Language is primarily a tool for communication rather than thought

2024 第16回最先端NLP勉強会

Transcript

1. Evelina Fedorenko, Steven T. Piantadosi & Edward A. F. Gibson (Nature) ࠷ઌ୺ NLP ษڧձ 2024 ಡΉਓɿཥ ྇פʢSB Intuitionsʣ Language is primarily a tool for communication rather than thought
2. ͲΜͳ࿦จʁ 2 👈 "Perspective" ͱ͸…* - Review ΑΓڱ͍ൣғΛѻ͏ - ҙݟ΍ߟ࡯ʹภ͍ͬͯͯ΋ྑ͍ *https://www.nature.com/nprot/content-types NLP Ͱ͍͑͹ɺACL ͷ Theme Paper Έ͍ͨͳ΋ͷ͔
3. ஶऀΒͷόοΫάϥ΢ϯυ 3 Google Scholar هࡌͷΩʔϫʔυΛྻڍ Evelina Fedorenko Cognitive Neuroscience / Computational Neuroscience Cognitive Science / Neurolinguistics Steven T. Piantadosi Cognitive Science / Language Acquisition Psycholinguistics / Machine Learning Edward A. F. Gibson Psycholinguistics / Language Processing Sentence Processing / Syntax / Semantics
4. ࿦จͷओு͸ʁ 4 ʮݴޠ͸ओʹίϛϡχέʔγϣϯͷͨΊʹ͋Δʯ 🐵 ٞ࿦ʹ͓͚Δࢹ఺ - ݱ୅ͷਓؒʹ͓͚Δݴޠػೳ - ݴޠͷܗ੒ʹ࡞༻͢Δྗ 🙈 ѻΘͳ͍ࢹ఺ - ਓؒͷਐԽ/ੜଘʹ͓͚Δݴޠͷ໾ׂ - ݴޠ͸Խੴʹ࢒Βͳ͍͔Β೉͗͢͠Δ
5. ओͳ࿦఺ 5 ݴޠ͸ࢥߟͷཁͰ͸ͳ͍ ࠜڌ: ݴޠ͸ࢥߟͷඞཁ৚݅Ͱ΋े෼৚݅Ͱ͸ͳ͍ ݴޠ͸ίϛϡχέʔγϣϯͷͨΊʹ͋Δ ࠜڌ: ݴޠ͸৘ใ఻ୡʹ࠷దԽ͞ΕͨΑ͏ͳಛ௃Λ࣋ͭ
6. ओͳ࿦఺ 6 ݴޠ͸ࢥߟͷཁͰ͸ͳ͍ ࠜڌ: ݴޠ͸ࢥߟͷඞཁ৚݅Ͱ΋े෼৚݅Ͱ͸ͳ͍ ݴޠ͸ίϛϡχέʔγϣϯͷͨΊʹ͋Δ ࠜڌ: ݴޠ͸৘ใ఻ୡʹ࠷దԽ͞ΕͨΑ͏ͳಛ௃Λ࣋ͭ
7. ೴׆ಈʹ΋ݴޠͱࢥߟͷ෼཭͕ݟͯऔΕΔ 7 - ଟ͘ͷछͷࢥߟͷࡍʹɺݴޠॲཧʹؔ࿈͢Δ෦Ґ͸׆ಈ͠ͳ͍ Tips: ϒϩʔΧ໺ɾ΢Σϧχοέ໺͸ݴޠॲཧͱ͍͏ΑΓɺൃ࿩ॲཧʹؔΘΔ෦Ґ
8. ݴޠ͸ࢥߟͷඞཁ৚݅Ͱ͸ͳ͍ 8 ਓؒ͸ɺݴޠػೳΛࣦͬͨ৔߹Ͱ΋ࢥߟ͕Ͱ͖Δ ࣦޠ঱Ͱ΋֤छͷࢥߟ͕Ͱ͖Δ͜ͱ͕ࣔ͞Ε͍ͯΔ - ࢉ਺ - ϓϥϯχϯά - ࿦ཧతਪ࿦ - ৺ͷཧ࿦
9. ࿦ཧతਪ࿦ͷྫ 9 *Fluid intelligence is supported by the multiple-demand system not the language system (Woolgar, 2018) [Ҿ༻͞Ε͍ͯΔୈҰஶऀͷ࿦จ*ͷ಺༰] ࢥߟςετͷείΞͱಛఆ෦Ґͷଛইׂ߹ͷ૬ؔ Λଌఆ - ݴޠத਻ͱͷਖ਼ͷ૬ؔ͸ͳ͍ - ͦͷଞͷ෦Ґͱ͸૬͕ؔݟΒΕͨ ➡︎ ࢥߟͱݴޠ͕ಠཱͰ͋Δ ࢥߟςετͷΠϝʔδਤ Cattell Culture Fair Intelligence Test
10. ݴޠ͸ࢥߟͷे෼৚݅Ͱ͸ͳ͍ 10 ਓؒ͸ɺݴޠػೳ͕े෼อͨΕͨ··Ͱ΋ɺࢥߟೳྗΛࣦ͏৔߹͕͋Δ - ౷߹ࣦௐ঱ - ೴ͷଛইΛෛͬͨਓʑͷ֤छέʔε
11. ݴޠ͸ࢥߟͷͨΊʹ͋ΔͷͰ͸ͳ͍ - ͦͷଞͷαϙʔτ - 11 - ΄ͱΜͲݴޠʹ৮Εͣ੒௕ͨ͠Ζ͏ࣇͰ͋ͬͯ΋ɺࢉ਺ͳͲͷ࿦ཧతࢥߟ͕ Ͱ͖Δ - ݴޠ֫ಘҎલͷࢠͲ΋΍ಈ෺͸ݴ༿Λ࣋ͨͳ͍͕ɺߴ౓ͳࢥߟ͕Ͱ͖Δ © davidevison, stock.adobe.com © The New York Times
12. ओͳ࿦఺ 12 ݴޠ͸ࢥߟͷཁͰ͸ͳ͍ ࠜڌ: ݴޠ͸ࢥߟͷඞཁ৚݅Ͱ΋े෼৚݅Ͱ͸ͳ͍ ݴޠ͸ίϛϡχέʔγϣϯͷͨΊʹ͋Δ ࠜڌ: ݴޠ͸৘ใ఻ୡʹ࠷దԽ͞ΕͨΑ͏ͳಛ௃Λ࣋ͭ
13. ݴޠ͸৘ใ఻ୡʹ࠷దԽ͞Ε͍ͯΔ 13 ݴޠͷ͞·͟·ͳಛ௃Λڍ͛ɺ৘ใ఻ୡͷͨΊʹ࠷దԽ͞Ε͍ͯΔͱ͍ͯ͠Δɻ - Ի੠ͷߏ଄ - ୯ޠͷ௕͞ - ҙຯγεςϜ - จ๏ - ᐆດੑ ඞͣ͠΋ʮࢥߟͷͨΊͷݴޠʯΛ൓ূ͢Δ΋ͷͰ͸ͳ͍ɻ
14. Ի੠ͷߏ଄ 14 ֤छݴޠʹଘࡏ͢Δ฼Ի͸ɺ۠ผ͠΍͍͢Α͏ʹɺՄೳͳ฼ԻͷൣғΛे෼ʹ ޿͘औΔΑ͏ʹଘࡏ͍ͯ͠Δ* ➡︎ ฉ͖औΓ΍͢͞ʹ࠷దԽ͞Ε͍ͯΔ * The Dispersion-Focalization Theory of vowel systems (Schwartz, 1997) ͦΕʹੜཧతʹൃԻ͠΍͍͢ ➡︎ ࿩͠΍͢͞ʹ࠷దԽ͞Ε͍ͯΔ
15. ୯ޠͷ௕͞ 15 Zipf ୹ॖʹΑΕ͹ɺස౓ͷߴ͍୯ޠ΄Ͳɺ୹͍܏޲ʹ͋Δɻ ୈࡾஶऀͷ࿦จ*ʹΑΕ͹ɺ৘ใྔ͕௿͍୯ޠ΄Ͳɺ୹͍܏޲ʹ͋Δɻ * Word lengths are optimized for efficient communication (Piantadosi, 2011)
16. ҙຯγεςϜ 16 ͋Δछͷޠኮ͸ɺcomplexity ͱ informativeness ͷؒʹτϨʔυΦϑ͕͋Δɻ ৭࠼ޠኮͷྫ શͯͷΧϥʔίʔυຖʹҟͳΔ໊લΛ͚ͭΕ͹ informative ͕ͩɺcomplexɻ ͔ͱ͍ͬͯɺޠኮΛݮΒ͗͢͠Δͱɺnot complex Ͱ΋ not informativeɻ * Efficient compression in color naming and its evolution (Zaslavsky et al., 2018) ࣮ࡏͷݴޠͷ৭࠼ޠኮ͸ɺཧ࿦తʹ࠷దͳγεςϜ ʹ͍ۙ΋ͷ͹͔Γ*
17. จ๏ 17 - จ๏ͱ͸ complexity Λ཈͑ͭͭɺinformativeness ΛߴΊΔඞཁੑͷதͰੜ·Ε ͨ΋ͷͰ͸ͳ͍͔ʁ - ௕͗͢Δ܎Γड͚͸࢖ΘΕͳ͍ɺͱ͍͏ͷ΋ཧղͷ͠΍͢͞ʹ࠷దԽ͞Ε͍ͯΔͱ ͍͑Δɻ - SOV ͕ੈքͷݴޠͷதͰ࠷΋Α͘ݟΒΕΔޠॱʢ47%ʣɻ͜Ε͸ਓ͕ؒδΣε νϟʔͷͱ͖ʹ࠾༻͕ͪ͠ͳޠॱ*Ͱ͋Γɺೝ஌తʹ΋ࣗવͳޠॱͳͷͰ͸ͳ͍͔ɻ * The natural order of events: How speakers of different languages represent events nonverbally (Goldin-Meadow et al., 2008) ͜ͷลΓ͸ίϛϡχέʔγϣϯ͚ͩͰͳ͘ʮݴޠ͸ࢥߟ͠΍͍͢Α͏ʹ࡞ΒΕ͍ͯΔʯ ͱ͍͏ओுͷαϙʔτʹ΋ͭͳ͕Δؾ͕͢Δɻ ࣮ࡍɺචऀΒ΋ʮจ๏͸͜ͷٞ࿦ʹ͓͍ͯ controversial ͩʯͱݴ͍ͬͯΔɻ
18. ᐆດੑ 18 Chomsky ͸ʮݴޠʹ͸ᐆດੑ͕͋ΔɻͦΜͳίϛϡχέʔγϣϯΛ๦͛Δಛ௃͕ ͋Δͷ͔ͩΒɺݴޠ͸ࢥߟͷͨΊͳͷͩʯͱ͍ͯͨ͠ɻ චऀΒʹݴΘͤΕ͹ɺʮࢥߟʹᐆດੑ͸ͳ͍ͷʹɺͳͥݴޠʹʁʯ ίϛϡχέʔγϣϯʹ͓͍ͯᐆດੑ͕خ͍͠ྫ - ฉ͖खʹจ຺͔Βิ׬ͤ͞Δ͜ͱͰɺ࿩͠खͷ࿑ྗ͕ݮΔ - ᐆດੑʢଟٛੑʣ͕ͳ͔ͬͨΒɺͱΜͰ΋ͳ͍ྔͷޠኮ΍จ๏͕ඞཁʹͳΔ͸ͣ
19. ࿦จͷ·ͱΊ 19 ݴޠ͸ࢥߟͷඞཁ৚݅Ͱ΋े෼৚݅Ͱ͸ͳ͍ - ೴෦Ґຖͷ׆ಈΛΈΔͱݴޠͱࢥߟ͸෼཭͞Ε͍ͯΔ - ݴޠ͕ͳͯ͘΋ࢥߟ͕Ͱ͖Δ / ݴޠ͕͋ͬͯ΋ࢥߟ͕Ͱ͖ͳ͍ ݴޠ͸৘ใ఻ୡʹ࠷దԽ͞ΕͨΑ͏ͳಛ௃Λ࣋ͭ - Ի੠ͷߏ଄ / ୯ޠͷ௕͞ / ҙຯγεςϜ / จ๏ / ᐆດੑ ʮݴޠ͸ओʹίϛϡχέʔγϣϯͷͨΊʹ͋Δʯ
20. ײ૝ 20 🙌 ͍ΖΜͳτϐοΫ͕ࢄΓ͹ΊΒΕ͓ͯΓɺָ͍͠࿦จ - Box Ͱఏࣔ͞Ε͍ͯΔิ଍৘ใ΋๛෋ - ຊจதͰͷ৮ΕΒΕํ͸͋ͬ͞Γ͍ͯ͠Δͱ͜Ζ΋͋ΔͷͰɺҾ༻͞Ε͍ͯ ΔݪయΛ͋ͨΔͱ໘ന͍ 🗣 ݴޠΛܗ࡞ΔओཁҼʹɺίϛϡχέʔγϣϯ͕͋Δͱ͍͏఺͸ಉҙ - ͨͩࢥߟͱίϛϡχέʔγϣϯ͸දཪҰମͳͱ͜Ζ΋ʁ - ࢥߟ͸ࣗ෼͔Βࣗ෼΁ͷίϛϡχέʔγϣϯʁ - Ͱ΋΍ͬͺΓୈҰ໨త͸ଞऀͱͷίϛϡχέʔγϣϯͳؾ͕͢Δ
21. ײ૝ 21 💭 ࢥߟʢthoughtʣͷఆٛʹ͍ͭͯ΋ͬͱٞ࿦͕ಡΈ͍ͨͱࢥͬͨ - ຊ࿦จͰ঺հ͞Ε͍ͯΔࣄྫ͸ʮࢥߟͷ࠷খجຊ୯Ґ͕ݴޠʹґଘ͍ͯ͠ͳ ͍ʯ͜ͱΛ͍ࣔͯ͠ΔΑ͏ͳҹ৅ - ݴޠͷิॿ͕ͳ͍ͱߦ͑ͳ͍Α͏ͳࢥߟͷྫ΋ݟͯΈ͍ͨ
22. ײ૝ 22 🔬 ओޠͷͰ͔͍࿩͸ɺ౴͑Λग़͢͜ͱʹͩ͜ΘͬͨΓɺ೿ൊ૪͍Λ͢Δͷ͸ ҙຯ͕ͳ͍͕ɺϦαʔνΫΤενϣϯͷੜ੒૷ஔͱͯ͠໾ʹཱͭɻ - ;Θͬͱͨ͠ʮࢥߟʯ΍ʮίϛϡχέʔγϣϯʯΛఆٛͯ͠ɺݕূՄೳͳ ϦαʔνΫΤενϣϯΛߟ͑Α͏ʂ - ʮࢥߟʯͷϝλϑΝʔͱͯ͠ͷ൚༻ੑΛར༻ͯ͠ɺ͍ΖΜͳέʔε ʢਓؒҎ֎ɺLLM ͱ͔ʣʹ͍ͭͯɺࠓճͷٞ࿦Λద༻ͯ͠ΈΑ͏ʂ
23. ײ૝ 23 🤖 ຊ࿦จΛண૝ʹ LLM ʹࢥ͍Λ஘ͤͯ΋໘ന͍͔΋ ʮݴޠ͸ࢥߟͷे෼৚݅Ͱ͸ͳ͍ʯ → ݴޠΛഔհͱ͠ͳ͍ࢥߟͷଘࡏ͸ɺݴޠΛੜ੒͢Δ͚ͩͷ LLM ͷݶքΛࣔ ͔͢ʁ → ൱ɺݴޠ͸ࢥߟͷΤϛϡϨʔλͱͯ͠ৼΔ෣͑ΔͷͰ͸ͳ͍͔ɻਓؒ͸ ݴޠΛհ͞ͳ͍Ͱߦ͏ࢥߟʹؔͯ͠΋ɺLLM ͸ think step by step తʹɺ LLM ͳΓͷํ๏ͰΤϛϡϨʔτͯ͠ߦ͏͜ͱ͕Ͱ͖Δ͔΋ɻ ݴޠ͸৘ใ࿦తͳҙຯͰͷίϛϡχέʔγϣϯʹ࠷దԽ͞Ε͍ͯΔ → LLM ͷΑ͏ͳ౷ܭతϞσϧʹͱͬͯ΋ֶश͠΍ͦ͢͏ɻLLM ੒ޭͷҰҼʁ