³×À̹ö(´ëÇ¥ ÇѼº¼÷)°¡ ÀÚ¿¬¾îó¸® ºÐ¾ß ÇÐȸ ‘EMNLP(Empirical Methods in Natural Language Processing) 2021’¿¡¼ ÃÊ´ë±Ô¸ð AI ¿¬±¸ ¼º°ú¸¦ °øÀ¯ÇÑ´Ù.
Ŭ·Î¹Ù ¹× AI·¦ ¿¬±¸ÁøÀº ³×À̹öÀÇ ÃÊ´ë±Ô¸ð AI ‘ÇÏÀÌÆÛŬ·Î¹Ù’ÀÇ ÇÙ½É ¿¬±¸ ³í¹®À» ºñ·ÔÇØ, ÃÑ 7°³ÀÇ ³í¹®À» ¹ßÇ¥ÇÒ ¿¹Á¤ÀÌ´Ù.
25³â ÀüÅëÀ» °¡Áø EMNLP´Â ACL(Association for Computational Linguistics)°ú ´õºÒ¾î ÀÚ¿¬¾îó¸®(NLP) ºÐ¾ßÀÇ ÃÖ°í AI ÇÐȸ·Î ²ÅÈù´Ù.
¿ÃÇØ´Â µµ¹Ì´ÏÄ«°øȱ¹¿¡¼ 11¿ù 7~11ÀϱîÁö ¿¸®¸ç, ¿Â¶óÀÎÀ¸·Îµµ µ¿½Ã ÁøÇàµÈ´Ù.
À̹ø¿¡´Â ³×À̹öÀÇ ÃÊ´ë±Ô¸ð AI ÇÏÀÌÆÛŬ·Î¹ÙÀÇ ÇÙ½É ¿¬±¸ ³í¹®ÀÌ ¸ÞÀÎ ÄÁÆÛ·±½º ¹ßÇ¥¿¡ äÅõǴ ¼º°ú¸¦ °Åµ×´Ù.
ÇÏÀÌÆÛŬ·Î¹Ù´Â ³×À̹ö°¡ Áö³ 5¿ù ±¹°ø°³ÇÑ ÃÊ´ë±Ô¸ð AIÀÌÀÚ, GPT-3º¸´Ù Çѱ¹¾î µ¥ÀÌÅ͸¦ 6500¹è ÀÌ»ó ÇнÀÇÑ, ÇöÀç °ø°³µÈ ´ÜÀÏ ¸ðµ¨ Áß Àü¼¼°è¿¡¼ °¡Àå Å« Çѱ¹¾î ÃÊ´ë±Ô¸ð ¾ð¾î¸ðµ¨À̱⵵ ÇÏ´Ù.
ÀÌ ¿¬±¸´Â Çѱ¹¾î ÃÊ´ë±Ô¸ð ¾ð¾î¸ðµ¨ ‘ÇÏÀÌÆÛŬ·Î¹Ù’¿Í ±× ÇнÀ¿¡ »ç¿ëµÈ µ¥ÀÌÅ͸¦ ¼Ò°³ÇÏ°í, ´Ù¾çÇÑ Å©±âÀÇ ¸ðµ¨µéÀÌ °®´Â ¼º´ÉÀ» °ËÁõÇÏ´Â ³»¿ëÀÌ´Ù.
³í¹®¿¡¼´Â ÇÏÀÌÆÛŬ·Î¹Ù°¡ ´Ù¾çÇÑ Çѱ¹¾î °úÁ¦(task)¿¡ ´ëÇØ Á¦ÇÑµÈ ¿¹Á¦¸¸À¸·Îµµ ¶Ù¾î³ ÇнÀ(in-context learning) ¼º´ÉÀ» º¸Àδٴ °ÍÀ» Áõ¸íÇß´Ù.
ÀÌ¿Í ´õºÒ¾î, ÃÊ°Å´ë ¾ð¾î¸ðµ¨¿¡¼ÀÇ ÇÁ·ÒÇÁÆ® ÃÖÀûÈ(prompt optimization), µ¥ÀÌÅÍ ÅäÅ«È(tokenization) µî¿¡ ´ëÇÑ ³íÀǸ¦ ¹ßÀü½ÃÅ°°í, ‘ÇÏÀÌÆÛŬ·Î¹Ù ½ºÆ©µð¿À’¸¦ ÅëÇØ ‘³ë ÄÚµå AI’(No Code AI) µî ÃÊ´ë±Ô¸ð AI°¡ °¡Á®¿Ã AI ¼ºñ½º °³¹ß ¹æ¹ý·ÐÀÇ Çõ½ÅÀûÀÎ º¯È¿¡ ´ëÇؼµµ ¼³¸íÇÑ´Ù.
¿¬±¸¿¡´Â ³×À̹öÀÇ Å¬·Î¹Ù ¹× AI·¦ÀÇ ¿¬±¸Áø »Ó¸¸ ¾Æ´Ï¶ó, ¼ÒÇÁÆ®¿þ¾î Ç÷§Æû ¿£Áö´Ï¾î, °Ë»ö ¿£Áö´Ï¾î µî ´Ù¾çÇÑ ÆÀ¿¡ °ÉÃÄ ÃÑ 37¸íÀÌ ÀúÀÚ·Î Âü¿©Çß´Ù.
- ³×À̹ö EMNLP 2021 äÅà ³í¹® ¸®½ºÆ® - |
1. What Changes Can Large-scale Language Models Bring? Intensive Study on Billions-scale Korean Generative Pretrained Transformers
±èº¸¼·*, ±èÇü¼®*, ÀÌ»ó¿ì* ¿Ü 34¸í
2. Cost-effective End-to-end Information Extraction for Semi-structured Document Images
Ȳ¿ø¼®, ÀÌÇöÁö, ±èÁø¿µ, ±è±â¿í, ¼¹ÎÁØ(KAIST)
3. Can Language Models be Biomedical Knowledge Bases?
¼º¹«Áø (°í·Á´ë), ÀÌÁøÇõ(Princeton Univ.), À̼®¿ø (°í·Á´ë), Àü¹ÎÁö(Icahn School of Medicine at Mount Sinai), ±è¼ºµ¿, °Àç¿ì(°í·Á´ë)
https://arxiv.org/abs/2104.08041
4. GPT3Mix: Leveraging Large-scale Language Models for Text Augmentation.
À¯°¹Î, ¹Úµ¿ÁÖ, °Àç¿í, ÀÌ»ó¿ì, ¹Ú¿ì¸í
https://arxiv.org/abs/2104.08826
5. Reasoning Visual Dialog with Sparse Graph Learning and Knowledge Transfer.
°±âõ(¼¿ï´ë), ¹ÚÁؼ®(¼¿ï´ë), ÀÌȶõ, À庴Ź†(¼¿ï´ë), ±èÁøȆ
6. Devil’s Advocate: Novel Boosting Ensemble Method for Text Classification
Á¶ÈÖ¿, ÀÓÀç¼(¼¿ï´ë), À庴Ź(¼¿ï´ë)
7. Understanding Mathematical Notation Semantics in Scientific Papers.
Á¶ÈÖ¿, °µ¿¿±(Univ. of Minnesota), Andrew Head (Univ. of Pennsylvania), Marti A. Hearst (Univ. of California, Berkeley) |
ÇÏÁ¤¿ì ³×À̹ö AI·¦ ¼ÒÀåÀº “ÀÛ³â ÇϹݱâºÎÅÍ ÃÊ´ë±Ô¸ð AI ±â¼ú¿¡ ÁýÁßÀûÀ¸·Î ÅõÀÚÇÏ°í ¿¬±¸ ¿ª·®À» °áÁý½ÃŲ °á°ú, ‘ÇÏÀÌÆÛŬ·Î¹Ù’ ±â¼úÀ» °ø°³ÇÏ°í ¼º°øÀûÀ¸·Î »ó¿ëÈÇÑ µ¥ À̾î ÇÐȸ¿¡¼ ±× ±â¼ú·ÂÀ» ÀÎÁ¤¹Þ´Â Äè°Å¸¦ °Åµ×´Ù”¸é¼ “À̹ø ³í¹®Àº ¿µ¾î Áß½ÉÀ̾ú´ø ±âÁ¸ÀÇ ¾ð¾î¸ðµ¨ ¿¬±¸¸¦ ³Ñ¾î, Çѱ¹¾î AIÀÇ °¡Ä¡¸¦ ±Û·Î¹ú ÀÚ¿¬¾îó¸® Çа谡 ÀÎÁ¤Çß´Ù´Â Á¡¿¡¼µµ ¸Å¿ì ÀÇ¹Ì ÀÖ´Â ¼º°ú”¶ó°í °Á¶Çß´Ù.
ÀÌ ¿Ü¿¡µµ ³×À̹ö´Â ÇÏÀÌÆÛŬ·Î¹Ù¿Í °ü·Ã, ÃÊ°Å´ë ¾ð¾î¸ðµ¨À» È°¿ëÇØ µ¥ÀÌÅ͸¦ ÀÚµ¿ »ý¼º ¹× ¶óº§¸µ ÇØ µ¥ÀÌÅ͸¦ È¿À²ÀûÀ¸·Î Áõ° ¹× Áõ·ù(distillation)½ÃÅ°´Â ±â¹ý¿¡ ´ëÇÑ ¿¬±¸µµ ¼Ò°³ÇÒ ¿¹Á¤ÀÌ´Ù.
µ¡ºÙ¿© OCR °úÁ¦¿¡¼ ¹®¼ÀÇ Á¤º¸¸¦ ´õ¿í È¿À²ÀûÀ¸·Î ÃßÃâÇÒ ¼ö ÀÖ´Â ¹æ½ÄÀ» Á¦¾ÈÇÏ´Â ³í¹®, AI ¾ð¾î¸ðµ¨ÀÌ ¹ÙÀÌ¿À¸ÞµðÄà ºÐ¾ßÀÇ Áö½Ä º£À̽º(knowledge base)·Î¼ È°¿ëµÉ °¡´É¼ºÀ» Ž±¸ÇÏ´Â ³í¹® µî ´Ù¾çÇÑ ÁÖÁ¦¸¦ ¸Á¶óÇÏ´Â ¼±Ç࿬±¸ °á°ú¸¦ ¹ßÇ¥ÇÒ ¿¹Á¤ÀÌ´Ù.
ÀÌ´Â Ä«À̽ºÆ®, °í·Á´ë, ¼¿ï´ë µî ´Ù¾çÇÑ ±¹³»¿Ü ´ëÇеé°ú Àû±ØÀûÀ¸·Î Çù·ÂÇÑ °á°ú¶ó°í ³×À̹ö ÃøÀº µ¡ºÙ¿´´Ù.
ÇÑÆí, ³×À̹ö´Â À̹ø EMNLP 2021¿¡ ³×À̹ö·¦½ºÀ¯·´°ú ÇÔ²² ½Ç¹ö(Silver) µî±Þ ½ºÆù¼·Îµµ Âü¿©ÇÏ¸ç ±Û·Î¹ú IT ±â¾÷ ¹× Çаè¿Í Àû±Ø ±³·ùÇÏ°í, AI ±â¼ú ¹ßÀü¿¡µµ ±â¿©ÇÑ´Ù´Â °èȹÀÌ´Ù.
<±èµ¿±â ±âÀÚ>kdk@bikorea.net < ÀúÀÛ±ÇÀÚ © BI KOREA ¹«´ÜÀüÀç ¹× Àç¹èÆ÷±ÝÁö > |