Upgrade to Pro — share decks privately, control downloads, hide ads and more …

論文紹介 Improving Medical Reasoning through Retrie...

Takashi Nishibayashi
February 19, 2024
760

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

医療×LLM勉強会の発表資料です (公開版)

Takashi Nishibayashi

February 19, 2024
Tweet

More Decks by Takashi Nishibayashi

Transcript

  1. ࿦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).
  2. ֓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౴ ベ ϯνϚʔΫ デ ʔληοτͰධՁ w #ҎԼͷύϥϝʔλαΠζͰΦʔϓϯͳج൫ϞσϧΑΓ΋ฏۉϙΠϯ τ"DDVSBDZ͕૿Ճ
  3. 4FMG3"( ઌߦݚڀ Asai, Akari, et al. "Self-rag: Learning to retrieve,

    generate, and critique through self-reflection." arXiv preprint arXiv:2310.11511 (2023).
  4. 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 ͯ͠͠·͏
  5. 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
  6. 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༗༻͞
  7. 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
  8. .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.
  9. 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։
  10. (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&
  11. औಘͨ͠FWJEFODFͷબ୒ w ൷ධείΞ4ʹج͍ͮͯऔಘͨ͠UPQ,FWJEFODFͷத͔ΒબͿ͜ͱͰ 
 ҎԼͷ༷ʹҩࢣͷ਍அΞϓϩʔνʹ͍ۙग़ྗ͕ՄೳʹͳΔ fi H  w ଟ೯๔ੑཛ૥঱ީ܈

    1$04 ͷ਍அྫ w ױऀ͸1$04ͷయܕతͳ঱ঢ় で ͋ΔχΩ ビ ͱංຬ が ͋Δ w ױऀ͸͠ ば ͠ ば 1$04ͱؔ࿈͢Δܕ౶೘පͷՈ଒ྺ が ͋Δ w ױऀ͸1$04ͷಛ௃ で ͋Δ ブド ΢౶ෛՙࢼݧΛड͚ͨ ϋΠύʔύϥϝʔλw(Ͱਪ࿦࣌ ͷ;Δ·͍Λม͑Δ
  12. 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