
çæAIãçšããã·ã¹ãããã£ãã¯ã¬ãã¥ãŒ
ð æç®æ å ± ãš æé²åèš³
ã·ã¹ãããã£ãã¯ã¬ãã¥ãŒã«ãããå€§èŠæš¡èšèªã¢ãã«é§ååã¹ã¯ãªãŒãã³ã°ã®ããã®ããã³ãããã³ãã¬ãŒãã®éçº
ðCao, Christian, et al. "Development of prompt templates for large language modelâdriven screening in systematic reviews." Annals of Internal Medicine 178.3 (2025): 389-401. https://doi.org/10.7326/ANNALS-24-02189
ð DOI, PubMed, Google Scholar ð²MORE‎ >>> Connected Papers
â» Connected Papersãšã¯ïŒ >>> note.
[èæ¯ã»ç®ç] ã·ã¹ãããã£ãã¯ã¬ãã¥ãŒïŒSRïŒã¯ãæåã®å³å¯ãªè«æã¹ã¯ãªãŒãã³ã°ãé害ãšãªããä¿¡é Œã§ããæ å ±çµ±åãžã®ã¢ã¯ã»ã¹ãé ããŠãããç®çå€§èŠæš¡èšèªã¢ãã«ïŒLLMïŒäž»å°ã®æé²ã»ãã«ããã¹ãã¹ã¯ãªãŒãã³ã°ã®ããã®æ±çšããã³ãããã³ãã¬ãŒããéçºããæ§ã ãªã¬ãã¥ãŒã«é©å¿ã§ããããã«ããããšã
[æ¹æ³] ãã¶ã€ã³ïŒèšºæãã¹ãã®ç²ŸåºŠãèšå®ïŒ10ä»¶ã®SRã«ãããŠã48 425ä»¶ã®åŒçšæç®ãæé²ã¹ã¯ãªãŒãã³ã°ã®å¯Ÿè±¡ãšããŠãã¹ãããããã«ããã¹ãã¹ã¯ãªãŒãã³ã°ã§ã¯ããªãªãžãã«ã®æ€çŽ¢ããèªç±ã«å ¥æå¯èœãª12 690ã®è«æãã¹ãŠãè©äŸ¡ãããããã³ããã®éçºã«ã¯GPT4-0125-previewã¢ãã«ïŒOpenAIïŒã䜿çšãããåå è ãªããæž¬å®ïŒå€§èŠæš¡ãªèšèªã¢ãã«ãçšããŠãSRã®é©æ Œæ§åºæºã«åºã¥ããŠè«æãå å«ãŸãã¯é€å€ããããä¿ãããã¢ãã«åºåã¯ãããã©ãŒãã³ã¹ïŒç²ŸåºŠãæåºŠãç¹ç°åºŠïŒãè©äŸ¡ããããã«ããã«ããã¹ãã¹ã¯ãªãŒãã³ã°åŸã®ãªãªãžãã«ã®SRèè æ±ºå®ãšæ¯èŒãããã
â æåºŠãšç¹ç°åºŠã®å®çŸ©
ã»LLM ã®å€å®çµæããå ã®SR èè ãæçµçã«åã蟌ã¿ïŒé€å€ããå€å®ãšç §åããæåºŠïŒæŸãæŒãããªãåïŒãšç¹ç°åºŠïŒäžèŠè«æã匟ãåïŒã§æ§èœãæ¯èŒã
[çµæ] GPT4-0125-previewãçšããæé©åããã³ããã¯ãæé²ã¹ã¯ãªãŒãã³ã°ã«ãããŠéã¿ä»ãæåºŠ97.7%ïŒç¯å²86.7%ïœ100%ïŒãç¹ç°åºŠ85.2%ïŒç¯å²68.3%ïœ95.9%ïŒãå šæã¹ã¯ãªãŒãã³ã°ã«ãããŠéã¿ä»ãæåºŠ96.5%ïŒç¯å²89.7%ïœ100.0%ïŒãç¹ç°åºŠ91.2%ïŒç¯å²80.7%ïœ100%ïŒã10ä»¶ã®SRã§éæãããäžæ¹ããŒãã·ã§ããããã³ããã®æåºŠã¯äœãã£ãïŒæé²49.0ïŒ ãå šæ49.1ïŒ ïŒãLLMå šäœã§ã¯ãClaude-3.5ïŒAnthropicïŒãšGPT4ã®äºçš®ã¯åçšåºŠã®æ§èœã§ãã£ãããGemini ProïŒGoogleïŒãšGPT3.5ïŒOpenAIïŒã®ã¢ãã«ã¯æ§èœäžè¶³ã§ãã£ãã10,000 ä»¶ã®è¢«åŒçšã«å¯ŸããçŽæ¥çãªã¹ã¯ãªãŒãã³ã°ã³ã¹ãã¯å€§ããç°ãªãïŒ ããã®æé²ã®ã¹ã¯ãªãŒãã³ã°ã«ã¯83æé以äžã1666.67ç±³ãã«ãå¿ èŠã§ãããšæšå®ãããã®ã«å¯Ÿããæã ã®LLMããŒã¹ã®ã¢ãããŒãã¯1æ¥æªæºã157.02ç±³ãã«ã§ã¹ã¯ãªãŒãã³ã°ãå®äºããã

- æé©åããããã³ããã䜿ãã ãã§ããŒãã·ã§ããã®å以äžã®æåºŠãéæã
- 人æã«æ¯ã¹ãæé㯠1/3 以äžãè²»çšã¯çŽ 1/10 ã«ã
- Claude-3.5 ãš GPT-4 ç³»ã¯åçã®é«æ§èœãGemini Pro ãš GPT-3.5 ã¯ããå£åŸã
[çµè«] ä»ã®SRãLLMã«ãé©çšå¯èœãªãé«ãæåºŠãšç¹ç°åºŠãéæããæé²ããã³å šæã¹ã¯ãªãŒãã³ã°ã®æ±çšããã³ãããéçºããããæã ã®ããã³ããã®å·¥å€«ã¯ãSRã®ç ç©¶è ãå»åŠå šè¬ã«ããã£ãŠåæ§ã®åºæºããŒã¹ã®ã¿ã¹ã¯ã宿œããç ç©¶è ã«ãšã£ãŠäŸ¡å€ããããããããªããéçãšããŠãããªãè¿ éãªæé©åãååšããå¯èœæ§ããããåŸæ¹èŠçç ç©¶ãSRã®ã³ã³ãããšã³ã¹ãµã³ãã«ãã¹ã¯ãªãŒãã³ã°ã®å šæè©äŸ¡ã¯ãPubMed Centralã®ç¡æå šæèšäºã«éå®ããã
ð± So What?ïŒäœãé¢çœããšæãããïŒ
åã¯ãèªåã§ãã·ã¹ãããã£ãã¯ã¬ãã¥ãŒç ç©¶ã«åå ããããšãè€æ°åãã£ãŠããã®å€§å€ãã身ããã£ãŠç¥ã£ãŠããã
äžå£ã«ã·ã¹ãããã£ãã¯ã¬ãã¥ãŒãšãã£ãŠããããã¯è€æ°ã®ããã»ã¹ã«ãã£ãŠæ§æãããŠããã
ãã®äžã§ãæãäœæ¥çã§ãèšãæãããšãèŸããäœæ¥ã
ãããã¹ã¯ãªãŒãã³ã°ãã§ããã
ãã®ã¹ã¯ãªãŒãã³ã°ã¯æé²ã¹ã¯ãªãŒãã³ã°ãå
šæã¹ã¯ãªãŒãã³ã°ã®å€§ããïŒã€ã«åãããã
å人çã«ãç¹ã«å€§å€ãšæããã®ã¯åŸè
ã®å
šæã¹ã¯ãªãŒãã³ã°ã§ãåã蟌ã¿åºæºãé€å€åºæºã«ç
§ããããã詳现ãªãã§ãã¯ãå¿
èŠã«ãªãã
ä»åã®æèªç ç©¶ã§ã¯ãããã³ãããæé©åããŠããã®ã¹ã¯ãªãŒãã³ã°äœæ¥ãããããŠã¿ãã
ãã®çµæãšããŠãé«ãæåºŠãç¹ç°åºŠãèªããå®çšã«èããå¯èœæ§ã瀺ããã
ãã®ã³ã¹ãã¯ã人æã«æ¯ã¹ãŠãæé㯠1/80 以äžãè²»çšã¯çŽ 1/10 ã«ãªããšããé©ç°çãªãã®ã ã£ãã
çæAIããã©ã®ãããªé åã«äœ¿ããã䜿ãã¹ããããšããåæªã®è°è«ã¯ãããšæãã
å人çã«ã¯ãæé²ã»å
šæã¹ã¯ãªãŒãã³ã°ã®ãããªãæèãåµé æ§ã®å
¥ã蟌ãäœå°ã®å°ãªããäœæ¥çïŒå¹æåšçïŒãªäœæ¥é åã«ã€ããŠã¯ãã©ãã©ãå§ããŠããã¹ãã ãšèããŠããã
ãã®éšåã«ã€ããŠãçæAIã䜿ãåŽã®äººéã¯ãããããèãããå¿
èŠãããã
â¬ïž é¢é£ note & ðã§ã®æçš¿âš
ðçæAIãçšããã·ã¹ãããã£ãã¯ã¬ãã¥ãŒ
â çåŠçæ³å£«_海接éœäž Ph.D. (@copellist) May 21, 2025
ã»æé², å šæã¹ã¯ãªãŒãã³ã°ã«çæAIã䜿çš
ã»æåºŠâæŸãæŒãããªãå, ç¹ç°åºŠâäžèŠè«æã匟ãå
ð¹æé©åããã³ããã¯ãŒãã·ã§ããã®å以äžã®æåºŠ
ð¹äººæã«æ¯ã¹, æé㯠1/3 以äž, è²»çšã¯çŽ 1/10ã«ïŒ
ãã®äœæ¥ãçæAIã§èªååãããã®ã¯é©åœçã§ããð² pic.twitter.com/S7FnKXDysl
âââââââââââââã»ã»ã»â¥ ⥠⥠â¥
è¯è³ªãªãªãå»åŠé¢é£ã»è±è«ææèªãã¢ãªïŒARIã
ãã¡ãããâª
âââ

⥠⥠⥠â¥ã»ã»ã»ââââââââââââ
#ïžâ£ #çåŠçæ³ #èšåºç ç©¶ #ç ç©¶ #ãªãããªããŒã·ã§ã³ #è±è«æ #æç®æèª #è±ææèª #ãšããã³ã¹ #ãµã€ãšã³ã¹ #æ¯æ¥æŽæ° #æè¿ã®åŠã³