๐Ÿ“š ๋…ผ๋ฌธ

๐Ÿ“š ๋…ผ๋ฌธ

Beyond Candidates: Adaptive Dialogue Agent Utilizing Persona and Knowledge

Abstract dialogue agents์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ์ด์–ด์ง€๋Š” ๊ฐ€์šด๋ฐ, ์•ž์„  ์—ฐ๊ตฌ๋“ค์€ ํŽ˜๋ฅด์†Œ๋‚˜ (persona)์™€ ์ง€์‹ (knowledge)์„ ๋‚ด์žฅ๋œ ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ๊ฐ€์ ธ์™€ ๋‹ต๋ณ€ํ–ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค์ œ ์„ธ๊ณ„์—์„œ ์‚ฌ๋žŒ๋“ค์ด ๋Œ€ํ™”ํ•  ๋•Œ, ์‚ฌ๋žŒ๋“ค์€ ์ค€๋น„๋œ ํ›„๋ณด ๋ฌธ์žฅ๋“ค์„ ๊ฐ€์ง€๊ณ  ๋‹ต๋ณ€ํ•˜๊ธฐ๋ณด๋‹ค๋Š”, ๋Œ€ํ™”์— ๋งž๋Š” ์˜๋ฏธ์  concept์„ ๋งˆ์Œ์— ๊ฐ€์ง€๊ณ  ๋Œ€ํ™”ํ•œ๋‹ค. ์ด๋Ÿฐ ๋Œ€ํ™” ์–‘์‹์— ์ฐฉ์•ˆํ•˜์—ฌ, ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๋ฌธ์žฅ ํ›„๋ณด๋“ค์ด ์ฃผ์–ด์ง€์ง€ ์•Š์€ ์ƒํ™ฉ์—์„œ์˜ ์ ์‘์  ๋Œ€ํ™” ์‹œ์Šคํ…œ์„ ์ œ์•ˆํ•œ๋‹ค. ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•œ ๋ชจ๋ธ์€ ๋‹จํŽธ์ ์ธ ์ •๋ณด๋งŒ์„ ๊ฐ€์ง€๊ณ  ์ผ๊ด€์ ์ด๊ณ  ๊ด€๋ จ ์žˆ๋Š” persona ์„ค๋ช…์„ ์ƒ์„ฑํ•˜๋ฉฐ, ๋…ผ๋ฆฌ์ ์ธ ๋‹ต๋ณ€์„ ์œ„ํ•ด ๊ด€๋ จ๋œ ์ง€์‹์„ ํ™•์ธํ•œ๋‹ค. Introduction ์ผ๋ฐ˜์ ์ธ ๋Œ€ํ™”์—์„œ๋Š” ๋Œ€ํ™” ์ฃผ์ œ์™€ ๋Œ€ํ™”์ž๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋Œ€ํ™”์˜ ์˜๋ฏธ๋ก ์  ๊ฐœ๋…์„ ๋งˆ์Œ์— ํ’ˆ..

๐Ÿ“š ๋…ผ๋ฌธ

IDAS: Intent Discovery with Abstractive Summarization

์š”์ฆ˜ New Intent Discovery (NID) ํƒœ์Šคํฌ์— ๋น ์ ธ์žˆ๋‹ค. ์ง€๋‚œ๋ฒˆ์— MTP-CLNN ๋…ผ๋ฌธ์„ ์ฝ๊ณ ,  "ํ”„๋กฌํ”„ํŠธ๋กœ ๋ฐœํ™”์— ๋Œ€ํ•œ intent๋ฅผ ์ƒ์„ฑํ•˜๊ณ , ๊ทธ๊ฑธ ๊ธฐ๋ฐ˜์œผ๋กœ clustering ํ•˜๋ฉด unsupervised ๋ฐฉ์‹์— ์žˆ์–ด ์ข‹์€ ํ‰๊ฐ€๋ฅผ ๋ฐ›์„ ์ˆ˜ ์žˆ์„ ๊ฒƒ ๊ฐ™๋‹ค"๋ผ๋Š” ์ƒ๊ฐ์œผ๋กœ ๊ด€๋ จ ๋…ผ๋ฌธ๋“ค์„ ์ฐพ์•„๋ณด์•˜๋‹ค.๊ทธ์ค‘, NLP4CONVAI@ACL 2023์— ๊ฒŒ์žฌ๋œ ๋…ผ๋ฌธ๊ณผ ๋‚ด๊ฐ€ ์ƒ๊ฐํ•œ ์•„์ด๋””์–ด๊ฐ€ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์ด ์žˆ์–ด ์ฝ์–ด๋ณด๊ฒŒ ๋˜์—ˆ๋‹ค.์กฐ๊ธˆ๋งŒ ๋” ์ผ์ฐ ์ƒ๊ฐํ• ๊ฑธAbstractMethod์ถ”์ƒ์ ์ธ summary์— ๊ธฐ๋ฐ˜ํ•œ utternace๋“ค์„ clustering ํ•˜๋Š” ๊ฒƒ์ด ๊ธฐ์กด์˜ intent discovery ๋ฐฉ๋ฒ•๋“ค๋ณด๋‹ค ๋›ฐ์–ด๋‚  ์ˆ˜ ์žˆ์ŒIDAS: LLM์— prompting์„ ํ†ตํ•ด ๋ฐœํ™”์˜ label์„ ์ƒ์„ฑํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ I..

๐Ÿ“š ๋…ผ๋ฌธ

New Intent Discovery with Pre-training and Contrastive Learning

์—ฐ๊ตฌ ๋ถ„์•ผ๋ฅผ ์ •ํ•˜๋ ค๊ณ  ๋…ผ๋ฌธ์„ ๋ณด๊ณ  ์žˆ๋Š”๋ฐ, NID (New Intent Classification) ๋…ผ๋ฌธ๋“ค์„ ๊ณ„์† ์ฝ๊ฒŒ ๋œ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์€ 2022 ACL ํ•™ํšŒ์— ์ˆ˜๋ก๋œ ๋…ผ๋ฌธ์ด๋ฉฐ, ์ฃผ ์ €์ž๋Š” Yuwei Zhang ์ด๋‹ค.AbstractProblem๊ธฐ์กด์˜ ๋ฐฉ๋ฒ•๋“ค์€ ๋‹ค๋Ÿ‰์˜ labeled data์— ์˜์กดํ•˜๊ฑฐ๋‚˜ pseudo-labeling์„ ํ†ตํ•œ clustering ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋„ˆ๋ฌด label์— ์˜์กด์ ์ด๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” NID ๋ถ„์•ผ์— ์žˆ์–ด ๋‹ค์Œ ์งˆ๋ฌธ๋“ค์— ๋Œ€ํ•œ ๋‹ต์„ ์–ป๊ณ ์ž ํ–ˆ๋‹ค:์–ด๋–ป๊ฒŒ ์˜๋ฏธ์  ๋ฐœํ™” ํ‘œํ˜„์„ ํ•™์Šต์‹œํ‚ฌ ์ˆ˜ ์žˆ๋Š”์ง€๋ฐœํ™”๋“ค์„ ์–ด๋–ป๊ฒŒ ๋” ์ž˜ clustering ํ•  ์ง€MethodMulti-task pre-training(MTP) ์ „๋žต ์‚ฌ์šฉrepresentation learning์„ ์œ„ํ•ด ๋งŽ์€ ์–‘์˜ un..

๐Ÿ“š ๋…ผ๋ฌธ

Two Birds One Stone: Dynamic Ensemble for OOD Intent Classification

์ง€๋‚œ๋ฒˆ DeepAligned Clustering ๋…ผ๋ฌธ์— ์ด์–ด ์ด๋ฒˆ์—๋Š” OOD intent classification์— ๊ด€ํ•œ ๋…ผ๋ฌธ์„ ์ฝ์—ˆ๋‹ค.๋ณธ ๋…ผ๋ฌธ์€ 2023๋…„๋„ ACL ํ•™ํšŒ์— ์ˆ˜๋ก๋œ ๋…ผ๋ฌธ์ด๋ฉฐ, ์ง€๋‚œ๋ฒˆ ์ €์ž์™€ ๊ฐ™์€ Xipeng Qiu ๊ฐ€ ์ €์ž๋กœ ์ฐธ์—ฌํ–ˆ๋‹ค.AbstractTODS์—์„œ OOD intent classification์€ ์ •๋ง ํ™œ๋ฐœํ•˜๊ฒŒ ์—ฐ๊ตฌ๋˜๋Š” ์ฃผ์ œ์ด๋‹ค. ์ด ๋ถ„์•ผ์—์„œ๋Š” 2๊ฐ€์ง€๋ฅผ ์š”๊ตฌํ•˜๋Š”๋ฐ,๋ฐ”๋กœ ๋ชจ๋ธ์ด '๋ฌด์—‡์„ ์•„๋Š”๊ฐ€'์™€ '๋ชจ๋ธ์ด ๋ฌด์—‡์„ ์•Œ์ง€ ๋ชปํ•˜๋Š”๊ฐ€'์ด๋‹ค.๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” open-world scenario์—์„œ์˜ overthinking๊ณผ OOD intent classification ๋ถ„์•ผ์—์„œ ๊ทธ๊ฒƒ์˜ ์˜ํ–ฅ๋ ฅ์— ๋Œ€ํ•ด ํƒ๊ตฌํ•œ๋‹ค.๊ฒฐ๊ณผ์ ์œผ๋กœ, ์ด ๋ชจ๋ธ์€ ์ถ”๋ก  ๊ณผ์ •์—์„œ OOD classification์„ ์ผ์ฐ ๋งˆ๋ฌด..

์žฅ์˜์ค€
'๐Ÿ“š ๋…ผ๋ฌธ' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๊ธ€ ๋ชฉ๋ก