ウェブ・ソーシャルメディア論文読み会 第36回: The Stepwise Deception: Simulating the Evolution from True News to Fake News with LLM Agents (EMNLP, 2025)

Transcript

1. The Stepwise Deception: Simulating the Evolution from True News to Fake News with LLM Agents (EMNLP 2025)  ଜࢁଠҰ ԣ඿ࠃཱେֶ 第36回ウェブ・ソーシャルメディア論⽂読み会
2. 2 ֓ཁ The Stepwise Deception: Simulating the Evolution from True News to Fake News with LLM Agents, (EMNLP, 2025) Yuhan Liu, Zirui Song, Juntian Zhang, Xiaoqing Zhang, Xiuying Chen, Rui Yan l Empirical Methods in Natural Language Processing (EMNLP 2025) l ⾃然⾔語処理分野のトップ会議の1つ l ݴޠཧղɾੜ੒͔ΒࣾձݴޠֶతԠ༻ͳͲͷɺ/-1෼໺ͷτϐοΫ͕த৺ l ஶऀ
   Yuhan Liu l .#;6"*
   1PTUEPD
   3FONJO
   6OJWFSTJUZ
   PG
   $IJOB
   1I% l ޡ৘ใʴLLMγϛϡϨʔγϣϯͷݚڀΛத৺ l From Skepticism to Acceptance: Simulating the Attitude Dynamics Toward Fake News, IJCAI, 2024. l The Truth Becomes Clearer Through Debate! Multi-Agent Systems with Large Language Models Unmask Fake News, SIGIR, 2025. l ͲΜͳ࿦จʁ ਅ࣮ͷχϡʔε͕ͲͷΑ͏ʹِ৘ใ΁ͱม࣭͍͔ͯ͘͠Λ࠶ݱ͢ΔͨΊͷγϛϡϨʔγϣϯϑϨʔϜϫʔΫz'64&z
   ΛఏҊɻ৘ใͷมԽΛ௥੻ɾ࠶ݱ͢Δ͚ͩͰͳ͘ɺِ৘ใରࡦʹؔ͢ΔհೖޮՌ΋ݕূՄೳͳ؀ڥΛߏஙɻ χϡʔ ε಺༰ͷײ৘ɾ֬৴౓ɾ৽ن৘ใɾελΠϧͳͲ࣍ݩͷมԽΛఆྔతʹޮՌΛଌఆ͢Δํ๏ΛఏҊɻ
3. 3 l ޡ৘ใ໰୊ͷਂࠁԽ l ଟ͘ͷطଘݚڀͷண؟఺ɿʮطʹଘࡏ͢Δʯޡ৘ใΛݕग़ɾ֦ࢄͷ࢓૊ΈΛ ཧղ͢Δ͜ͱ 㱺 ޡ৘ใͷଟ͘͸ɺਅ࣮ͷχϡʔε͕࣌ؒͷܦաͱͱ΋ʹ࿪ۂ͞ΕͨΓɺޡͬ ͯղऍ͞ΕͨΓͯ͠ɺঃʑʹਐԽ͍ͯ͘͠ϓϩηεΛܦͯੜ·ΕΔՄೳੑʁ എܠ
4. 4 എܠ How does Truth Evolve into Fake News? An Empirical Study of Fake News Evolution (WebConf, 2021) Mingfei Guo, Xiuying Chen, Juntao Li, Dongyan Zhao, Rui Yan l χϡʔε͸֦ࢄͷաఔͰվม͞ΕͨΓɺ࿪ΊΒΕͨΓ͠ɺʮਅ࣮
   ˠޡ৘ใ
   ˠ
   ͞Βʹվม͞Εͨޡ৘ใʯͱ͍ͬͨมԽ͢Δɻͦͷաఔͷ݅ͷ'BLF
   /FXT
   &WPMVUJPO
   '/&
   EBUBTFUΛެ։ l จ๏తʹ͸େ͖ͳมԽ͸ଘࡏ͠ͳ͍ l վม͞Εͨޡ৘ใ͸ɺ1PTJUJWF͔ͭ٬؍ తʹهड़͞ΕΔ܏޲
5. 5 എܠ Lost in translation: using global fact-checks to measure multilingual misinformation prevalence, spread, and evolution (EPJ DataScience, 2025) Dorian Quelle, Calvin Yixiang Cheng, Alexandre Bovet & Scott A. Hale l ݴޠɺ ݅ͷ
   GBDUDIFDLهࣄΛ ༻͍ͯɺݴޠΛ௒͑ͨޡ৘ใͷ֦ࢄɾ຋༁ ͷύλʔϯൃݟ l શମͷ໿ʢ໿ 
   ݅ʣͷओு ͸ෳ਺ճϑΝΫτνΣοΫ l ࣌ؒܦաͱͱ΋ʹओுͷมԽ͕ൃੜ
6. 6 2VFTUJPO ਅ࣮ͷχϡʔε͕ɺιʔγϟϧωοτϫʔΫ಺Ͱͷਓʑ ͷهԱɾ໾ׂɾ૬ޓ࡞༻Λ௨ͯ͡ͲͷΑ͏ʹِ৘ใ΁ͱ ਐԽͯ͠͠·͏ͷ͔ͱ͍͏ϓϩηεΛ࠶ݱͰ͖ͳ͍͔ʁ  ͳͥɺ৘ใ͕ม༰͍ͯ͘͠ͷ͔ʁ  Ͳͷཁૉ͕ɺ৘ใͷม༰ΛਐΊΔͷ͔ʁ  ͲͷΑ͏ʹհೖ͢Ε͹ɺ৘ใͷม༰Λ๷͛Δͷ͔ʁʢհೖޮՌʣ 㱺 --.ΤʔδΣϯτͰ৘ใΛ֦ࢄ͢ΔϢʔβΛ໛฿͢Δ؀ڥΛߏங
7. 7 '64&ʢ'BLF
   OFXT
   FWPMVUJPO
   4JNVMBUJPO
   GSBN&XPSLʣ ఏҊϑϨʔϜϫʔΫ l χϡʔε͕5SVF
   /FXT㱺 1BSUJBMMZ
   FWPMWFE
   GBLF
   OFXT㱺&WPMWFE
   GBLF
   OFXT ΁ͱมԽ͢ΔྲྀΕΛ࠶ݱ͢ΔΤʔδΣϯτϕʔεγϛϡϨʔγϣϯۭؒΛߏங
8. 8 '64&ͷྲྀΕ ఏҊϑϨʔϜϫʔΫ  χϡʔε͕ೖྗ  ωοτϫʔΫ͔ΒΤʔδΣϯτ͕ϥϯμϜʹબ୒  ϥϯμϜʹબ୒͞ΕͨΤʔδΣϯτ͕χϡʔεΛ ಡΈɺ.FNPSZΛߋ৽  ΤʔδΣϯτ͕3PMFʹԠͯ͡ɺχϡʔεΛ࠶ൃ৴ 㱺 ͜ΕΛ܁Γฦͯ͠ɺχϡʔε͕ਐԽ
9. 9 '64&ʢ'BLF
   OFXT
   FWPMVUJPO
   4JNVMBUJPO
   GSBN&XPSLʣ ఏҊϑϨʔϜϫʔΫ アラスカ州で 巨⼤なネコが発⾒！ ニュース記事 メモリー ・⼯場から危険な薬品が流出 ・アラスカ州は⼤きい アラスカ州で、 ⼯場の薬品の影響で ネコが巨⼤化！
10. 10 1SPQBHBUJPO
    3PMF"XBSF
    BHFOUT
    13" ఏҊϑϨʔϜϫʔΫ l 3PMF
    *EFOUJUZ l 4QSFBEFS
    ৘ใΛ୯७ʹ఻ൖ l $PNNFOUBUPS
    ҙݟ΍ղऍΛఏڙɻ l 7FSJGJFS
    ࣄ࣮֬ೝΛߦ͏໾ׂ l #ZTUBOEFS
    ؍࡯͠ɺੵۃతʹؔ༩͠ͳ͍ɻ l 1FSTPOBM *OGPSNBUJPO BOE
    5SBJUT l ໊લɺ೥ྸɺੑผɺֶྺͳͲ l #JH
    'JWFಛੑ l )JFSBSDIJDBM .FNPSZ l 4IPSU
    .FNPSZ
    ͦͷ೔ͷ૬ޓ࡞༻ͷهԱ l -POH .FNPSZ
    ͜Ε·ͰͷཤྺͷهԱ
11. 11 13"ͷϓϩϯϓτྫ ఏҊϑϨʔϜϫʔΫ l 4IPSU
    .FNPSZ΁ͷอଘ ”Summarize the opinions you have heard in a few sentences, including their own perspective on the news.” l χϡʔεൃ৴ “As a [role], you combine your [previous personal opinion] with the new information stored in your [long memory]. You process this information in the following manner: [role behavior], and then reintroduce the [news].”
12. 12 '64&&7"- χϡʔεͷมԽͷධՁ l χϡʔεͷมԽΛଌఆ͢ΔͨΊͷఆྔతͳࢦඪ l (15PNJOJͰධՁʢਓؒͱͷ1FSBTPO૬ؔ ʣ l 
    r 
    ͷ཭ࢄείΞ l $POUFOU
    %FWJBUJPO
    .FUSJDTͱ4UBUJTUJDBM
    %FWJBUJPO .FUSJDTͷछྨ l $POUFOU
    %FWJBUJPO
    .FUSJDT l 4FOUJNFOU
    4IJGU
    ײ৘ͷมԽ l /FX
    *OGPSNBUJPO
    *OUSPEVDFE ݩͷχϡʔεʹଘࡏ͠ͳ͍௥Ճ৘ใͷఔ౓ l $FSUBJOUZ
    4IJGU
    ֬৴౓ɾஅఆͷϨϕϧͷมԽ l 4UZMJTUJD
    4IJGU
    จମɺτʔϯɺݴޠతಛ௃ͷมԽ l 5FNQPSBM 4IJGU
    ࣌ؒతදݱͷมԽ l 1FSTQFDUJWF
    %FWJBUJPO
    ओ؍తͳҙݟɾࢹ఺ͷಋೖ 㱺 5PUBM
    %FWJBUJPOͱͯ͠ɺͭͷείΞͷฏۉΛܭࢉ
13. 13 '64&&7"- γϛϡϨʔγϣϯ݁Ռ l χϡʔεͷมԽΛଌఆ͢ΔͨΊͷఆྔతͳࢦඪ l (15PNJOJͰධՁʢਓؒͱͷ1FSBTPO૬ؔ ʣ l 
    r 
    ͷ཭ࢄείΞ l $POUFOU
    %FWJBUJPO
    .FUSJDTͱ4UBUJTUJDBM
    %FWJBUJPO .FUSJDTͷछྨ l 4UBUJTUJDBM
    %FWJBUJPO
    .FUSJDTɿ5PUBM
    %FWJBUJPOείΞ͔Β೿ੜͨ͠ࢦඪ l %FWJBUJPO ࠷ऴ೔ͱॳ೔ͷฏۉဃ཭ͷࠩ l "WFSBHF
    %FWJBUJPO
    ֤࣌ؒεςοϓʹ͓͚ΔશΤʔδΣϯτͷ5%ͷฏۉ l 'JOBM
    %FWJBUJPO ࠷ऴ࣌ؒʹ͓͚Δ5% l 1FBL %FWJBUJPO
    5JNF ࠷େ5%ʹୡ͢Δ·Ͱͷ࣌ؒ l )BMG %FWJBUJPO
    5JNF
    5%͕࠷େͷ൒෼ʹୡ͢Δ࣌ؒ
14. 14 τϐοΫͷൺֱ χϡʔεͷมԽͷධՁ l ੓࣏తͳχϡʔε͕࠷΋ဃ཭౓ɺဃ཭ εϐʔυͱ΋ʹߴ͍ l Պֶͱۚ༥ͷχϡʔεͷဃ཭εϐʔυ ͸Ώͬ͘ΓͰɺͦ͜·Ͱဃ཭͕ߴ͘ͳ Βͳ͍ l -B[FS
    %BWJE
    .+
    FU
    BM
    5IF
    TDJFODF
    PG
    GBLF
    OFXT 4DJFODF 
     
    
    ͱҰக͢Δ݁Ռ
15. 15 #JH
    'JWFͷൺֱ χϡʔεͷมԽͷධՁ l ײԽ͞Ε΍͍͢
    *NQSFTTJPOBCMF
    ΤʔδΣϯτͷάϧʔϓ͸ɺ༻৺ਂ͍
    7JHJMBOU
    ΤʔδΣϯτͷάϧʔϓͱൺֱͯ͠ɺဃ཭౓͕ߴ͍ l 7JHJMBOUάϧʔϓ͸ɺ҆ఆͨ͠৴೦Λ࣋ͭ܏޲ l .JS[BCFJHJ
    .BIEJFI
    .BITB
    5PSBCJ
    BOE
    5BIFSFI
    +PXLBS
    5IF
    SPMF
    PG
    QFSTPOBMJUZ
    USBJUT
    BOE
    UIF
    BCJMJUZ
    UP
    EFUFDU
    GBLF
    OFXT
    JO
    QSFEJDUJOH
    JOGPSNBUJPO
    BWPJEBODF
    EVSJOH
    UIF
    $07*%
    QBOEFNJD -JCSBSZ
    )J
    5FDI  
    ͱҰக͢Δ݁Ռ
16. 16 ࣮ࡍͷχϡʔεͱಉ༷ͷ౤ߘΛ࠶ݱ χϡʔεͷมԽͷධՁ l τϥϯϓऻܸࣄ݅ͷχϡʔεͰ͸ɺz
    5SVNQ
    XBT
    OPU
    BUUBDLFE
    JU`T
    B
    ESBNBUJD
    FGGFDU zͳͲ͕ੜ੒ l Ξϧθϯνϯ੓෎ͱ*.'ͷ৽ͨͳ࠴຿࠶ฤڠఆͷχϡʔεͰ͸z"SHFOUJOBT
    
    *.'
    EFBM
    JT
    KVTU
    BOPUIFS
    DPSQPSBUF
    TDIFNF
    JO
    EJTHVJTFz l ʮͷϑΣΠΫχϡʔεΛ࠶ݱͰ͖ͨʯͱओுʢͲ͏΍ͬͯʁʣ
17. 17 Ͳͷཁૉ͕ဃ཭ʹد༩͍ͯ͠Δͷ͔ʁ χϡʔεͷมԽͷධՁ l 1FSTQFDUJWF
    %FWJBUJPOʢ1%ʣ͕࠷΋େ͖͘د༩ɺ 5FNQPSBM
    4IJGUʢ54ʣ͕খ͍͞د༩ l 4DJFODFΧςΰϦʔͷχϡʔε͸ɺ 1FSTQFDUJWF
    %FWJBUJPOʢ1%ʣ
    ͱ/FX
    *OGPSNBUJPO
    *OUSPEVDUJPO ʢ/**ʣͰେ͖͘มԽ l 6SCBO -FHFOE
    'JOBODF͸ɺ /FX
    *OGPSNBUJPO
    *OUSPEVDUJPOʢ/**ʣʹґଘ
18. 18 ൓࣮Ծ૝తγφϦΦͷݕূɿհೖ χϡʔεͷมԽͷධՁ l հೖͷτϦΨʔɿဃ཭౓͕ҰఆҎ্ൃੜ ͨ͠৔߹ʹհೖ l հೖઓུɿ৴པͰ͖Δ৘ใݯʹج͍ͮͨ ੠໌Λൃߦ͠ɺ֦ࢄʹର߅ l հೖͳ͠ͱൺֱͯ͠ɺ'JOBM
    %FWJBUJPO͕ ݮগ l ૣظհೖͷॏཁੑΛࣔࠦ
19. 19 ·ͱΊ ·ͱΊ l ଞʹ΋ɺωοτϫʔΫઃఆʹΑΔൺֱɺ"CMBUJPO 4UVEZͳͲͷ࣮ݧ΋͋Γ l جຊతͳ--.ΤʔδΣϯτγϛϡϨʔγϣϯͷઃఆͱಉ͡Ͱ͋Γɺγϛϡ Ϩʔγϣϯ؀ڥͷ৽نੑͱ͍͏ΑΓ΋ɺϑΣΠΫχϡʔεͷมԽͱ͍͏఺ʹ ஫໨ͨ͠఺͕ධՁ͞Εͨ఺ʁ l ΤʔδΣϯτͷ໾ׂ͕ݻఆͰ͋Δ఺΍ɺ৘ใͷ֦ࢄΛ΄ͱΜͲߦ͏఺ͳͲɺ վળͰ͖Δ఺͸·ͩ͋Γͦ͏ l ϑΣΠΫχϡʔεʹ͔͔ΘΒͣɺʮ৘ใ$POUFOUͷਐԽʯͱ͍͏ҙຯͰҰൠ ੑΛ΋ͨΕͤͦ͏ͳςʔϚ