·Î ÀÌ·ç¾îÁø ¹®¼­¸¦ ÀÌ¿ëÇÏ¿© ¹Ì¼¼Á¶Á¤ (fine-tuning)À» ÁøÇàÇÏ¿´´Ù. ¶ÇÇÑ Á¦¸ñ »ý¼º ŽºÅ©·Î ¹üÀ§°¡ Á¦ÇѵǴ ¹Ì¼¼Á¶Á¤ °úÁ¤¿¡¼­ ÀÔ·Â ¹®¼­¿¡ ³ªÅ¸³ª´Â ¾îÀýÀÇ »ý¼º ºñÀ²À» Áõ°¡½ÃÅ°±â À§ÇØ ´Ü¾î ¼Õ½ÇÇÔ¼ö¸¦ Ãß°¡ÇÏ°í ÅäÅ«ÀÌ ¹Ýº¹ÀûÀ¸·Î »ý¼ºµÇ´Â ¹®Á¦¸¦ °³¼±Çϱâ À§ÇÑ ¹Ýº¹ ÆгÎƼ¸¦ ¸ðµ¨ Ãß°¡ÇÏ´Â ¹æ¹ýÀ» Á¦¾ÈÇÑ´Ù. 25,564°³ÀÇ ³í¹® µ¥ÀÌÅ͸¦ »ç¿ëÇÑ ½ÇÇè¿¡¼­ ´Ü¾î ¼Õ½ÇÇÔ¼ö¿Í ¹Ýº¹ ÆгÎƼ¸¦ °³º°ÀûÀ¸·Î Àû¿ë½ÃŲ ¸ðµ¨ÀÇ ¼º´ÉÀÌ ±âÁ¸ ¸ðµ¨¿¡ ºñÇØ °³¼±µÇ°í, µÎ Á¦¾È ¹æ¹ýÀ» ¸ðµÎ Àû¿ëÇÑ ¸ðµ¨¿¡¼­´Â Rouge-LÀÇ ¼º´ÉÀÌ 2.7% Çâ»óµÇ´Â °ÍÀ» È®ÀÎÇÏ¿´´Ù.
¿µ¹®³»¿ë
(English Abstract)
The title can be defined as a phrase or sentence the represents the document. We propose a transformer encoder-decoder model to generate the title of the document. The transformer model is pre-trained based on a the usage of a large document, and fine-tuning is performed using the data comprising of the body and title. Also, in the fine-tuning process, the scope of which is limited to the title generation task, a Word Loss is added to increase the generation ratio of words appearing in the input document and ground truth title. We propose a method of adding a Repeat Penalty to the model to reduce the problem that tokens are repeatedly generated. In an experiment conducted using data from 25,564 papers, the performance of the model that individually applied the Word Loss and the Repeat Penalty was improved compared to the baseline. It was confirmed that Rouge-L's performance improved by 2.7% in the model to which both the proposed methods were applied.
Å°¿öµå(Keyword) Æ®·£½ºÆ÷¸Ó ÀÎÄÚ´õ-µðÄÚ´õ ¸ðµ¨   ÀÚµ¿ Á¦¸ñ »ý¼º   ÀÚµ¿ Á¦¸ñ »ý¼º   ´Ü¾î ¼Õ½ÇÇÔ¼ö   ¹Ýº¹ Æä³ÎƼ   transformer encoder-decoder   automatic title generation   word loss   repeat penalty        
ÆÄÀÏ÷ºÎ PDF ´Ù¿î·Îµå