論文紹介 Improving Medical Reasoning through Retrieval and Self-Reflection with Retrieval-Augmented Large Language Models

࿦จ঺հ Improving Medical Reasoning through Retrieval and Self-Re fl ection
with Retrieval- Augmented Large Language Models ҩྍº--.ษڧձ! ੢ྛ޹ 5BLBTIJ/JTIJCBZBTIJIBHJOP 1 Jeong, Minbyul, et al. "Improving Medical Reasoning through Retrieval and Self-Reflection with Retrieval-Augmented Large Language Models." arXiv preprint arXiv:2401.15269 (2024).

֓ཁ w όΠΦϝσΟΧϧྖҬͷ࣭໰Ԡ౴ʹಛԽͨ͠3"(ͷϑϨʔϜϫʔΫͰ͋Δ 4FMG#JP3"(ΛఏҊ w ࣭໰ʹճ౴͢Δͷʹ֎෦஌͕ࣝඞཁ͔Ͳ͏͔ͷ൑அ͓ΑͼɺSFUSJFWBMNPEFM ͷݕࡧ݁Ռ͕࣭໰ʹؔ࿈͍ͯ͠Δ͔ɺݕࡧ݁ՌΛར༻ͨ͠ग़ྗ͕࣭໰ͷճ౴ ͱͯ͠༗༻͔Ͳ͏͔Λਪ࿦࣌ʹ൑அͭͭ͠ճ౴Λੜ੒ 4FMG3"(
w ੜ੒Ϟσϧͷ܇࿅ʹόΠΦϝσΟΧϧಛԽJOTUSVDUJPOTFUΛར༻ w ֎෦஌ࣝͱͯ͠1VC.FEͳͲͷจॻ͔Β࡞੒ͨ͠ίʔύεΛ࢖༻ w ݁Ռͭͷҩྍ࣭໰Ԡ౴ ベ ϯνϚʔΫ デ ʔληοτͰධՁ w #ҎԼͷύϥϝʔλαΠζͰΦʔϓϯͳج൫ϞσϧΑΓ΋ฏۉϙΠϯ τ"DDVSBDZ͕૿Ճ

*OUSPEVDUJPO

όΠΦϝσΟΧϧྖҬʹ͓͚Δ--.ͷద༻ʹ͍ͭͯ w ࣭໰Ԡ౴΍ςΩετੜ੒ʹ͓͍ͯɺױऀͷ৘ใͳͲ͸--.ࣗ਎ͷ஌ ࣝͰΧόʔ͢Δࣄ͸Ͱ͖ͳ͍ w ϋϧγωʔγϣϯΛ༠ൃ w ͜ͷͨΊճ౴ͷཪ෇͚ࠜڌͱͳΔ৘ใΛఏڙ͢Δ3"( SFUSJFWBM BVHNFOUFEHFOFSBUJPO
͕༻͍ΒΕΔ w ͔͠͠όΠΦϝσΟΧϧྖҬʹ͓͍ͯ͸൚༻తͳख๏Ͱ্ख͍͘ ͔ͳ͍

ఏҊख๏ͷ֓ཁ

4FMG3"( ઌߦݚڀ Asai, Akari, et al. "Self-rag: Learning to retrieve,
generate, and critique through self-reflection." arXiv preprint arXiv:2310.11511 (2023).

୯७ͳ3"(ͷ໰୊఺ Asai, Akari, et al. "Self-rag: Learning to retrieve, generate,
and critique through self-reflection." arXiv preprint arXiv:2310.11511 (2023). ֎෦஌͕ࣝෆཁͳͱ͖ʹ ΋༩͑ͯ͠·͍֎෦஌ࣝ ͕ϊΠζͱͳΔ ࣭໰ͷճ౴ͱͯ͠Ұ؏ੑ ͷແ͍ճ౴Λͯ͠͠·͏ ࣭໰ʹؔ࿈͠ͳ͍಺༰΋ ༩͑ͯ͠·͏͜ͱͰճ౴ ʹͳ͍ͬͯͳ͍จষΛฦ ͯ͠͠·͏

4FMG3"(ͷखଓ͖ Asai, Akari, et al. "Self-rag: Learning to retrieve, generate,
and critique through self-reflection." arXiv preprint arXiv:2310.11511 (2023). ֎෦஌͕ࣝෆཁͳͱ͖͸ ࢖Θͳ͍ ճ౴ʹࠜڌ͕ඞཁ͔Ͳ͏͔ ͷ൑ఆ݁ՌΛද͢ಛघτʔ ΫϯΛੜ੒͠ɺΦϯσϚϯ υͰSFUSJFWBMNPEFMΛݺͼ ग़͢ ݸʑͷݕࡧ݁Ռ͕࣭໰ʹ ؔ࿈͕͋Δ͔Ͳ͏͔ɺճ ౴͢ΔͨΊͷ৘ใΛؚΜ Ͱ͍Δ͔൱͔Λ൑ఆ   ݕࡧ݁ՌΛݩʹੜ੒ͨ͠ ग़ྗ͕࣭໰ͷճ౴ʹͳͬ ͍ͯΔ͔ࣗݾ൷ධ͢Δ

4FMG3"(ͷ*OTUSVDUJPO5VOJOH Asai, Akari, et al. "Self-rag: Learning to retrieve, generate,
and critique through self-reflection." arXiv preprint arXiv:2310.11511 (2023). 8BMLJOHEFBEͷ์ૹ։࢝೔Λ ஌Δʹ͸֎෦஌͕ࣝඞཁͳͷͰ <3FUSJFWF:FT> 1SFNJFSFEPO0DUPCFSͱ͋ ΔͷͰؔ࿈͋Γ <*T3&-3FMFWBOU> ग़ྗʮ0DUPCFS ʯ͸࣭ ໰ͷճ౴ʹͳ͍ͬͯΔͷͰ <*T461'VMMZ4VQQPSUFE> ճ౴ͷ༗༻͞

4FMG3"(ͷSF fl FDUJPOUPLFO Asai, Akari, et al. "Self-rag: Learning to
retrieve, generate, and critique through self-reflection." arXiv preprint arXiv:2310.11511 (2023). w SF fl FDUJPOτʔΫϯΛਪ࿦࣌ʹग़ྗࣗ͠ݾ੍ޚΛߦͳ͏ w ֎෦஌ࣝEΛऔಘ͢Δ͔Ͳ͏͔ɺग़ྗZΛଓ͚Δ͔ࢭΊΔ͔ʜ

.FE$15 ؔ࿈ݚڀ w όΠΦϝσΟΧϧྖҬಛԽ*OGPSNBUJPO3FUSJFWBMNPEFM w 1VC.FEͷݕࡧΫΤϦͱΫϦοΫϩάΛར༻ͯ͠܇࿅ Jin, Qiao, et al.
"MedCPT: Contrastive Pre-trained Transformers with large-scale PubMed search logs for zero-shot biomedical information retrieval." Bioinformatics 39.11 (2023): btad651.

ఏҊख๏4FMG#JP3"( w 4FMG3"(͔Βͷมߋ఺ w *OTUSVDUJPO5VOJOHͷͨΊͷ*OTUSVDUJPO4FUΛόΠΦϝσΟΧϧ ಛԽςΩετ͔Βੜ੒ w &WJEFODFऔಘઌͱͳΔίʔύεΛ1VC.FE΍1.$ͱ͍ͬͨόΠ ΦϝσΟΧϧσʔλʹ w
3FUSJFWBM.PEFMʹ.FE$15Λ࠾༻ w ౴͑ʹؔ͢Δઆ໌ΛఏڙͰ͖Δਪ࿦ SFBTPOJOH ೳྗΛ΋ͭ

ߩݙ w ੜ෺ҩֶ͓ΑͼྟচࢦࣔʹΑͬͯ܇࿅͞Εͨ4FMG#JP3"(ϑϨʔϜ ϫʔΫΛಋ͍ͨ w υϝΠϯݻ༗ͷίϯϙʔωϯτ SFUSJFWFS EPDVNFOUT JOTUSVDUJPOTFUT
͕ͦͷυϝΠϯͷࢦࣔʹରॲ͢ΔͨΊʹඞཁͰ͋ ΔࣄΛূ໌ͨ͠ w ͭͷΦʔϓϯυϝΠϯੜ෺ҩֶ࣭໰Ԡ౴ϕϯνϚʔΫσʔληο τͰͷ༗ޮੑΛ࣮ূ͠ฏۉͷ"DDVSBDZϙΠϯτͷվળΛୡ੒ w ࣮ݧʹར༻ͨ͠JOTUSVDUJPOTFU͓Αͼ܇࿅ίʔυɺXFJHIUΛެ։

.FUIPE

*OTUSVDUJPOTFU w υϝΠϯಛԽͷࢦࣔηοτΛར༻

%PDVNFOUDPSQVT w &WJEFODFऔಘઌͱͳΔίʔύεΛҩֶσʔλΛݩʹ੔උ w 1VC.FEͷ"CTUSBDU 1.$GVMMUFYU ਍ྍΨΠυϥΠϯ $1( ڭՊॻ

3F fl FDUJPO5PLFOTPG4FMG#JP3"( w 4FMG3"(ͱಉ͡

(FOFSBUPS.PEFMͱ$SJUJD.PEFM w ࠷ऴతʹ࢖͏ͷ͕(FOFSBUPS.PEFM w $SJUJD.PEFM w (FOFSBUPS.PEFMͷ*OTUSVDUJPO5VOJOH܇࿅σʔλΛΞϊςʔγϣϯ͢Δͷ ʹར༻ w ·ͣΞϊςʔγϣϯ༻ͷ$SJUJD.PEFM
-.$ Λ༻ҙ w $SJUJD.PEFMΛ*OTUSVDUJPO5VOJOH w $SJUJD.PEFM༻ͷ*OTUSVDUJPO4FU͸(15ͰΞϊςʔγϣϯ w (FOFSBUPS.PEFMͷ܇࿅ w $SJUJD.PEFMͰ܇࿅σʔλʹSF fl FDUJPOUPLFOΛ෇༩ w 3&5 3&- 461 64&

औಘͨ͠FWJEFODFͷબ୒ w ൷ධείΞ4ʹج͍ͮͯऔಘͨ͠UPQ,FWJEFODFͷத͔ΒબͿ͜ͱͰ   ҎԼͷ༷ʹҩࢣͷ਍அΞϓϩʔνʹ͍ۙग़ྗ͕ՄೳʹͳΔ fi H w ଟ೯๔ੑཛ૥঱ީ܈
1$04 ͷ਍அྫ w ױऀ͸1$04ͷయܕతͳ঱ঢ় で ͋ΔχΩ ビ ͱංຬ が ͋Δ w ױऀ͸͠ ば ͠ ば 1$04ͱؔ࿈͢Δܕ౶೘පͷՈ଒ྺ が ͋Δ w ױऀ͸1$04ͷಛ௃ で ͋Δ ブド ΢౶ෛՙࢼݧΛड͚ͨ ϋΠύʔύϥϝʔλw(Ͱਪ࿦࣌ ͷ;Δ·͍Λม͑Δ

࣮ݧ݁Ռͱߟ࡯

Կ͕ੑೳ޲্ʹد༩ͨ͠ͷ͔ όΠΦϝσΟΧϧಛԽ*OTUSVDUJPO4FUTͷಋೖ࣌ͷੑೳ޲্෯ ͕Ұ൪େ͖͍

ར༻ͨ͠FWJEFODFͷιʔεͷൺֱ FWJEFODFऔಘઌͷൺ཰Λطଘख๏ͱൺֱɻఏҊख๏͸਍ྍΨΠυϥ ΠϯͱڭՊॻͷׂ߹͕૿͍͑ͯΔ

֎෦஌ࣝར༻ʹΑΔੑೳ޲্ ֎෦஌ࣝͷར༻͸໌֬ʹੑೳʹد༩͍ͯ͠Δ

BEBQUJWFSFUSJFWBMͷޮՌ "<ৗʹ֎෦஌ࣝΛར༻>ɺ#<ৗʹར༻͠ͳ͍>ɺ$<ඞཁ͔൑அΛͯ͠ར ༻>ͷύλʔϯΛൺֱɻ#͸"ΑΓѱ͘ͳΔࣄ͋Δ͕$ΑΓྑ͍࣌΋͋ Γ݁Ռ͸ෆ҆ఆ

ײ૝ w ਖ਼ղ཰ׂ͕ఔ౓ͱ͍͏ࣄ͸3"(Ͱૠೖͨࠜ͠ڌ͕ਖ਼͘͠ͳ͍ w *3ͷਫ਼౓͕ؾʹͳΔ w ධՁσʔληοτ͕ଟࢶબ୒ܗࣜͳͷͰɺબ୒ࢶͦΕͧΕʹ͍ͭͯ FWJEFODFΛऔಘͨ͠Βਖ਼ղͰ͖ͯ͠·͏ͷͰ͸
w ࣮຿ͷྟচ਍அ͸ଟࢶબ୒Ͱ͸ͳ͍ͷͰɺ0" 0QFO2VFTUJPO Ͱͷੑ ೳ͕ؾʹͳΔͱಉ࣌ʹ0"ͰFWJEFODFΛҾ͖౰ͯΔͷ͸͔ͳΓ೉ͦ͠͏ w ਍ྍΨΠυϥΠϯɾڭՊॻͷ࢖༻ׂ߹͕૿͑ͯਖ਼ղ͍ͯ͠ΔͷͰ঱ྫใࠂ ʹଟ͍ϨΞͳ঱ྫΑΓ΋DPNNPOدΓͷDBTF͕ࢼݧʹଟ͍

ײ૝ w ίʔύεͷFNCFEEJOHޙͷαΠζͰݟΔͱ1VC.FEͷ΋ͳ͍࡭ ͷڭՊॻͷޮ཰͕ѹ౗త w ڭՊॻɾ਍ྍΨΠυϥΠϯͷॏཁੑ

論文紹介 Improving Medical Reasoning through Retrie...

論文紹介 Improving Medical Reasoning through Retrieval and Self-Reflection with Retrieval-Augmented Large Language Models

Takashi Nishibayashi

More Decks by Takashi Nishibayashi

Featured

Transcript

࿦จ঺հ Improving Medical Reasoning through Retrieval and Self-Re fl ection

*OUSPEVDUJPO

όΠΦϝσΟΧϧྖҬʹ͓͚Δ--.ͷద༻ʹ͍ͭͯ w ࣭໰Ԡ౴΍ςΩετੜ੒ʹ͓͍ͯɺױऀͷ৘ใͳͲ͸--.ࣗ਎ͷ஌ ࣝͰΧόʔ͢Δࣄ͸Ͱ͖ͳ͍ w ϋϧγωʔγϣϯΛ༠ൃ w ͜ͷͨΊճ౴ͷཪ෇͚ࠜڌͱͳΔ৘ใΛఏڙ͢Δ3"( SFUSJFWBM BVHNFOUFEHFOFSBUJPO

ఏҊख๏ͷ֓ཁ

4FMG3"( ઌߦݚڀ Asai, Akari, et al. "Self-rag: Learning to retrieve,

୯७ͳ3"(ͷ໰୊఺ Asai, Akari, et al. "Self-rag: Learning to retrieve, generate,

4FMG3"(ͷखଓ͖ Asai, Akari, et al. "Self-rag: Learning to retrieve, generate,

4FMG3"(ͷ*OTUSVDUJPO5VOJOH Asai, Akari, et al. "Self-rag: Learning to retrieve, generate,

4FMG3"(ͷSF fl FDUJPOUPLFO Asai, Akari, et al. "Self-rag: Learning to

.FE$15 ؔ࿈ݚڀ w όΠΦϝσΟΧϧྖҬಛԽ*OGPSNBUJPO3FUSJFWBMNPEFM w 1VC.FEͷݕࡧΫΤϦͱΫϦοΫϩάΛར༻ͯ͠܇࿅ Jin, Qiao, et al.

ఏҊख๏4FMG#JP3"( w 4FMG3"(͔Βͷมߋ఺ w OTUSVDUJPO5VOJOHͷͨΊͷOTUSVDUJPO4FUΛόΠΦϝσΟΧϧ ಛԽςΩετ͔Βੜ੒ w &WJEFODFऔಘઌͱͳΔίʔύεΛ1VC.FE΍1.$ͱ͍ͬͨόΠ ΦϝσΟΧϧσʔλʹ w

ߩݙ w ੜ෺ҩֶ͓ΑͼྟচࢦࣔʹΑͬͯ܇࿅͞Εͨ4FMG#JP3"(ϑϨʔϜ ϫʔΫΛಋ͍ͨ w υϝΠϯݻ༗ͷίϯϙʔωϯτ SFUSJFWFS EPDVNFOUT JOTUSVDUJPOTFUT