• 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 : 8 / 17 ÀÌÀü°Ç ÀÌÀü°Ç   ´ÙÀ½°Ç ´ÙÀ½°Ç

ÇѱÛÁ¦¸ñ(Korean Title) Bidirectional GRU-CRF ±â¹ÝÀÇ Çѱ¹¾î °³Ã¼¸í ÀνÄÀ» À§ÇÑ ¾îÈÖ »çÀü ÀÚÁú Àû¿ë ³×Æ®¿öÅ© ÅäÆú·ÎÁö ¿¬±¸
¿µ¹®Á¦¸ñ(English Title) Research on the Various Neural Network Topologies for Korean NER based on Bidirectional GRU-CRF applying Lexicon Features
ÀúÀÚ(Author) ±è¼±¿ì   ÃÖ¼ºÇÊ   Seon-Wu Kim   Sung-Pil Choi  
¿ø¹®¼ö·Ïó(Citation) VOL 25 NO. 02 PP. 0099 ~ 0105 (2019. 02)
Çѱ۳»¿ë
(Korean Abstract)
°³Ã¼¸í ÀνÄÀº ¹®Çå ³»¿¡ Ç¥ÇöµÈ °³Ã¼¸íÀ» ½Äº°ÇÏ´Â °úÁ¤À¸·Î, °³Ã¼¸í °£ÀÇ °ü°è¸¦ ÅëÇÑ Á¤º¸ÃßÃâ °úÁ¤¿¡ ¼±ÇàµÇ¾î¾ß ÇÏ´Â Áß¿ä °úÁ¤ÀÌ´Ù. ÃÖ±Ù ½ÉÃþÇнÀ ±â¼úÀ» °³Ã¼¸í ÀνĿ¡µµ Àû¿ëÇÏ´Â ¿¬±¸°¡ È°¹ßÈ÷ ÀÌ·ïÁö°í ÀÖÀ¸¸ç, °³Ã¼¸í »çÀü Á¤º¸¸¦ ¸ðµ¨ ¿ÜºÎ¿¡¼­ Ãß°¡ÀûÀ¸·Î È°¿ëÇÏ¿© °³Ã¼¸í ÀνĿ¡ ´ëÇÑ ¼º´ÉÇâ»óÀÌ ÀÌ·ïÁö°í ÀÖ´Ù. ±×·¯³ª °³Ã¼¸í »çÀüÀ̶ó´Â ÁÖ¿ä ÀÚÁú Á¤º¸¸¦ ½ÉÃþÇнÀ ¸ðµ¨ ³»ºÎ¿¡ Àû¿ëÇÏ´Â ¹æ½Ä¿¡ ´ëÇÑ ¿¬±¸´Â ã¾Æº¸±â ¾î·Æ´Ù. ÀÌ¿¡ º» ³í¹®Àº °³Ã¼¸í »çÀü Á¤º¸¸¦ ½ÉÃþÇнÀ ¸ðµ¨¿¡ Àû¿ëÇÏ´Â ¹æ¹ý·ÐÀ» ÃÑ 4°¡Áö·Î ±¸¼ºÇÏ°í, ÀÌ¿¡ µû¶ó ÅäÆú·ÎÁö¸¦ °¢°¢ ±¸¼ºÇÏ°í ºñ±³ºÐ¼®ÇÏ¿© °¡Àå ÀûÇÕÇÑ °³Ã¼¸í »çÀü Á¤º¸ Àû¿ë ½ÉÃþÇнÀ ÅäÆú·ÎÁö¸¦ µµÃâÇÑ´Ù. 2016³â ±¹¾îÁ¤º¸Ã³¸®´ëȸÀÇ °³Ã¼¸í ÀÎ½Ä ÄÚÆÛ½º¸¦ È°¿ëÇÑ ½ÇÇè°á°ú, ÇÕ»ê Àû¿ë ÅäÆú·ÎÁö°¡ 80.22%ÀÇ F1 Á¡¼ö¸¦ º¸À̸ç, °¡Àå ³ôÀº ¼º´ÉÀ» ³ªÅ¸³»¾ú´Ù. ÀÌ´Â °³Ã¼¸í »çÀü Á¤º¸¸¦ Àû¿ëÇÏÁö ¾ÊÀº ¸ðµ¨¿¡ ºñÇØ ¾à 1.5% ³ôÀº ¼º´ÉÀ̸ç, ÈÄó¸® Àû¿ë ¹æ¹ý¿¡ ºñÇØ 0.9% ³ôÀº ¼º´ÉÀÌ´Ù.
¿µ¹®³»¿ë
(English Abstract)
Named entity recognition is a process that identifies named entity expressed in a document. It is an important process that must be preceded by the process of extracting information through relationships between entities. Recently, studies have been actively done to apply the deep-learning technology to named entity recognition. The performance of the named entity recognition was improved by using the lexicon information outside the model. Nevertheless, it is difficult to find a method to apply the main feature information that is known as the lexicon to the deep-learning model. In this study, we constructed four methodologies in which we can apply the lexicon information to the deep-learning model. The study also constructs the topology, and compares and analyzes each methodology to derive the best in deep-learning topology through the application of the lexicon information. Experiments using the named entity recognition corpus of the Korean Language Information Processing Conference in 2016. The lexicon information for the input sentence was configured as a binary vector and input form as input to the CRF (conditional random field) before inputting into CRF. The topology showed the highest performance with a score of 80.22% F1. This is about 1.5% higher performance than the model without the lexicon information. Moreover, it is generally 0.9% higher than the method of applying the lexicon in post-processing.
Å°¿öµå(Keyword) °³Ã¼¸í ÀνĠ  Á¤º¸ ÃßÃâ   ÅؽºÆ® ¸¶ÀÌ´×   ÀÚ¿¬¾î 󸮠  ½ÉÃþÇнÀ   NER(Named Entity Recognition)   information extraction   text mining   NLP(Natural Language Processing)   deep-learning  
ÆÄÀÏ÷ºÎ PDF ´Ù¿î·Îµå