PyCon Korea 2019 키노트 발표 자료입니다.
ॆҗ ழޭפ౭৬ ೠҴয য়ؘఠ2019-08-17߅
View Slide
߅ a.k.a. lucypark, echojuliett, e9t• ӝ҅ߣਸ ೞח ݠन۞ ূפয• ؘఠ۽ ध ߷ਸ ծ୶Ҋ र ࢎۈ• Yak shaver2
ٸח 2019֙ 5ਘ• ӝഥо ػݶ ೞҊ र ݈ য!• "ઁউ೧࣊ࢲ хࢎפ! যڃ ফӝܳ ೧ঠೡ ഛन ٜח ঋ݅ ਬೠ ফӝܳ ೡ ࣻ ب۾ ળ࠺೧ࠅѱਃ."• ߊө 3ѐਘա ਵפ ળ࠺ೡ दрب ࠙ೞ!3
ԝԝೠ ҅ദ• ೠҴয NLP য়ࣗझ/ؘఠী ೠ ࢲ߬ܳ ೧ࢲ, അടਸ ҕਬೞҊ э ߊदெաоҊ ઁউ೧ঠѷয!• ః֢ۄݶ ࠁܳ ׳ೞӝࠁח ನਊinclusion, নࢿdiversity, ழޭפ౭৬э оܳ ࠗпೞח ߊݶ જਸ Ѫ эই• Ӓؘ۠ ژ פө ॆী ೠ ফӝب ೧ঠ• Ӓۧݶ 3֙݅ী konlpyب ખ সؘ ೧ࠅө? пઙ ؘఠࣇਸkonlpy.downloadী ನೣ೧ࢲ ࢎۈٜ औѱ ۽٘ ߉ѱ ೞݶ જਸ Ѫэ!4
ӒܻҊ 2019֙ 8ਘ• konlpyח Ҋசѱ ցޖ ݆... э ٜ݅যоਃ!• ই ߊܐب হ֎... ƕƕ• ೯ ܲ ইஜ ߊפө ࢎۈ ߹۽ উৢѢঠ...5
য়ט ೡ ঠӝ• ޙઁ ध ҕਬ• ழޭפ౭ী ѱա݃ ӝৈೡ ࣻ ؍ ѐੋੋ ҃ ҕਬ ঢ়զ ফӝ• ೠҴয য়ؘఠ ঠӝب ઑӘ ߸ ࢎۈٜ ݔզ ٜ؍ ফӝ*NBHFTPVSDF4BOESBBOE8PP 3JDIBSETHVJEFUPTPGUXBSFEFWFMPQNFOU 6
ৈ۞࠙ ۽Ӓې߁ਸ ৵ ೞदաਃ?7
۽Ӓې߁ ৵ ೞחо?• ஂ• ݡҊࢎפ્• ѐੋ ࢿ• ա ழޭפ౭ী ӝৈೞӝ೦ݾٜ࢚ഐߓఋঋ 8
ঢ়զ ঠӝ ೞա: ನಯॆਸ જইೠ द࠻ ೧ழٜ9
ٸח 2011֙10
11
12
ਫ਼Ӭ, ܐо ࠗೞҊਃ?13
ৈӝী חؘਃ? http://likms.assembly.go.kr14
ೞ݅ ցޖ য۵Ҋ হ ־о ־ҵب ݽܰѷҊ֙दചݶੑפӘউࠁदझమખ؊ಞܻೞѱ߄Շযਃ!15
ޙࢲח Ӕࢿ ڄয ֤ޙ ו՝ ޛঁ աח PDF16
Ѩ࢝ূ Ӕೡ ࣻ হח ଼ ೞ݅ Ҵഥ ܐח ҕҕ. Ҵ Ѩ࢝ೡ ࣻ যঠ!17
ܻо о ӝࣿਸ ਊ೧ࢲ ࣁ࢚ਸ ߄Լࠁ!18
ನಯ http://popong.com• ࠺ ٣ց, ѐߊ, ӝദ ١ਵ۽ ҳࢿػ ࢤ 10ৈ ݺ• ӝࣿ۽ॄ ೠҴ ܳ ٍൔ٘חѱ ݾ• ࢲ࠺झܳ ٜ݅যࠁݶࢲ ܻо ࢿೞח Ѫ ؒ! (ژח ؊ ݾ)1010/(1VCMJD0QFO10MJUJDBMFOHJOFFSJ/( 19
ࢲ࠺झ ۽షఋೝਸ ೞҊद೮؍۽షఋೝో#BMTBNJR!۽షఋೝҗҾӘೞݶৈӝ۽оࣁਃ 20
ਗ ֎ਕ ࠙ࢳب ೞҊ3FBENPSFBCPVUUIJTIFSF 21
ݒ షਃੌ݃ ݽৈ ӝദҗ షۿҗ झఠ٣ࣻо ੑߨҗ ೧ ইחѱ ೞաب হѢٚਃӒۧѱ 2ߣ ܻझ݃झܳ ೣԋ ݏೠ റ...22
ਢࢲ࠺झ "ೠҴ ݽٚ Ѫ" ࢤ Ѫ ٸޙী ܀݂ب ৌब ೞҊইऔѱبӘബਊ࠺࠺ਊ݆ػҊ౸ױ೧ࢲ֙ਘࠗఠPVUPGTFSWJDF ೞ݅ನಯӝоदबತࣗࢤۄदࣁ࢚ীաৢࣻبযਃ 23
ਢࢲ࠺झ "ೠҴ ݽٚ Ѫ" ࢤ ۿূ٘ب ߓҊ24
ਢࢲ࠺झ "ೠҴ ݽٚ Ѫ" ࢤ D3৬ NLPীب ҙबਸ оѱ غҊ25
ਢࢲ࠺झ "ೠҴ ݽٚ Ѫ" ࢤ PDF यب Ҋ26
South Korea/Seoul Maps ؘఠ ҕѐ ݫܰషܰ э ైߨب ߓshpੌਸTopoJSONਵ۽߄Բחؘ݅֙աѦ۷যਃ!27
Ҵഥਗ/উ ؘఠ REST API• ࢎप ਢࢲ࠺झח ࣳாझ• ѱ ਗې ೯ೞ۰؍ ۽ં• ؊ ݆ ࢎۈٜ ؘఠܳ ࠙ࢳೞѢաഝਊ೧ࢲ ࢲ࠺झܳ ٜ݅ӝܳ ߄ۗইऔѱبӘѪبPVUPGTFSWJDFҊ٘݅թই݅ ઁחҴഥীࢲؘఠܳઁҕפഥ۾ উ ਗ 28
য়ࣗझ, য়ؘఠ: ݽٚ Ѫਸ ҕਬೞ۰ח ֢۱29
ೠҴ ܳ ٍൔٜѷҊ ࢤпೠ ಁӝ ֈח ٜ֙• ೠಞਵ۽ח ߸ ࢎഥ۽ࠗఠ ੋਸ ߉ӝب ೮݅• ࣁ࢚ Ӓۧѱ औѱ ߸ೞחѱ ইפۄח Ѫਸ ߓਛҊ ࢲߡ ࠺ਊ ࠺ऱחѪب ঌѱ غҊ• ޖࠁب ࣗೠ જ ҳܳ Ҋ30
Ӓ যڃ ҃ࠁ ч જ ѐߊ ޙചܳ ߓਛणפ• ޙࢲച ೧ࣁਃ. ղ৻৬ ࣗాೞח ߑߨ• Git જ ۑ౭झח Ѣېਃ. ղ৻৬ ࣗాೡ ٸ ఃݶ જ ݒց• ࢜ ۽ંীࢲח Flaskۄח ۨਕܳ ॄࠅөਃ? ࢜۽ بҳܳ ח ߑߨ• ই٣যח working prototypeਵ۽ ࠁৈࣁਃ. ఌ࢚ҕۿਸ ߩযա ਸ ח ߑߨ• Ӓ ٘/ؘఠ ۄࣃझо ޤীਃ? ఋੋ ޛਸ ઓೞח ࣁ31
ࣁ࢚ਸ ߄Բח Ѫա৬ ղ ߸ਸ ߄Բח Ѫ32
दр ખ Ѧ۰ب ҡଳݽܰח Ѥ ߓݶ ػ33
য়ࣗझо ܳ о٠ؘఠܳ оҕ೧ࢲ য়ؘఠ۽ܾܻૉೞח Ѫب 34
ঢ়զ ঠӝ ل: KoNLPyਃо ࢤदఅ ॆ ۄ࠳۞ܻ35
ٸח 2014֙• ઁ ҕ ਗې ఫझی ࢚ҙহҊ ंܳ ࠙ࢳೞח ؘఠ݃• Ү ۽ંܳ ࣻ೯ೞ؍ ೠҴয ࠙ࢳਸ ಞܻೞѱ ೞӝ ਤ೧ द• ೠҴয ࠙ࢳਸ ೞ۰ݶ ੌױ షաਸ ೧ঠೞחؘ যڌѱ ೞ?• ഋకࣗ ࠙ࢳਸ ೞݶ ػחؘ য়ࣗझח ա?• ݻ оо ח Ѫ эؘ ࢿמ যڌѱ ܰ?36
Enter KoNLPy>>> from konlpy.tag import Okt>>> okt = Okt()>>> okt.pos('݅աࢲ ߈оਕਃ!')[('݅աࢲ', 'Verb'), ('߈оਕਃ', 'Adjective'), ('!', 'Punctuation')]• пઙ য়ࣗझ ഋకࣗ ࠙ࢳӝܳ ݽ• nltk ١ ܲ ۄ࠳۞ܻ৬ э ಞܻೞѱ ॶ ࣻ ѱ ੋఠಕझܳ ాੌ• ೠӖਸ ܙ ٸ ਃೠ пઙ utilਸ ୶оೣ37
KoNLPy৬ ߣ૩ PyCon KR• द ॆ झथ: " ೠҴীࢲ ৌܽ!"• ա: "ա৬ ࠺तೠ ࠛಞೣਸ ѻҊ ח ࢎۈ ݆ ঋਸө?"• ٜ݅؍ Ѧ ੜ ಁః೧ࢲ ҕѐ೧ࠁ!38
39
ࢎۈٜ ৵ KoNLPyܳ ࢎਊ೮ਸө?• ୡࠁ: ࢎਊߨ एਕࢲ? ؊ एݶ જѷ֎ਃ• ࢤ: ࠁҊ ٮۄೡ दо যࢲ?• पޖ: নೠ ҳഅ р ࢿמ ࠺Үо ಞ೧ࢲ?• ৻Ҵੋ: ޙࢲо য۽ب ॳৈ যࢲ?• ജ҃ ਃੋ: റ ॆ Ҵղ৻ীࢲ ೠହ ੋӝށܳ ೧ࢲ?40
ղо ਃೠ بҳחղо ٜ݅য ҕਬೠӒ۞ݶ ࢤпب ޅೠ بਸ Ҋ ߉ਸ ࣻ !41
ٞ೧ী فߣ૩ PyCon KR• ѐੋ زӝ: ҃ ցޖ જও• द ҅ীࢲח representation learning ೧ द• word2vec, doc2vecਸ Ҵղ ॆ ழޭפ౭ী ࣗѐ೧ࠅө?• ֙ী য়ࣗझ۽ KoNLPyܳ ҕѐ೮݅ ݄࢚ షؘఠо ߹۽ হ.ੌױ ೠҴয۽ ػ ؘఠࣇਸ ٜ݅যࠁ!42
ೠҴয ചಣ ؘఠ nsmc$ head ratings_train.txtid document label9976970 ই ؊ࡂ.. ૐա֎ਃ ݾܻࣗ 03819312 ൙...ನझఠࠁҊ ୡ٬ച....য়ߡোӝઑର оߺ ঋҳա 110265843 ցޖӒېࢲࠁחѪਸ୶ୌೠ 09045019 Үبࣗ ঠӝҳݢ ..ࣛ ח হ..ಣ ઑ 06483659 ࢎހಕӒ झ۠ োӝо ثࠁ؍ ച!झ؊ݔীࢲ יযࠁӝ݅ ೮؍ ழझ౯ ؍झо ցޖաب ࠁ 15403919 ݄ Ѧ݃ ڏ 3ࣁࠗఠ ୡ١Ү 1֙ࢤੋ 8ਊച.ƀƀƀ...߹߈ѐب ইө. 07797314 ਗ ӟхਸ ઁ۽ ۰ղޅ೮. 09443947 ߹ ߈ѐب ইӰ աৡ ҃ ӡਊ োӝࢤഝݻ֙ੋ..݈ ߊ۽೧ب ӒѪࠁױ իѶ ժ.хӘ݅߈ࠂ߈ࠂ.. 07156791 ঘ࣌ হחؘب ח ݻউغח ച 1• য ചಣਸ ӛ ژח ࠗਵ۽ ࠙ܨೞח ؘఠࣇ• Maas et al. 2011 IMDB ؘఠࣇ(য)ਸ ߮݃43
ղо о ӝࣿ ױೞ ঋইبழޭפ౭ী ӝৈೡ ࣻ ־ҳա ޖٚ!44
݄݃ਵ۽ਃ્ ೠҴয য়ؘఠ ঠӝ45
ٜযоӝী খࢲԙ ೞҊ र؍ ঠӝ46
ঌ߳ যо ইפ!Alphabet NLPо ইפۄ English NLP47
ೠӖ ೠҴযо ইפ!Hangul NLPо ইפۄ Korean NLPೠӖ ޙ, ೠҴযо য48
ೠҴয য়ؘఠ അട: Sequence classification / labelingҕѐदӝ ܴ ݾ ӝ ۄࣃझ2007 ࣁઙಌझ ಿࢎకӦ 838k sentences CC BY-NC-ND 4.02012 KOSAC хࢿ࠙ܨ(хࢿয ࢎب ߓನ)7.7k sentences Custom2015 nsmc хࢿ ࠙ܨ 200k sentences(Train: 150k, Test: 50k)CC0(Public Domain)2016 KoreanNERCorpus NER 3.5k sentences -2018 nlp-challenge NER 90k sentences -2018 nlp-challenge SRL 35k sentences -2018 Question pair Paraphrase detection 7k sentence pairs -ۄࣃझоউഃחۄࣃझоݺदغযঋѢաઁоޅ҃ੑפ 49
ೠҴয য়ؘఠ അട: Sequence generationҕѐदӝ ܴ ݾ ӝ ۄࣃझ2015 JPO patent corpus ӝ҅ߣja-ko, ౠೲ بݫੋ257ksentence pairsCustom(࢚স ਊ ࠛо)2017 Korean parallel corpora ӝ҅ߣko-{en, fr, ja}п যह߹0.7k, 95k, 0.2ksentence pairsCC BY-NC-ND 3.0(࢚স ࢎਊ ࠛо)2018 KSS ࢿ ࢿ/ੋध 12.9k pairs CC BY-NC-SA 4.0(࢚স ਊ ࠛо)2018 Chatbot data ച 12k pairs -2018 OpenSubtitles2018 ӝ҅ߣko-*1.3m sentence pairs(For ko-en only)-2019 AIHub ೠҴয-য ߣ ߽۳݈ޡӝ҅ߣko-en16k sentence pairs(To be 1.6m within 2019)-ۄࣃझоউഃחۄࣃझоݺदغযঋѢաઁоޅ҃ੑפ"*)VC҃ۄࣃझоݺदغযঋ݅ ࣗӝস ߮ӝস झఋস ѐੋѐߊ োҳ١ఋѶਬੋࠗࢎস۽CC0ژחCC BYঋਸө୶ஏ೧ࠇפژ ઁоӝ҅ߣীҙबযࢲܲ࠙ঠী࠺೧࢚ਵ۽؊աৌೡࣻחѪэؘ֬જؘఠࣇݶBXFTPNFLPSFBOOMQэܻझীҕਬ೧ࣁਃ 50
ೠҴয য়ؘఠ അട: Othersҕѐदӝ ܴ ݾ ӝ ۄࣃझ2018 KorQUAD MRC 66k Q/A pairs -2019? AIHub ӝ҅ة೧ MRC 450k Q/A pairs -2019 cc-kedict ೠ ࢎ 13.7k entries CC BY-SA 3.02019 kosentences Self-supervisedlearning31m sentences MIT+ GNU FreeDocumentation+ CC BY-NC-SAۄࣃझоউഃחۄࣃझоݺदغযঋѢաઁоޅ҃ੑפݻоחઁоਊ೧ࠁޅ೮णפ 51
ࢤпࠁ ݆!!؊ ݆ਵݶ જѷ݅...52
જ ࢎ۹ 1: KorQUAD https://korquad.github.io/• ࠙ೠ ন ؘఠ ҕѐ• बয ܻ؊ࠁ٘ө ҕਬ!53
જ ࢎ۹ 2: KSS https://kaggle.com/bryanpark/korean-single-speaker-speech-...• Ҵղ ୭ୡ ࢿ য়ؘఠ• ઁߨ ݆ ࠙• ۄࣃझо ڢ۶ೞѱ ݺदغয যࢲ ޖਸ ೡ ࣻ Ҋ হחо ݺഛೣ54
৵ য়ؘఠо ਃೠоਃ?1. ߮݃о ؼ ࣻ ӝ ٸޙ!• ML ݽ؛ٜ ؘఠ৬ о эইঠ ݽ؛ р ࢲ۽ ࠺Үо оמ• ࠺Үо оמ೧ݶ ӝࣿ ߊ ৡ2. ־ҳա ߄۽ ࠙ࢳա ݽ؛݂ਸ दೡ ࣻ ӝ ٸޙ!• ۽Ӓې߁ীࢲ reinventing the wheel ઙઙ ҃҅ ࢚ غ٠• ݽفо ؘఠ ஂٙҗ ઁܳ ೡ ਃח হਗ਼ইਃ55
ؘఠܳ ҕѐೡ ٸ ഛੋೞݶ જ 1. ࢎਊо ߄۽ ۽٘೧ࢲ ࢎਊೡ ࣻ חо?• оәݶ ഥਗоੑ, ਊز ର হ ߄۽ ࢎਊೡ ࣻ ѱ ೣ2. ਗޙী ѐੋࠁ/ӂ ޙઁח হחо?• য়ؘఠח ӝࣿ ߊਸ ਤ೧ ݒ ਃೞ݅ ѐੋࠁ৬ ӂبઓ߉Ҋ ெઉঠ ೣ!3. оәݶ ۄࣃझܳ ԙ ݺद೧ࣁਃ!• ਊ ߹ب ޙহ оמೠ Ѫ, ࠛоמೠ Ѫਸ ঌܻӝ ਤೣ56
য়ؘఠܳ ࢎਊೡ ٸ ഛੋೞݶ જ 1. ؘఠо ࠙ о? ࠙ ח Ѿҗܳ յ ࣻ ח ب۽ ࠙ೠо?2. ۄࣃझо ޖੋо? ߓನо оמೠо? ࢚সਊਵ۽ ਊ೧ب غחо?57
Call for partipation• ೠҴয NLP э ߊदெਃ!• য়ࣗझ۽ب, য়ؘఠ۽ب ӝৈೡ ࣻ णפ• য়ࣗझח ॆযب જ݅ ԙ ॆ ইפযب ؾפ• য়ؘఠח ࠗ, ӝস, Ү ݽف ೣԋ೧ਃ ޛۿ ѐੋ ѐߊب• ೠҴয NLP ݈Ҋ ܲ ߀ח ࠙ঠب ݆ ӝৈೞҊ ࢲ۽ ҕਬ೧ਃ!58
хࢎפ@echojulietthttps://lucypark.krठۄ٘ܳࠁҊӈೠೖ٘ߔਸन!MPWJU !TIVSBJOשԋхࢎ٘݀פ 59