iOSエンジニアのためのNLP基礎

iOSDC2017での登壇資料です。

Transcript

1. iOSΤϯδχΞͷͨΊͷ NLPجૅ @akatsuki174
2. ࣗݾ঺հ • ਢ౻ຘ • iOSΤϯδχΞ • גࣜձࣾZaimॴଐ • iOSDC 2017ίΞελοϑ • Twitter, Qiita, SlideShare, Github, SpeakerDeck →akatsuki174
3. ࠓ೔ͷεϥΠυ ɾSpeakerdeckͷϦϯΫΛઌఔTwitterͰྲྀ͠·ͨ͠ ɾݟͮΒ͍ɺޙͰݟฦ͍ͨ͠౳͕͋Ε͹ͪ͜Βࢀর ɹΑΖ͓͘͠ئ͍͠·͢ ʢ࣌ؒͳ͍ͷͰεϥΠυαΫαΫਐΊ·͢ʣ ΞΧ΢ϯτ͸ @akatsuki174Ͱ͢
4. ࠓ೔ͷεϥΠυ ※iOS11ͷ࿩΋ग़͖ͯ·͕͢ɺެ։͞Ε͍ͯΔ৘ใΛ ݩʹεϥΠυΛ࡞͍ͬͯ·͢ɻ
5. ࢲͱࣗવݴޠॲཧ ɾେֶͷଔۀݚڀͰΏΔΏΔ1೥ؒʢॏཁޠ൑ఆʣ ɾ࠷ॳͷձࣾͰ2,3ϲ݄ʢτϐοΫ൑ఆʣ ɾ2ࣾ໨ͰΏΔΏΔยखؒ൒೥ʢݕࡧνϡʔχϯάʣ
6. ͜ͷηογϣϯͷ໨త ɾNLP API͸ͲͷΑ͏ͳ΋ͷͳͷ͔஌ͬͯ΋Β͏ ɾࣗવݴޠॲཧʹগ͠Ͱ΋ڵຯΛ࣋ͬͯ΋Β͏
7. ͜ͷηογϣϯͷେ࿮ NLP APIΛઆ໌ͭͭͦ͠ͷཪଆͰ࢖ΘΕ͍ͯΔ ࣗવݴޠॲཧͷٕज़ʹ͍ͭͯ঺հ͠·͢
8. ͦ΋ͦ΋ࣗવݴޠॲཧͬͯʁ
9. ࣗવݴޠͱ͸ ʮਓ͕͓ؒޓ͍ʹίϛϡχέʔγϣϯΛߦ͏ͨΊͷ ࣗવൃੜతͳݴޠʯ by Wikipedia 㱻ਓ޻ݴޠʢϓϩάϥϛϯάݴޠ౳ʣͳͲ https://ja.wikipedia.org/wiki/ࣗવݴޠ
10. ࣗવݴޠॲཧͱ͸ ʮਓ͕ؒ೔ৗతʹ࢖͍ͬͯΔࣗવݴޠΛίϯϐϡʔλʹ ॲཧͤ͞ΔҰ࿈ͷٕज़Ͱ͋Γɺਓ޻஌ೳͱݴޠֶͷҰ෼໺ʯ by Wikipedia https://ja.wikipedia.org/wiki/ࣗવݴޠॲཧ
11. ࣗવݴޠॲཧͷ׆༻ྫ ɾػց຋༁ ɾ͔ͳ׽ࣈม׵ ɾର࿩γεςϜʢνϟοτϘοτ౳ʣ ɾ৘ใநग़ɺ৘ใݕࡧ
12. ࣗવݴޠॲཧͷશମ૾ େن໛σʔλ࣌୅ʹٻΊΒΕΔࣗવݴޠॲཧ https://www.slideshare.net/pfi/ss-11474303
13. ࣗવݴޠॲཧͷશମ૾ େن໛σʔλ࣌୅ʹٻΊΒΕΔࣗવݴޠॲཧ https://www.slideshare.net/pfi/ss-11474303 ΈΜͳͷେ޷͖ͳ ܗଶૉղੳ͸ ୯ޠ෼ׂͷҰ෼໺
14. ࣗવݴޠॲཧͷશମ૾ େن໛σʔλ࣌୅ʹٻΊΒΕΔࣗવݴޠॲཧ https://www.slideshare.net/pfi/ss-11474303 NLP API͕ѻ͏ ൣғ͸͜͜Β΁Μ
15. WWDC2017Ͱͷ ηογϣϯ಺༰ʹ͍ͭͯ
16. [WWDC 2017] Natural Language Processing and your Apps ɾϑΝʔετύʔςΟΞϓϦͰͲͷΑ͏ͳ׆༻͕͞ΕΔ͔ ɾࣗવݴޠॲཧͷجૅతͳઆ໌ ɾNLP APIͱ͸ ɾNLP API׆༻ྫ ɾNLP APIΛ࢖͏ϝϦοτ ɾσόοά࣌ͷώϯτ
17. [WWDC 2017] Natural Language Processing and your Apps ɾϑΝʔετύʔςΟΞϓϦͰͲͷΑ͏ͳ׆༻͕͞ΕΔ͔ ɾࣗવݴޠॲཧͷجૅతͳઆ໌ ɾNLP APIͱ͸ ɾNLP API׆༻ྫ ɾNLP APIΛ࢖͏ϝϦοτ ɾσόοά࣌ͷώϯτ
18. Natural Language Processing and your Apps / https://developer.apple.com/videos/play/wwdc2017/208/ iMessageΛ։͘ ݻ༗໊ࢺ͕ม׵ީิ ʹग़ͯ͜ͳ͍
19. Natural Language Processing and your Apps / https://developer.apple.com/videos/play/wwdc2017/208/ χϡʔεΞϓϦΛ։͘ ͖ͬ͞ม׵͠Α͏ͱ ͨ͠ݻ༗໊ࢺؚ͕·Ε ͍ͯΔχϡʔεΛಡΉ
20. Natural Language Processing and your Apps / https://developer.apple.com/videos/play/wwdc2017/208/ iMessageʹ໭Δ ͖ͬ͞ग़ͯ͜ͳ͔ͬͨ ݻ༗໊ࢺ͕ม׵ ީิʹʂʂ
21. ͦͷଞηογϣϯ಺༰ʹؔ͢Δ ৄ͍͠৘ใ͸QiitaͰɻ
22. ͔͜͜Β͸NLP APIʹԊͬͯ ࣗવݴޠॲཧͦͷ΋ͷʹ͍ͭͯ ࿩͍ͨ͠ͱࢥ͍·͢ɻ
23. NLP APIͷॲཧͷྲྀΕ ݴ ޠ ൑ ఆ τ ʛ Ϋ ϯ Խ ඼ ࢺ λ ά ෇ ͚ Ϩ ϯ Ϛ Խ ݻ ༗ ද ݱ ந ग़
24. NLP APIͷॲཧͷྲྀΕ ݴ ޠ ൑ ఆ τ ʛ Ϋ ϯ Խ ඼ ࢺ λ ά ෇ ͚ Ϩ ϯ Ϛ Խ ݻ ༗ ද ݱ ந ग़ ೔ຊޠඇରԠ
25. NLP APIͷॲཧͷྲྀΕ ݴ ޠ ൑ ఆ τ ʛ Ϋ ϯ Խ ඼ ࢺ λ ά ෇ ͚ Ϩ ϯ Ϛ Խ ݻ ༗ ද ݱ ந ग़
26. ݴޠ൑ఆ ɾԿΛ͢Δʹͯ͠΋ݴޠ͕Θ͔Βͳ͚Ε͹Կ΋Ͱ͖ͳ͍ ɾͦͷͨΊʢݴޠͷ൑ఆ͕ඞཁͳ৔߹͸ʣҰ൪࠷ॳʹ ɹߦΘΕΔλεΫ ɾจࣈN-gramΛ׆༻ͯ͠൑ఆΛߦ͏͜ͱ͕ଟ͍༷ࢠ
27. N-gramʹΑΔݴޠ൑ఆ ɾݴޠ൑ఆͩͬͯͦΜͳ؆୯͡Όͳ͍ ɾɾಉ͡ޠ೿ɺޠ܈ͷݴޠ͸Α͘ࣅ͍ͯΔ ɾͦ͜ͰN-gramͷొ৔ ɾNจࣈ(or N୯ޠ)͝ͱʹ۠੾Γɺͦͷմ͕ͲͷݴޠͰΑ͘ ɹݟΒΕΔ͔Λ൑ఆʢ㲈෼ྨʣ
28. จࣈN-gramʹΑΔݴޠ൑ఆͷྫ ɾN = 1 ɹcafé ʢϑϥϯεޠʣ ɹˠ ӳޠʹ͸ʮéʯͱ͍͏දه͸ͳ͍͔Β… etc ɾN = 2 ɹthe peopleʢӳޠʣɹ ɹˠʮthʯͱ͍͏ฒͼ͸ӳޠʹଟ͍͕υΠπޠʹ͸গͳ͘… etc
29. iOSʹ͓͚Δݴޠ൑ఆ iOS11Ͱ৽ొ৔ ݴޠίʔυͷจࣈྻͰग़ྗ͞ΕΔ
30. ɾʮHelloʯͱ͍͏จࣈΛݴޠ൑ఆ͢Δͱ”it”͕ฦͬͯ͘Δ ɾɾ͔͠͠”en”ͱ͍͏Մೳੑ΋͋Γ͏Δ ↓ ༧Ίݴޠ͕Θ͔͍ͬͯΔͳΒͪ͜Β͔ΒݴޠΛࢦఆͯ͠ ͋͛ͨํ͕ྑ͍ ؾΛ͚ͭΔ΂͖͜ͱ
31. NLP APIͷॲཧͷྲྀΕ ݴ ޠ ൑ ఆ τ ʛ Ϋ ϯ Խ ඼ ࢺ λ ά ෇ ͚ Ϩ ϯ Ϛ Խ ݻ ༗ ද ݱ ந ग़
32. ʮτʔΫϯʯʮτʔΫϯԽʯ ɾτʔΫϯʢ㲈ࣈ۟ʣ ɹςΩετதͷ୯ޠͱ۟ಡ఺ͷ͜ͱ ɾτʔΫϯԽʢ㲈ࣈ۟ղੳʣʢ㱭 ܗଶૉղੳʣ ɹࣗવݴޠ΍ϓϩάϥϛϯάݴޠͳͲͷจࣈྻΛղੳͯ͠ ɹτʔΫϯΛಘΔ͜ͱ ɹ˞ܗଶૉղੳͷղੳର৅͸ࣗવݴޠͷΈ
33. iOSʹ͓͚ΔτʔΫφΠζ
34. ೔ຊޠʹ΋௅ઓʂ
35. ·͋͜Μͳ͜ͱ΋͋ΔΑͶ…ʁ
36. ࡶஊɿίϯύΠϥ͕ίʔυΛʮཧղʯ͢Δʹ͸ ࣈ۟ղੳɺߏจղੳɺ࣮ߦίʔυੜ੒ͳͲͷաఔΛܦͯ ίϯύΠϥ͸ίʔυΛղऍ͢Δ
37. ࡶஊɿίϯύΠϥ͕ίʔυΛʮཧղʯ͢Δʹ͸ ࣈ۟ղੳͷաఔͰ͸ιʔείʔυΛϓϩάϥϜ্ͷ ࠷খ୯Ґʢҙຯͷ͋Δ୯Ґʣʹ෼ׂ͢Δ ྫɿwhile (abc < 10) { abc += 1 } (͜ͷྫจڐ͍ͯͩ͘͠͞) while / ( / abc / < / 10 / ) / { / abc / += / 1 / } /* hoge */ ͳͲͷίϝϯτ΋Ұմͱͯ͠ѻ͏
38. ࡶஊɿίϯύΠϥ͕ίʔυΛʮཧղʯ͢Δʹ͸ ߏจղੳͷաఔͰ͸ɺ༩͑ΒΕͨτʔΫϯྻ͔Βߏจنଇʹ ͕ͨͬͯ͠ߏจΛղੳ͢Δ 1 + 3 * ( 9 - 3 ) + - *
39. NLP APIͷॲཧͷྲྀΕ ݴ ޠ ൑ ఆ τ ʛ Ϋ ϯ Խ ඼ ࢺ λ ά ෇ ͚ Ϩ ϯ Ϛ Խ ݻ ༗ ද ݱ ந ग़
40. ඼ࢺλά෇͚ ɾ୯ޠʹରͯ͠඼ࢺΛ͚͍ͭͯ͘͜ͱ ɾ׆༻ྫʢࣙॻΞϓϦʣɿςΩετதͷ͋Δ୯ޠ͕ ɹಈࢺͳͷ໊͔ࢺͳͷ͔͕Θ͔͍ͬͯΕ͹ ɹΑΓద੾ͳ༁ΛৼΔ͜ͱ͕Ͱ͖Δɹ
41. ͲͷΑ͏ʹ඼ࢺΛܾఆͯ͠Δʁ ɾܗଶૉతͳख͕͔Γ͔Β ɹhappiness → ܗ༰ࢺ, establishment → ໊ࢺ ɾ౷ޠʢ㲈ߏจʣతͳख͕͔Γ͔Β ɹӳޠͰ͸ܗ༰ࢺ͸໊ࢺͷ௚લɺbeಈࢺͷ௚ޙʹஔ͔ΕΔ ɹ͜ͱ͕ଟ͍ɺͳͲ ɾҙຯతͳख͕͔Γ͔Β ɹ໊ࢺͱ͸ɺਓɺ৔ॴɺ෺ࣄͷ໊લͰ͋ΔɺͳͲ
42. iOSʹ͓͚Δλά෇͚
43. NLP APIͷॲཧͷྲྀΕ ݴ ޠ ൑ ఆ τ ʛ Ϋ ϯ Խ ඼ ࢺ λ ά ෇ ͚ Ϩ ϯ Ϛ Խ ݻ ༗ ද ݱ ந ग़
44. ϨϯϚԽ ࣙॻͷݟग़͠ޠΛநग़͢Δॲཧ ྫɿpresided → presideʢաڈܗ͔Βݱࡏܗ΁ʣ ɹɹhours → hourʢෳ਺ܗ͔Β୯਺ܗ΁ʣ
45. iOSʹ͓͚ΔϨϯϚԽ աڈܗ͔Βݱࡏܗʹͳ͍ͬͯΔ খจࣈʹͳ͍ͬͯͳ͍
46. NLP APIͷॲཧͷྲྀΕ ݴ ޠ ൑ ఆ τ ʛ Ϋ ϯ Խ ඼ ࢺ λ ά ෇ ͚ Ϩ ϯ Ϛ Խ ݻ ༗ ද ݱ ந ग़
47. ݻ༗දݱநग़ ɾݻ༗໊ࢺʢਓ໊ɺ஍໊౳ʣɺ࣌ؒදݱʢ೔෇ɺ࣌ࠁʣͳͲ ɹΛςΩετ͔Βநग़͢Δٕज़ ɾςΩετʹ͓͍ͯݻ༗දݱ͸ॏཁ৘ใͳͷͰ ɹ৘ใநग़΍νϟοτϘοτͳͲͷର࿩γεςϜʹ͓͍ͯ ɹ׆༻͞Ε͍ͯΔ
48. ݻ༗දݱநग़ͷख๏ ɾ஌ࣝϕʔεख๏ ɾ౷ܭతख๏
49. ஌ࣝϕʔεख๏ ɾنଇʢύλʔϯʣΛਓखͰ࡞੒͠ɺͦΕʹ౰ͯ͸·Δ ɹ΋ͷΛநग़͢Δ ɹྫɿ໊ࢺʴʮ͞Μʯˠਓ໊ ɹɹɹ໊ࢺʴʮࢢʯˠ஍໊ ɾنଇΛ໢ཏతʹߏங͠Α͏ͱ͢Δͱലେͳίετ͕͔͔Δ
50. ౷ܭతख๏ ɾڭࢣ͋ΓֶशΛͤ͞ɺநग़نଇΛࣗಈֶश͢ΔΑ͏ʹ͢Δ ɾSVMʢSupport Vector Machine / ࣝผؔ਺ʣ΍CRF ɹʢConditional Random Fields / ࣝผϞσϧʣ͕༗໊
51. iOSʹ͓͚Δݻ༗දݱநग़ joinNamesΛઃఆ͍ͯ͠ΔͷͰ ෳ߹ޠͰ΋நग़Մʂ
52. ·ͱΊ
53. ·ͱΊ ɾNLP API͸ཪଆͷࣗવݴޠॲཧΛҙࣝ͠ͳͯ͘΋खܰʹ ɹ࢖͏͜ͱ͕Ͱ͖Δ ɾࣗવݴޠॲཧ͸Ԟਂָ͍ͯ͘͠ʂʂʂ
54. ิ଍ɿؔ࿈ࢿྉҰཡ ɾNLP APIͰ͍Ζ͍Ζࢼͯ͠Έͨʢۙ೔ެ։ʣ ɹʢNDAͷؔ܎Ͱʣ ɾNatural Language APIͱͦͷཪଆΛ೷͘[׬શ൛] https://speakerdeck.com/akatsuki174/natural-language- apitosofalseli-ce-wosi-ku-wan-quan-ban ɾࣗવݴޠॲཧͱΞϓϦ http://qiita.com/akatsuki174/items/9f273ecfd29245a9f39d
55. ऴΘΓɻ͋Γ͕ͱ͏͍͟͝·ͨ͠ɻ ͋ͱɺϦδΣΫτίϯདྷͯͶɻ
56. དྷͦ͏ͳ࣭໰ͱճ౴ ※͋͘·Ͱ΋ࢲͷ༧૝ͱߟ͑Ͱ͋Γɺ ճ౴ͷਖ਼֬ੑ͸อূ͠·ͤΜɻ
57. NLP APIͲ͜Ͱ࢖͏ͷʁ ͍͍࣭໰Ͱɺ೉͍࣭͠໰Ͱ͢Ͷɻ ͿͬͪΌ͚ϩʔΧϧͰࣗવݴޠॲཧΛ΍Δ৔໘͸ͦΜͳ ͳ͍ͱࢥ͍·͢ɻ Ͱ͕͢ɺ͋͑ͯڍ͛Δͱͨ͠Β࣍ͷϝϦοτ͕͋Δͱࢥ͍· ͢ɻ
58. ϩʔΧϧͰNLPΛ΍ΔϝϦοτ ɾΦϑϥΠϯͰ΋࢖͑Δ ɾɾจষೖྗ͕සൟʹൃੜ͢ΔΞϓϦͳΒ಺෦ʹࣙॻ࣋ͬͯ ɹɹNLP APIΛ࢖ͬͯ΋͍͍͔΋ ɾɾݴޠ൑ఆ͔͠࢖Θͳ͍৔໘ͳΒɺΘ͟Θ͟αʔόαΠυ ɹɹ͔Βݺ͹ͳ͍͍ͯ͘ͷ͸ྑ͍…͔΋ʁ ɾΩʔϘʔυΞϓϦͳΒࠓ·ͰΑΓॲཧָʹॻ͚Δ͔΋ʁ
59. ߟ͑ΔͳɺՄೳੑΛײ͡Ζ i M e s s a g e N e w s NLPͰॲཧͨ͠΋ͷΛڞ༗͢Δ෦෼
60. ߟ͑ΔͳɺՄೳੑΛײ͡Ζ i M e s s a g e N e w s NLPͰॲཧͨ͠΋ͷΛڞ༗͢Δ෦෼ y o u r a p p