Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
±â¾ïÀçÇö Áö¼ÓÇнÀÀ» È°¿ëÇÑ ´ÙÁß µµ¸ÞÀÎ ÀÀ´ä »ý¼º |
¿µ¹®Á¦¸ñ(English Title) |
Multi-Domain Response Generation Using Memory-Replay Continual Learning |
ÀúÀÚ(Author) |
¹ÚÇüÁØ
È«Ãæ¼±
¹Ú¼º¹è
¼ÛÇöÁ¦
Hyeong-Jun Park
Choong Seon Hong
Seong-Bae Park
Hyun-Je Song
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 49 NO. 03 PP. 0153 ~ 0159 (2022. 03) |
Çѱ۳»¿ë (Korean Abstract) |
´ÙÁß µµ¸ÞÀÎ ÀÀ´ä »ý¼ºÀº Çϳª ÀÌ»óÀÇ µµ¸ÞÀο¡ ´ëÇØ ÀԷµǴ ¹ßÈ¿¡ ÀÀ´äÀ» »ý¼ºÇÏ´Â ¹®Á¦ÀÌ´Ù. ¿©·¯ µµ¸ÞÀÎÀ» ´ëÀÀÇØ¾ß Çϴ ȯ°æ¿¡¼ ´ÙÁß µµ¸ÞÀÎ ÀÀ´ä »ý¼ºÀº ±âÁ¸¿¡ ÇнÀµÈ µµ¸ÞÀλӸ¸ ¾Æ´Ï¶ó »õ·ÎÀÌ Ãß°¡µÇ´Â µµ¸ÞÀο¡ ´ëÇؼµµ ÀûÀýÈ÷ ÀÀ´äÇÒ ¼ö ÀÖ¾î¾ß ÇÑ´Ù. ±âÁ¸ ¿¬±¸µéÀº ±âÇнÀµÈ ÀÀ´ä »ý¼º ¸ðµ¨¿¡ »õ·ÎÀÌ Ãß°¡µÇ´Â µµ¸ÞÀÎÀÇ µ¥ÀÌÅÍ·Î ¹Ì¼¼ Á¶Á¤ÇÏ¿´´Ù. ÇÏÁö¸¸, À§ ¹æ¹ýµéÀº ±âÇнÀµÈ µµ¸ÞÀο¡ ´ëÇØ ÀûÀýÈ÷ ÀÀ´äÀ» »ý¼ºÇÏÁö ¸øÇϰųª µµ¸ÞÀΰú ¹«°üÇÑ ÀϹÝÀûÀÎ ´ë´äÀ» »ý¼ºÇÏ´Â ¹®Á¦°¡ ÀÖ´Ù. º» ³í¹®¿¡¼´Â ±â¾ïÀçÇö Áö¼ÓÇнÀ ±â¹Ý ´ÙÁß µµ¸ÞÀÎ ÀÀ´ä »ý¼º ¸ðµ¨À» ±¸ÃàÇÏ´Â ¹æ¹ýÀ» Á¦¾ÈÇÑ´Ù. ±â¾ïÀçÇö Áö¼ÓÇнÀ ±â¹Ý ÀÀ´ä »ý¼º ¸ðµ¨Àº ÀÌÀü µµ¸ÞÀÎ ¹ßÈ-ÀÀ´ä µ¥ÀÌÅÍ Áß ´ëÇ¥ µ¥ÀÌÅ͸¦ ¸Þ¸ð¸®¿¡ ÀúÀåÇصΰí, »õ·Î¿î µµ¸ÞÀÎÀÇ µ¥ÀÌÅÍ¿Í ¸Þ¸ð¸®¿¡ ÀúÀåµÈ µ¥ÀÌÅ͸¦ ÇÔ²² È°¿ëÇÏ¿© ÇнÀÇÑ´Ù. ±âÇнÀµÈ µµ¸ÞÀÎ Áß ÀϺκÐÀ» ¸Þ¸ð¸®¿¡ ÀúÀå ¹× ÇнÀ¿¡ È°¿ëÇÏ¿© »õ·Î¿î µµ¸ÞÀÎÀ» À§ÇØ ¸ðµ¨À» °»½ÅÇÒÁö¶óµµ µµ¸ÞÀο¡ Æ¯ÈµÈ ´ë´äÀ» »ý¼ºÇÒ ¼ö ÀÖ´Ù. µµ¸ÞÀο¡ Æ¯ÈµÈ ´ëÇ¥ µ¥ÀÌÅ͸¦ »ý¼ºÇϱâ À§ÇØ º» ³í¹®¿¡¼´Â k-Æò±Õ ±ºÁýȸ¦ È°¿ëÇÑ´Ù. ½ÇÇè¿¡¼ ÃÑ 5°³ µµ¸ÞÀο¡ ´ëÇØ ¼øÂ÷ÀûÀ¸·Î ÇнÀÇÏ°í ¸ðµç µµ¸ÞÀο¡ ´ëÇØ ÀÀ´äÀ» Á¦´ë·Î »ý¼ºÇÏ´ÂÁö¸¦ Æò°¡ÇÏ¿´´Ù. ½ÇÇè °á°ú Á¦¾ÈÇÑ ÀÀ´ä »ý¼º ¸ðµ¨ÀÌ ¸ðµç µµ¸ÞÀο¡¼ ±âÁ¸ ºñ±³ ¸ðµ¨º¸´Ù ¿ì¼öÇÑ ¼º´ÉÀ» º¸¿´´Ù. Á¦¾ÈÇÑ ¹æ¹ýÀº °¢ µµ¸ÞÀκ° ¼º´É ºñ±³¿¡¼ ºñ±³ ¸ðµ¨ ´ëºñ ÃÖ´ë 4.45 BLEU ¼º´É Çâ»óÀ» ¾ò¾ú´Ù. »Ó¸¸ ¾Æ´Ï¶ó Á¤¼ºÀû Æò°¡¸¦ ÅëÇØ Á¦¾ÈÇÑ ¹æ¹ýÀÌ ºñ±³ ¸ðµ¨¿¡ ºñÇØ µµ¸ÞÀο¡ Æ¯ÈµÈ ÀÀ´äÀ» »ý¼ºÇÏ´Â °ÍÀ» È®ÀÎÇÒ ¼ö ÀÖ¾ú´Ù. |
¿µ¹®³»¿ë (English Abstract) |
Multi-domain response generation is the task of generating responses to utterances that cover more than one domain. As the number of domains increases, multi-domain response generation must be able to respond appropriately to newly added domains as well as previously learned domains. However, previous studies based on fine-tuning with a new domain dataset have a problem that involves an inadequate response generation to pre-learned domains or a generation of domainindependent general responses. To solve this problem, the proposed model adopts a memory-replay continual learning to evolve a single-domain response generator to a multi-domain response generator. The model first learns a specific domain and retains some training instances in this domain for the next step. Then, the model is trained again with training instances for a new domain as well as the remaining instances. Since the remaining instances should represent the previous domains, the proposed method selects representative instances using clustering specialized to response generation. Through intensive experiments, the proposed method outperforms the baselines. |
Å°¿öµå(Keyword) |
ÀÀ´ä »ý¼º
´ÙÁß µµ¸ÞÀÎ ÀÀ´ä »ý¼º
±â¾ïÀçÇö Áö¼ÓÇнÀ
response generation
multi-domain response generation
memory-replay continual learning
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|