Introduction to Dialog System

Transcript

1. ؀ച दझమਸ ଺ইࢲ 2019/05/15 Final draft NLP Engineer ҳ ࢚ ળ koosangjun@scatterlab.co.kr
2. ؀ച दझమ੄ ҳࢿ Dialog System
3. Construction of Dialog System • ؀ച दझమ੄ ҳࢿ • (ޛܻ੸ੋ नഐܳ ఫझ౟۽ ߸ജೞח ࠗ࠙) • ਺ࢿ ੋधӝ = Automatic Speech Recognition (ASR) • Ӗ੗ ੋधӝ = Optical Character Recognition (OCR) • ఫझ౟ܳ ੉೧ೞח ࠗ࠙ = Natural Language Understanding (NLU) unit • ੉೧ೠ ఫझ౟۽ ؀ച੄ ਽׹ਸ ੉যաоѱ ೞח ࠗ࠙ = Dialog Manager (DM) • ࢎۈ੉ ੉೧ೡ ࣻ ੓ח ഋక۽ ਽׹ਸ ࢤࢿ = Natural Language Generation (NLG) • (਽׹ਸ ޛܻ੸ੋ नഐ۽ ߸ജೞח ࠗ࠙) • ਺ࢿ ୹۱ӝ = Text-To-Speech (TTS) • ҳ࠙਷ ੺؀੸੉૑ ঋ਺ • NLU + DM / DM + NLG / NLU + DM + NLG (End-to-End Dialog System) /-6 %. /-(
4. Construction of Dialog System ASR : Automatic Speech Recognition NLU : Natural Language Understanding DM : Dialog Management NLG : Natural Language Generation TTS : Text to Speech User Utterance Text System Utterance Text ࢎਊ੗ 어벤져스 어때? 어벤져스 봤어요! 토르 알아요? ؀ച दझమ • ؀ച दझమ੄ ҳࢿ (Black Box) • ࢎਊ੗о ߄ۄࠁח ੑ੢ীࢲ੄ ؀ച दझమ • ࢎਊ੗ח যڃ ݽٕ੉ ٜযоࢲ ਽׹ਸ ղח૑ ੋ૑ೞ૑ ঋ਺
5. Construction of Dialog System • ؀ച दझమ੄ ҳࢿ (White Box) • ӝദ੗о / ѐߊ੗о ߄ۄࠁח ੑ੢ীࢲ੄ ؀ച दझమ • ӝദ੗ա ѐߊ੗ח п ݽٕ੉ যڌѱ ҳࢿغযঠೞח૑ ౵ঈ೧ঠೣ ASR : Automatic Speech Recognition NLU : Natural Language Understanding DM : Dialog Management NLG : Natural Language Generation TTS : Text to Speech User Utterance Text Semantic Frame System Action System Utterance Text User UA : ask-opinion NE : movie = 어벤져스 SA : say-opinion, ask-actor Movie = য߮ઉझ <movie> 봤어요! <actor> 알아요? 어벤져스 어때?
6. Construction of Dialog System • ؀ച दझమ੄ ҳࢿ (White Box) • ӝദ੗о / ѐߊ੗о ߄ۄࠁח ੑ੢ীࢲ੄ ؀ച दझమ • ӝദ੗ա ѐߊ੗ח п ݽٕ੉ যڌѱ ҳࢿغযঠೞח૑ ౵ঈ೧ঠೣ ASR : Automatic Speech Recognition NLU : Natural Language Understanding DM : Dialog Management NLG : Natural Language Generation TTS : Text to Speech User Utterance Text Semantic Frame System Action System Utterance Text User UA : ask-opinion NE : movie = 어벤져스 SA : say-opinion, ask-actor Movie = য߮ઉझ <movie> 봤어요! <actor> 알아요? 어벤져스 어때?
7. Construction of Dialog System (NLU, DM) • ؀ച दझమ੄ ҳࢿ (NLU) • ޙ੢ਸ ஹೊఠо ੉೧ೞח ഋక۽ ߸ജೞח ݽٕ • ؀ۚ੸ੋ NLU੄ ҳࢿ • Normalizer + Tokenizer (ఫझ౟ ੿ઁ) • Intent Classification (੄ب ౵ঈ) • Named Entity Recognition (ѐ୓ݺ ੋध) • Sentiment Analyzer (ࢎਊ੗ ӝ࠙ ౵ঈ) • Coreference Resolution (ഐட ೧ࣗ) • Topic Extraction (઱ઁ ੋध) • … • ؀ച दझమ੄ ҳࢿ (DM) • NLU੄ Ѿҗ۽ ਽׹ਸ ٜ݅যղח ݽٕ • ؀ۚ੸ੋ DM੄ ҳࢿ • System act selection (दझమ ੄ب ࢤࢿ) • Context manager (ޙݓ ݒפ੷) • External Knowledge Graph (৻ࠗ ૑ध Ӓې೐) • যו ࠗ࠙੉ NLUҊ যו ࠗ࠙੉ DMੋо? (VOSPDL
   /-6 (VOSPDL
   /-6
8. Construction of Dialog System (NLG, ASR, TTS) • ؀ച दझమ੄ ҳࢿ (NLG) • ࢎਊ੗о ੉೧ೡ ࣻ ੓ח ޙ੢ਸ ٜ݅যղח ݽٕ • ؀ۚ੸ੋ NLG੄ ҳࢿ • Answer Template DB (਽׹ మ೒݁ DB) • Answer Sentence Decoder (਽׹ ޙ੢ ٣௏؊) • Post Processor (റ୊ܻ ࠗ࠙) • Prosody Manager (ߊ਺ ୊ܻӝ, TTS) • …. • ؀ച दझమ੄ ҳࢿ (ASR) • ࢎਊ੗ ߊ਺ਸ ੋधೞৈ ఫझ౟۽ ٜ݅যղח ݽٕ • ؀ച दझమ੄ ҳࢿ (TTS) • ఫझ౟ܳ ੋр੉ ੋधೡ ࣻ ੓ח ਺ࢿਵ۽ ٜ݅য઱ח ݽٕ
9. Construction of Dialog System (Quiz) • ؀ച दझమ੄ ҳࢿ (ೝಯ੄ ৘द) • (ޛܻ੸ੋ नഐܳ ఫझ౟۽ ߸ജೞח ࠗ࠙) : • ఫझ౟ܳ ੉೧ೞח ࠗ࠙ = Natural Language Understanding (NLU): • ੉೧ೠ ఫझ౟۽ ؀ച੄ ਽׹ਸ ੉যաоѱ ೞח ࠗ࠙ = Dialog Manager (DM): • ࢎۈ੉ ੉೧ೡ ࣻ ੓ח ഋక۽ ਽׹ਸ ࢤࢿ = Natural Language Generation (NLG): • (਽׹ਸ ޛܻ੸ੋ नഐ۽ ߸ജೞח ࠗ࠙): • ݏ୾ࠁب۾ ೤द׮
10. Construction of Dialog System (Quiz) • ؀ച दझమ੄ ҳࢿ (ೝಯ੄ ৘द) • (ޛܻ੸ੋ नഐܳ ఫझ౟۽ ߸ജೞח ࠗ࠙) : হ਺, ਗې ఫझ౟פө! • i.e.,) ೲࣁ઺, ܖա := ҳӖ ਺ࢿੋधӝ • ఫझ౟ܳ ੉೧ೞח ࠗ࠙ = Natural Language Understanding (NLU): • Reply Matching Model • Reaction Model • ੉೧ೠ ఫझ౟۽ ؀ച੄ ਽׹ਸ ੉যաоѱ ೞח ࠗ࠙ = Dialog Manager (DM): • ID-OOD Ѿ੿, RvR Ѿ੿ • ࢎۈ੉ ੉೧ೡ ࣻ ੓ח ഋక۽ ਽׹ਸ ࢤࢿ = Natural Language Generation (NLG): • RMM మ೒݁ ਽׹ ࢤࢿ • Reaction మ೒݁ ਽׹ ࢤࢿ • (਽׹ਸ ޛܻ੸ੋ नഐ۽ ߸ജೞח ࠗ࠙): হ਺, ਗې ఫझ౟פө! • I.e.,) ೲࣁ઺, ܖա := ҳӖ TTS • Pingpong਷ যڌѱ ࠁݶ End-to-End System: • ੑ۱ীࢲ ୹۱ө૑ ݽ؛ 1ױ҅ (RMM, Reaction)ਸ ా೧ࢲ ੉ܖয૗
11. Construction of Dialog System (Quiz) • য٣ী ೧׼غח ݽٕੌө? (ઙਮש ই੉٣যীࢲ ߊ஀) • ҕх, ૕ޙ, ٜয઱ӝ ܳ ੜೞח ݽٕ੉ ੓঻ਵݶ જѷ׮. • ࢎਊ੗ী ؀ೠ ౠ੿ೠ ੿ࠁܳ ׮਺ী ࠁյ ݫࣁ૑۽ ߸ജೞח ݽ؛ • ઱য૓ ޙ੢੄ ݈ైܳ ੄بա ॳ੐ী ٮۄ ߸ജदఃח ݽ؛ • ஶబஎ ݒפ૑ݢ౟ - ׮নೠ ઱ઁ৬ ഋకܳ ӝദ੗о ֍ਵݶ ੸੺ೞѱ ߜয઱ח ݽ؛ • ޙ੢੄ п ਃࣗо য٣ী োѾغয੓ח૑ܳ ׳ইષ • ࢤп೧ࠁب۾ ೤द׮
12. Construction of Dialog System (Quiz) • য٣ী ೧׼غח ݽٕੌө? (ઙਮש ই੉٣যীࢲ ߊ஀) • ҕх, ૕ޙ, ٜয઱ӝ ܳ ੜೞח ݽٕ੉ ੓঻ਵݶ જѷ׮. • Dialog Managementী оө਑ • Ӓ۞ա ࢎਊ੗੄ х੿ਸ ੍যղח ੘স੉ ೙ਃೞפө NLUীب ೧׼ؼ ࣻ ੓਺ • ࢎਊ੗ী ؀ೠ ౠ੿ೠ ੿ࠁܳ ׮਺ী ࠁյ ݫࣁ૑۽ ߸ജೞח ݽ؛ • Dialog Managementী ࣘೣ • NLGীب ࣘೡࣻب ੓ਸ૑ب? • ઱য૓ ޙ੢੄ ݈ైܳ ੄بա ॳ੐ী ٮۄ ߸ജदఃח ݽ؛ • NLGী ࣘೡ Ѫ э਺ • ஶబஎ ݒפ૑ݢ౟- ׮নೠ ઱ઁ৬ ഋకܳ ӝദ੗о ֍ਵݶ ੸੺ೞѱ ߜয઱ח ݽ؛ • Dialog Managementী ࣘೣ • ޙ੢੄ п ਃࣗо য٣ী োѾغয੓ח૑ܳ ׳ইષ • NLUী ࣘೣ • Ӓ۞ա ৮߷ೠ ੿׹਷ ইש. = NLU / DM / NLG ࢎ੉੄ ҳ࠙਷ ੼੼ ݽഐ೧૑ח ୶ࣁ
13. ؀ച दझమ੄ ಣо Evaluation of Dialog System
14. Evaluation Metrics • ؀ച दझమ਷ যڌѱ ಣоغযঠೞա? • ੉ दझమ੉ ਋ܻо ӝ؀ೠ ׹߸ਸ ೞחо? • ੿۝੸ ҙ੼ • Precision@N : ਋ܻо ӝ؀೮؍ ׹߸(ٜ)੉ ࣽਤ N উী ੓חо • ೤׼ೠ ׹߸੉ 3ѐо ࢚ਤ 10ѐী ੓঻ਵݶ P@10 = 0.3 • MRR : ӝ؀ೞח ׹߸੄ ࣽਤ੄ ৉ࣽ੄ ಣӐ • ৘ܳ ٜয 3ѐ੄ ௪ܻী ؀೧ࢲ ղ ੸੺ೠ ׹߸੉ 1ਤ, 3ਤ, 10ਤܳ ೮׮ݶ • (1/1+1/3+1/10) / 3 • NDCG : • ৘ܳ ٜয ղо մ ਽׹ 6ѐо ࣽࢲ؀۽ 3,2,3,0,1,2੼੉঻׮Ҋ о੿ • ୐ߣ૩۽ աৡ ਽׹੉ ੜ ݏҊ, ݃૑݄ਵ۽ աৡ ਽׹਷ ખ ؏ݏইب ػ׮Ҋ ೞݶ • ੼ࣻܳ ࣽਤ੄ ৉ࣻ (੿ഛ൤ח ࣽਤ੄ ৉ࣻ੄ log) ғ೧ࢲ ҅࢑ೡ ࣻ ੓਺ • ੉ۧѱ ҅࢑ೠ ੼ࣻо DCG੐ • NDCG = അ੤ DCGܳ о੢ ੉࢚੸ੋ ҃਋੄ DCG۽ ա׀ ҃਋ • ੿ࢿ੸ ҙ੼ • ҙ۲ࢿ ੼ࣻ ୋب ಣо • ੸೤ࢿ ੼ࣻ ୋب ಣо
15. Evaluation Metrics • ؀ച दझమ਷ যڌѱ ಣоغযঠೞա? • ੉ दझమ਷ ࢎۈҗب э਷ ਽׹ਸ ೞחо? • ࢎۈҗب э਷ ਽׹਷ যڌѱ ٜ݅૑? • Wizard-of-OZ (WOZ, aka ֢о׮): • ࢎਊ੗ח दझమ੉ۄҊ ࢤпೞҊ ੑ۱ೣ + दझమ ٍীח ѐߊ੗о ׹߸ೣ • ѐߊ੗ ׹߸ = ࢎۈэ਷ ਽׹ • ੿۝੸ ҙ੼ • ਬࢎب ಣо (ױয ੐߬٬ ױਤ, ޙ੢ ੐߬٬ ױਤ) • BLEU ١੄ ࠺Ү ಣо • ੿ࢿ੸ ҙ੼ • ࠶ۄੋ٘ పझ౟ (ౚ݂ పझ౟) • ࢎۈੋ૑ ࠙рೡ ࣻ ੓ח૑ܳ ಣо
16. ؀ച दझమ੄ ҳഅ Implementation of Dialog System
17. Revisited: Problem Description • ஹೊఠח झझ۽ ࢎҊೞ૑ ޅೣ • 1+1 ਸ 2۽ ҅࢑ೡ ࣻ ੓যب 1+1੉ ৵ 2ੋ૑ ࢸݺೞ૑ ޅೣ • ஹೊఠ۽ ޙઁܳ ಽҊ੗ ೞݶ ޙઁܳ ҳ୓੸ਵ۽ ੿੄೧ঠೣ (Algorithm) • উ જ਷ ਬഋ : • ੉ Ӗী х੿ ࠙ࢳ೧઻ • ੉ ੘о ݅ചೂ Ӓܿ Ӓ۰઻ • ӡ ଺ই઻ • જ਷ ਬഋ: • ఫझ౟ Tо ઱য઎ਸ ٸ, х੿ కӒ (न܉х, গ੿х, ஘޻х)ী ؀ೠ чਸ ب୹೧઻ • Ӓܿ ૘೤ Aо ઱য઎ਸ ٸ Ӓܿ Bܳ ૘೤ A۽ ੋधदఆ ࣻ ੓ب۾ ߸ജ೧઻ • ࢶҗ ੼ਵ۽ োѾػ ب۽ Ӓې೐ Gо ੓ਸ ٸ G੄ ೠ੼ X1ীࢲ ׮ܲ ੼ X2ө૑੄ о੢ ࡅܲ ҃۽ܳ ҳ೧઻
18. NLU੄ ҳഅ Implementation of NLU
19. Revisited (NLU) • ؀ച दझమ੄ ҳࢿ (NLU) • ޙ੢ਸ ஹೊఠо ੉೧ೞח ഋక۽ ߸ജೞח ݽٕ • ؀ۚ੸ੋ NLU੄ ҳࢿ • Normalizer + Tokenizer (ఫझ౟ ੿ઁ) • Intent Classification (੄ب ౵ঈ) • Named Entity Recognition (ѐ୓ݺ ੋध) • Sentiment Analyzer (ࢎਊ੗ ӝ࠙ ౵ঈ) • Coreference Resolution (ഐட ೧ࣗ) • Topic Extraction (઱ઁ ੋध) • … • ੉Ѫਸ যڌѱ ࣻ೟੸ਵ۽ աఋյ ࣻ ੓ਸө? • Intent Classification = ޙ੢੉ ઱য઎ਸ ٸ ௿ېझ ೞաܳ ଺ח ੘স • Named Entity Recognition = ޙ੢੉ ઱য઎ਸ ٸ ؀਽غח ױযৌਸ ଺ח ੘স (VOSPDL
    /-6
20. Revisited: What is a sentence? • ޙ੢੉ۄח Ѫ਷ ޖ঺ੋо? • ച੗о ಴അೞ۰ח ڷਸ աఋղח ױয੄ (ӏ஗੸ੋ) োࣧ • ױࣽೠ ױয੄ ૘ױ Ӓ ੉࢚੐: {ࡎ, ա, ݍ} • ӏ஗੉ হਵݶ ઁ؀۽ ػ ޙ੢੉ۄҊ ೞӝ য۰਑: ࡎਸ ղо ݍ੓঻׮ • ڷ੉ ݏ૑ ঋਵݶ ઁ؀۽ ػ ޙ੢੉ۄҊ ೞӝ য۰਑: աח ࡎਸ ޖପۥ׮ • ޙ੢ਸ ࠁח ف о૑ ҙ੼ • ഋध઱੄੸ ҙ੼: ޙ੢਷ ࢎۈ੉ ڷਸ "ޙߨ"ী ݏѱ ӝࣿೞ۰Ҋ ߊ੹ػ Ѫ੉׮ • ޙ੢ ҳޙ ࠙ࢳ ١ • ӝמ઱੄੸ ҙ੼: ޙ੢਷ ࢎۈ੉ "ڷ"ਸ ޙߨী ݏѱ ӝࣿೞ۰Ҋ ߊ੹ػ Ѫ੉׮ • ঱য ݽ؛ ߂ ޙ੢ ੐߬٬ ١ • ਋ܻח റ੗੄ ߑߨী ઑӘ ؊ ૘઺ೞҊ੗ ೣ
21. Revisited: Language Model • ঱য ݽ؛ (Language Model) • ޙ੢ਸ ࣻ೟੸ਵ۽ ӝࣿೞח ૒ҙ੸ੋ ߑߨ • ױয੄ ੄ઓҙ҅ח খ ױযীࢲࠗఠ ഋࢿػ׮Ҋ о੿ • LM(ա য়ט ࡎ ݡ঻য) = P(ա) * P(য়ט | ա) * P(ࡎ | য়ט, ա) * P (ݡ঻য | ࡎ, য়ט, ա) • ঱য ݽ؛੄ ഝਊ • ঱য ݽ؛ਸ ੉ਊೠ ৘ஏ • ޙઁ : ա য়ט _ ݡ঻য • ੿׹ : Argmax (P(ա) * P(য়ט | ա) * P(_ | য়ט, ա) * P (ݡ঻য | _, য়ט, ա)) • ঱য ݽ؛ਸ ੉ਊೠ ޙ੢੄ ੸೤ࢿ Ѩࢎ • ޙઁ : ա য়ט ࡎ ݡ঻য <=> ա ࡎਸ ޖପۥয • ੿׹ : LM (ա য়ט ࡎ ݡ঻য) > LM (ա ࡎਸ ޖପۥয) • ঱য ݽ؛੄ рۚച • P(A) * P(B | A) * P(C | B, A) ...= ҅࢑۝੉ ցޖ ݆਺ • P(X | Y, Z, . ..) ܳ P(X | Y)۽ Ӕࢎೡ ࣻ ੓਺ (૊, ّ ױযח ߄۽ খ ױযী ੄೧ࢲ Ѿ੿ؽ) • Markov Assumption (഑਷ Markov Property Assumption) ੉ۄҊ ܴࠗ
22. Revisited: Language Model • ঱য ݽ؛ (Language Model) җ ഛܫ ݽ؛ • LM(ա য়ט ࡎ ݡ঻য) = P(ա) * P(য়ט | ա) * P(ࡎ | য়ט) * P(ݡ঻য | ࡎ) • ঱য ഛܫ ݽ؛ਸ ా೧ ޙ੢੄ ਗੋਸ ӝࣿೡ ࣻ ੓਺ • ਗੋ • ޙ੢੄ ਗې ੄ب, ച੗੄ х੿, ݺ۸/ಣࢲ/૕ޙ ৈࠗ ١١ • ޙ੢ਸ աఋղח ೞա੄ ߭ఠ ա য়ט ࡎ ݡ঻য
23. Revisited: Language Model • ঱য ݽ؛ (Language Model) җ ഛܫ ݽ؛ • LM(ա য়ט ࡎ ݡ঻য) = P(ա) * P(য়ט | ա) * P(ࡎ | য়ט) * P(ݡ঻য | ࡎ) • ঱য ഛܫ ݽ؛ਸ ా೧ ޙ੢੄ ਗੋਸ ӝࣿೡ ࣻ ੓਺ • ਗੋ • ޙ੢੄ ਗې ੄ب, ച੗੄ х੿, ݺ۸/ಣࢲ/૕ޙ ৈࠗ ١١ • ޙ੢ਸ աఋղח ೞա੄ ߭ఠ • ૒ҙ • ޙ੢੄ ੄޷ܳ ঱য ݽ؛ਸ ా೧ Ѩ୹ೡ ࣻ ੓૑ ঋਸө? ա য়ט ࡎ ݡ঻য
24. NLU (Intention Classifier) • Intention Classifier = Class Classification • ঱য ݽ؛ਸ ా೧ࢲ ௿ېझ Ѩ୹ӝܳ ҳഅೡ ࣻ ੓਺ • ঱য ݽ؛੄ Ӓܿਸ Ӓ۰ࠁ੗
25. NLU (Intention Classifier) • Intention Classifier = Class Classification • ঱য ݽ؛ਸ ా೧ࢲ ௿ېझ Ѩ୹ӝܳ ҳഅೡ ࣻ ੓਺ • যڌѱ ঱য ݽ؛ਸ ҳഅೡ ࣻ ੓ਸ Ѫੋо? • ഛܫݽ؛۽? न҃ݎਵ۽? ա য়ט ࡎ ݡ঻য &BU 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 Regression Layer PGM Inference (HMM, MaxEnt) DNN Regression (Bi-LSTM, GRU)
26. NLU (Named Entity Recognition) • Named Recognition = Sequence Classification • ঱য ݽ؛ਸ ా೧ࢲ ޙ੗ৌ Ѩ୹ӝب ҳഅೡ ࣻ ੓਺ • ঱য ݽ؛ਸ Ӓ۰ࠁ੗
27. NLU (Named Entity Recognition) • Named Recognition = Sequence Classification • ঱য ݽ؛ਸ ా೧ࢲ ޙ੗ৌ Ѩ୹ӝب ҳഅೡ ࣻ ੓਺ • যڌѱ ঱য ݽ؛ਸ ҳഅೡ ࣻ ੓ਸ Ѫੋо? • ഛܫݽ؛۽? न҃ݎਵ۽? ա য়ט ࡎ ݡ঻য 0 0 #'PPE 0 1 2 3 4 5 6 7 1 2 3 4 5 6 7 PGM Inference (CRF) DNN Seq-to-Seq (Bi-LSTM, GRU) 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7
28. NLU (Coreference Resolution) • Coreference Resolution = Pair-wise Classification • ঱য ݽ؛ਸ ా೧ࢲ ഐட ಴അب ੜ ೧ࣗೡ ࣻ ੓਺ • ঱যݽ؛ਸ Ӓ۰ࠁ੗
29. NLU (Coreference Resolution) • Coreference Resolution = Pair-wise Classification • ঱য ݽ؛ਸ ా೧ࢲ ഐட ಴അب ੜ ೧ࣗೡ ࣻ ੓਺ 4IF -PWFT )JN 7FSZ .VDI .Z 4JTUFS )BT " %PH
30. DM੄ ҳഅ Implementation of DM
31. Revisited (DM) • ؀ച दझమ੄ ҳࢿ (DM) • NLU੄ Ѿҗ۽ ਽׹ਸ ٜ݅যղח ݽٕ • ؀ۚ੸ੋ DM੄ ҳࢿ • System act selection (दझమ ੄ب ࢤࢿ) • Context manager (ޙݓ ݒפ੷) • External Knowledge Graph (৻ࠗ ૑ध Ӓې೐) • যڌѱ ਽׹ਸ ࢤࢿೡ Ѫੋо? • ৘ઁ ӝ߈ ೐ۨ੐ਕ௼ (= ࢎۈҗ оӰѱ ਽׹ਸ ٜ݅যࠁ੗.) • ఃਕ٘: Supervised Learning / Accuracy / Similarity • ੼ࣻ ӝ߈ ೐ۨ੐ਕ௼ (= ੼ࣻܳ ֫ѱ ӝ۾ೞח ਽׹ਸ ٜ݅যࠁ੗.) • ఃਕ٘: Reinforcement Learning / Reward / Task Success Rate • যڌѱ ೞݶ ਽׹ਸ ੜ ٜ݅ ࣻ ੓ਸө?: • যڌѱ ೞݶ ޙݓਸ ߈৔ೠ ਽׹ਸ ٜ݅য յ ࣻ ੓ਸө? • যڌѱ ೞݶ ߓ҃૑धਸ ߈৔ೠ ਽׹ਸ ٜ݅য յ ࣻ ੓ਸө? (VOSPDL
    /-6
32. DM Learning (Single Turn) • ؀ചী ؀೧ࢲ ૕ޙ (Query)৬ ਽׹ (Reply) ह੉ ੓׮Ҋ о੿ • যڃ Queryо ٜযৡ ҃਋ : જ਷ ؀ച ݽ؛਷ Queryী ؀ೠ оө਍ ਽׹ਸ ଺਺ • ؀ۚ੸ੋ Ӓܿਸ Ӓ۰ࠁݶ
33. DM Learning (Single Turn) • ؀ചী ؀೧ࢲ ૕ޙ (Query)৬ ਽׹ (Reply) ह੉ ੓׮Ҋ о੿ • যڃ Queryо ٜযৡ ҃਋ : જ਷ ؀ച ݽ؛਷ Queryী ؀ೠ оө਍ ਽׹ਸ ଺਺ • ؀ۚ੸ੋ Ӓܿਸ Ӓ۰ࠁݶ ա য়ט ࡎ ݡ঻য য Ӓې աب ݡ঻য 2VFSZ 3FQMZ
34. DM Learning (Single Turn) • ؀ചী ؀೧ࢲ ૕ޙ (Query)৬ ਽׹ (Reply) ह੉ ੓׮Ҋ о੿ • যڃ Queryо ٜযৡ ҃਋ : જ਷ ؀ച ݽ؛਷ Queryী ؀ೠ оө਍ ਽׹ਸ ଺਺ • ؀ۚ੸ੋ Ӓܿਸ Ӓ۰ࠁݶ ա য়ט ࡎ ݡ঻য য Ӓې աب ݡ঻য 2VFSZ 3FQMZ
35. DM Learning (Single Turn) • ؀ചী ؀೧ࢲ ૕ޙ (Query)৬ ਽׹ (Reply) ह੉ ੓׮Ҋ о੿ • যڃ Queryо ٜযৡ ҃਋ : જ਷ ؀ച ݽ؛਷ Queryী ؀ೠ оө਍ ਽׹ਸ ଺਺
36. DM Learning (Single Turn + Constraint) • Query੄ ఃਕ٘ ੿ࠁܳ ഝਊೞӝ ਤ೧ࢲח যڌѱ ೧ঠೡө? • ց “ف࢑” જই ? —> ੷ب “ف࢑”੉ જইਃ . • Ӓܿਸ Ӓ۰ࠁݶ ց tف࢑u જই 2VFSZ 3FQMZ tف࢑u ੉ જইਃ ੷ب
37. DM Learning (Multi Turn) • ݣ౭ఢب Ӓܿਸ Ӓ۰ࠁ੗! • ؀ۚ੸ੋ Ӓܿਸ Ӓ۰ࠁݶ • ݣ౭ఢ ೟ण੄ য۰਑ • ݽٚ ੿ࠁо ز١ೞѱ ഝਊغח Ѫ੉ ইש • ૑դ ఢী ؀ೠ ੄ઓҙ҅о ݽ؛ ೟ण੸ਵ۽ ҳഅغӝ য۰਑ • ੿ࠁо ٜযоח ࣽࢲо ઺ਃೣ + ੿ࠁী ٮܲ ઑѤٜ ҳഅ੉ ೙ਃ ա য় ࡎ ݡ 2VFSZ
     য Ӓ ա ݡ 3FQMZ
    
    ց ޤ ݡ 2VFSZ
     ա ೫ ݡ ৳ 3FQMZ
     ೫ ա જ ൫ 2VFSZ
     3FQMZ
    
    
38. DM Learning (Multi Turn) • ݣ౭ఢ ঌҊ્ܻ ৘द (Yu et al. 2019, Computational Linguistics)
39. DM Learning (Multi Turn) • ݣ౭ఢ ঌҊ્ܻ ৘द (Yu et al. 2019, Computational Linguistics)
40. DM Learning (Knowledge) • ؀ചী যڌѱ ૑धਸ ׸ਸ ࣻ ੓ਸө? • ژ ৉द Ӓܿਸ Ӓ۰ࠁ੗
41. DM Learning (Knowledge) • ؀ചী যڌѱ ૑धਸ ׸ਸ ࣻ ੓ਸө? • ژ ৉द Ӓܿਸ Ӓ۰ࠁ੗ • ੉ Ӓܿ੄ ޙઁ੼਷ ޖ঺ੌө? ա য়ט ࡎ ݡ঻য য Ӓې աب ݡ঻য 2VFSZ 3FQMZ ,OPXMFEHF
    &NCFEEJOH
42. DM Learning (Knowledge) • ؀ചী যڌѱ ૑धਸ ׸ਸ ࣻ ੓ਸө? • ׮ܲ Ӓܿ (ױয пп੄ ߓ҃૑धਸ ߈৔) ա য়ט ࡎ ݡ঻য য Ӓې աب ݡ঻য 2VFSZ 3FQMZ
43. DM Learning (Knowledge) • ؀ചী যڌѱ ૑धਸ ׸ਸ ࣻ ੓ਸө? • ׮ܲ Ӓܿ (ױয пп੄ ߓ҃૑धਸ ߈৔) ա য়ט ࡎ ݡ঻য য Ӓې աب ݡ঻য 2VFSZ 3FQMZ X = hs G = sponse e prob- graphs ds in a in G 2. , τNgi } oted as the en- bridge nstruc- urpose: , t) = formed knowledge graphs. The entity is selected by attending on the graphs and the triples within each graph. 3.3 Knowledge Interpreter h t-1 Knowledge Interpreter h t Knowledge Interpreter h t+1 Knowledge Interpreter … … rays of sunlight Knowledge Graph Knowledge Graph Knowledge Graph Word Vector Key Entity Neighboring Entity Not_A_Fact Triple Retrieved Graph Figure 3: Knowledge interpreter concatenates a word vector and the national Joint Conference on Artificial Intelligence (IJCAI-18) gi = Ngi n=1 αs n [hn; tn], (4) αs n = exp(βs n ) Ngi j=1 exp(βs j ) , (5) βs n = (Wrrn)⊤tanh(Whhn + Wttn), (6) where (hn, rn, tn) = kn , Wh, Wr, Wt are weight matri- ces for head entities, relations, and tail entities, respectively. The attention weight measures the association of a relation rn to a head entity hn and a tail entity tn . Essentially, a graph vector gi is a weighted sum of the head and tail vectors [hn; tn] of the triples contained in the graph. 3.4 Knowledge Aware Generator Not_A_Fact Triple Vector Word Vector Key Entity Neighboring Entity Attended Entity Not_A_Fact Triple Attended Graph Previously Selected Triple Vector s t-1 s t s t+1 … … a Knowledge Aware Generator lack Knowledge Graph of Knowledge Graph lack of uv Knowledge Graph Knowledge Aware Generator Knowledge Aware Generator … Figure 4: Knowledge aware generator dynamically attends on the graphs and for final w attends on to compute is defined a where Vb/ of choosing vector cg t weight me st and a gr The mo K(gi) = { late the pro formally as Proceedings of the Twenty-Seventh International Joint Conference
44. DM Learning (Knowledge) • ੿݈۽ ޙઁо হח Ӓܿੌө? • Alexa prize 2017 winner / Alexa prize 2018 winner • Knowledgeܳ न҃ݎ੸ਵ۽ ׮ܖח Ѫ਷ ઱ചੑ݃ী ࡅ૑ח Ѫੌ ࣻب ੓਺ • ই૒ө૑੄ न҃ݎ ҳઑח ੋр੄ ߓ҃૑धਸ ׸ӝীח ցޖաب ੘਺ • न҃ݎ੸ ҳഅҗ ݽٕ੸ ҳഅ ࢎ੉ীࢲ੄ Ӑഋ ੟ח Ѫ੉ ೙ਃೣ *OUSPEVDUJPO 3FMBUFE
    8PSL $PODMVTJPO &YQFSJNFOUT .PEFM 4ZTUFN
    "SDIJUFDUVSF