Text-similarity analysis for Shiganai Radio

by Takaaki Tanaka

Slide 1

Slide 1 text

͕͠ͳ͍ϥδΦΛ෼ੳͯ͠Έͨ ʢ·ͩಓ൒͹ʣ

Slide 2

Slide 2 text

ࣗݾ঺հ @in_the_rye SIerʹ10೥ۙ͘ۈ຿ʢݱࡏ͸ௐ੔ۀ຿onlyʣˠ৭ʑࢥ ͏ͱ͜Ζ͕͋ͬͯస৬༧ఆɻ ਓੜͰॳΊͯͷLTͰ͢ʂ 2

Slide 3

Slide 3 text

෼ੳͷϞνϕʔγϣϯ ͕͠ͳ͍ϥδΦΛep͔̍Βௌ͖࢝Ίͯɺ࠷ۙ Α͏΍͘࠷৽ճʹ௥͍ͭ͘ɻˠ̍೥͘Β͍ ͔͔ͬͨɻ શ෦ฉ͘ͷ͸ָ͍͚͠Ͳɺͪΐͬͱେมɻ ஍ਤ͕͋ΔͱͪΐͬͱָͳͷͰ͸ʁ ͕͠ͳ͍ϥδΦͷ֤Τϐιʔυͷத਎͕ɺ ΋͏গ͠Θ͔ΔΑ͏ͳ஍ਤΛ࡞Ζ͏ʂ

Slide 4

Slide 4 text

෼ੳʁͬͯԿΛͨ͠ͷʁ Ի੠͔Βจࣈى͜͠ ܗଶૉղੳ TF-IDFม׵ͰϕΫτϧԽ ίαΠϯྨࣅ౓ܭࢉ

Slide 5

Slide 5 text

͓લ͸Ұମ ԿΛݴ͍ͬͯΔΜͩʁ

Slide 6

Slide 6 text

આ໌͠·͢ʂ ʢάάͬͯಘͨ஌ࣝͰɻɻɻʣ

Slide 7

Slide 7 text

จࣈى͜͠ Google Cloud Text-to-speech Api Λར༻ɻ ݄60෼·Ͱͷจࣈى͜͠͸ແྉɻͦΕ Ҏ߱͸ैྔ՝ۚɻ શΤϐιʔυ͸ਏ͍ͷͰɺϥϯμϜʹબ Μͩ19ΤϐιʔυΛจࣈى͜͠ɻ GCPͷແྉ࿮Ͱˇ5,406՝ۚɻɻɻ

Slide 8

Slide 8 text

ܗଶૉղੳ pythonͷϥΠϒϥϦΛར༻ͯ͠จষΛ୯ޠʹ෼ׂ

Slide 9

Slide 9 text

TF-IDFม׵ͰจষΛϕΫτϧԽ TFʢӳ: Term Frequencyɺ୯ޠͷग़ݱස౓ʣ IDFʢӳ: Inverse Document Frequencyɺٯจॻස౓ʣ TFͱIDFͷೋͭͷج४Ͱͷܭࢉ஋Λֻ͚ࢉͯ͠୯ޠʹॏΈΛ͚ͭΔɻˠ୯ ޠͱॏΈࣗମ͕ϕΫτϧͱͳΔͷͰɺυΩϡϝϯτΛϕΫτϧԽͰ͖Δʂ ࡾ࣍ݩϕΫτϧ(x, y, z) = (1, 2, 3) ࠲ඪͱͦΕͧΕͷ஋ υΩϡϝϯτͷϕΫτϧ(͕͠ͳ͍, ϥδΦ, SIer) = (1, 2, 3) ୯ޠͱͦΕͧΕͷ஋ X Y Z

Slide 10

Slide 10 text

TF-IDFม׵ͱ͸ཁ͢Δʹ ߴස౓ͷ୯ޠ΄Ͳ఺਺͕ߴ͍ɻ ϨΞͳ୯ޠ΄Ͳ఺਺͕ߴ͍ɻ ্ͷೋͭͷ఺਺ͷֻ͚ࢉͰ୯ޠͷ఺਺ΛܾΊΔΑʂ ୯ޠ͝ͱͷ఺਺͕͍ͭͨΒͦΕ͕ϕΫτϧͩʂ

Slide 11

Slide 11 text

ίαΠϯྨࣅ౓ ϕΫτϧͷͳ֯͢ΛϺͱͨ࣌͠ͷ cosΘ ಉ͡ํ޲Λ޲͍͍ͯΕ͹cosΘ = 1 90౓ͳΒcosΘ = 0 ຊ౰͸ਅٯ(180౓)ͳΒ-1ʹͳ Δ͚Ͳɺ͜͜Ͱ͸஋͸ਖ਼ͷ஋ ͷΈͳͷͰɺ0ʙ1ͷൣғͷΈɻ Θ

Slide 12

Slide 12 text

෼ੳ݁Ռʢ19ΤϐιʔυͷΈʣ cosྨࣅ౓ ͷϚοϓ

Slide 13

Slide 13 text

SP8A-TBPGR

Slide 14

Slide 14 text

SP38A-HIGUYUME

Slide 15

Slide 15 text

SP16A-INFRAGIRL755

Slide 16

Slide 16 text

Ͳͳ͕ͨήετͷճͰ͠ΐ͏͔ʁ

Slide 17

Slide 17 text

·ͱΊ OSSͷϥΠϒϥϦͳͲΛ࢖͕ͬͯ͠ͳ͍ϥδΦͷ෼ੳΛͨ͠ɻ ͲͷΤϐιʔυಉ͕࢜ࣅ͍ͯΔ͔ͳͲͷϚοϓΛ࡞ͬͨɻ ࢥ͍ͷ΄͔ɺͦΕͬΆ͍݁Ռ͕ग़͖ͯͨɻ ͚ͲΑ͘ݟΔͱมͳ୯ޠ͕ೖ͍ͬͯͯɺจࣈى͜͠ͷਫ਼౓͕ͦ͜· Ͱྑ͘ͳ͍ͨΊͱࢥΘΕΔɻ ର৅Τϐιʔυ਺͕·ͩ19Τϐιʔυ͚ͩͳͷͰɺGCPͷແྉΫϨ δοτ࿮಺Ͱ΋͏ͪΐͬͱ෼ੳͯ͠ΈΑ͏ͱࢥ͍·͢ɻ

Slide 18

Slide 18 text

͝੩ௌ͋Γ͕ͱ͏͍͟͝·ͨ͠