時間情報表現抽出とルールベース解析器のこれから / Temporal Expression Analysis in Japanese and Future of Rule-based Approach

࣌ؒ৘ใදݱநग़ͱϧʔϧϕʔεղੳثͷ͜Ε͔Β
2022/04/08

Ubieגࣜձࣾ @yag_ays

View Slide

2
Ԟా ༟थ

@yag_ays

Recruit → Sansan → Ubie

ࣗݾ঺հ
https://yag-ays.github.io/

View Slide

3
ࠓ೔͓࿩͢Δ͜ͱ
• ࣌ؒ৘ใදݱΛղੳ͢Δja-timexʹ͍ͭͯ

• ࣌ؒ৘ใදݱͷநग़/ن֨Խͱ͍͏λεΫʹ͍ͭͯઆ໌͠·͢

• ࢲ͕࡞੒ͨ࣌ؒ͠৘ใදݱΛղੳ͢ΔϥΠϒϥϦΛ঺հ͠·͢

• ͲͷΑ͏ʹ࣮૷͍͔ͯͬͨ͠ͷഎܠ΍ۤ࿑΋؆୯ʹ঺հ͠·͢

• ϧʔϧϕʔεղੳثͱ࣌ؒ৘ใදݱͷ͜Ε͔Β

• ϧʔϧϕʔεͬͯࠓͲ͖Ͳ͏ͳͷʁେن໛ݴޠϞσϧʹஔ͖׵Θ͍ͬͯ͘ͷʁͱ͍͏ٙ໰Λɹ
ߟ͑ͯΈ·͢

• طଘݚڀͳͲ΋౿·͑ͯɺ࣌ؒ৘ใදݱͷࠓޙʹ͍ͭͯߟ͑ͯΈ·͢

View Slide

4
ja-timexͱ͸
• ࣗવݴޠͰॻ͔Εͨ࣌ؒ৘ใදݱΛநग़/ن֨Խ͢Δϧʔϧϕʔεͷղੳث

• Pythonύοέʔδͱͯ͠pipͰΠϯετʔϧͰ͖ΔܗͰPyPIͰެ։

• ࣌ؒ৘ใදݱͱ͸ʁ

• ਓ͕ؒݴ༿ͱͯ͠දݱ͢Δ೔෇΍࣌ࠁɺظؒɺස౓ͱ͍ͬͨ࣌ؒʹؔ࿈͢Δ֓೦ͷ͜ͱ
https://github.com/yagays/ja-timex/

https://ja-timex.github.io/docs/

View Slide

5
൴͸2008೥4݄͔Βिʹ3ճͷδϣΪϯάΛ

ே8͔࣌Β1࣌ؒߦ͖ͬͯͨ
࣌ؒ৘ใදݱͰλά෇͚͢ΔTIMEX3ܗࣜ [1]

View Slide

6
2008೥4݄ ே8࣌ 1࣌ؒ िʹ3ճ
நग़


View Slide

7

2008೥4݄ ே8࣌ 1࣌ؒ िʹ3ճ
"2008-04-XX" "T08-XX-XX" "PT1H" "P1W", "3X"
DATE

೔෇දݱ
TIME

࣌ؒදݱ
DURATION

࣋ଓ࣌ؒදݱ
SET

ස౓ू߹දݱ
ن֨Խ
நग़

View Slide

8
• ࠷ॳͷ͖͔͚ͬ (2021/07ࠒ)

• NAISTߥ຀ݚओ࠵ͷNTCIR16 Real-MedNLPͱ͍͏γΣΞʔυλεΫ͕ެ։͞Εͨ

• ຊλεΫͷݻ༗දݱநग़ͷΤϯςΟςΟͷ1ͭʹ࣌ؒ৘ใදݱ͕͋ͬͨ

• Real-MedNLPͰͷར༻΋ߟ͑ɺ͍ͣΕࣗࣾͷϓϩμΫτ։ൃ౳ͰඞཁʹͳΔͩΖ͏ͳͱࢥͬͯ
நग़ΤϯδϯΛ࡞Γ࢝ΊͨΒɺ൚༻తͳ΋ͷʹͨ͘͠ͳͬͨ (खஈͷ໨తԽ)

• ҩྍυϝΠϯʹ͓͚Δ࣌ؒ৘ใදݱͷॏཁ͞

• ిࢠΧϧςจॻɺ࿦จͷ঱ྫใࠂͳͲͰ͸ɺ࣌ؒ৘ใදݱͱࣄ৅͕ϖΞͰهࡌ͞ΕΔ

• ױऀ͕ૌ͑Δ঱ঢ়ͷൃੜ࣌ظ΍͔ͦ͜Βͷܦա (׮ղ/૿ѱ) ͸ɺ਍࡯ʹ͓͍ͯॏཁ

• Temporal Relation Extraction, Temporal reasoningͳͲͱݺ͹ΕΔλεΫ [2] [3]
ja-timex࡞੒ͷϞνϕʔγϣϯ - ཧ೦ฤ

View Slide

9
ҩྍυϝΠϯͷจॻʹ͓͍ͯ࣌ؒ৘ใදݱ͸Ͳ͏දΕΔ͔
22ࡀͷஉੑɻΉ͘ΈΛओૌʹདྷӃͨ͠ɻ2 िલɺٸʹإ໘ͱԼḰͷΉ͘Έٴͼഉ೘ޙͷ
೘ͷࡉ͔͍๐ཱͪʹؾ෇͍ͨɻΉ͘Έ͸࣍ୈʹ૿ѱ͠ɺ͜ͷؒʹମॏ͸໿ 20 kg૿Ճ͠
ͨɻ5 ೔લ͔Β৯ཉ͕ͳ͘৯ࣄྔ͸൒ݮ͠ɺԼཀྵؾຯͰɺશ਎᷺ଵײ͕ѱԽ͍ͯ͠Δɻ
Ұࡢ೔͔Βഉ೘ճ਺͕ݮগ͠ɺࡢ೔͸৭͕ೱ͘๐ཱͪͷڧ͍೘͕ 2 ճɺ͍ͣΕ΋গྔग़
ͨͷΈͰɺຊ೔͸ىচޙ 10 ࣌ؒͰ·ͩഉ೘͕ͳ͍ɻ

ҩࢣࠃՈࢼݧ ୈ115ճ A໰୊ 26໰໨ΑΓ

View Slide

10
ҩྍυϝΠϯͷจॻʹ͓͍ͯ࣌ؒ৘ใදݱ͸Ͳ͏දΕΔ͔
22ࡀͷஉੑɻΉ͘ΈΛओૌʹདྷӃͨ͠ɻ2 िલɺٸʹإ໘ͱԼḰͷΉ͘Έٴͼഉ೘ޙͷ
೘ͷࡉ͔͍๐ཱͪʹؾ෇͍ͨɻΉ͘Έ͸࣍ୈʹ૿ѱ͠ɺ͜ͷؒʹମॏ͸໿ 20 kg૿Ճ͠
ͨɻ5 ೔લ͔Β৯ཉ͕ͳ͘৯ࣄྔ͸൒ݮ͠ɺԼཀྵؾຯͰɺશ਎᷺ଵײ͕ѱԽ͍ͯ͠Δɻ
Ұࡢ೔͔Βഉ೘ճ਺͕ݮগ͠ɺࡢ೔͸৭͕ೱ͘๐ཱͪͷڧ͍೘͕ 2 ճɺ͍ͣΕ΋গྔग़
ͨͷΈͰɺຊ೔͸ىচޙ 10 ࣌ؒͰ·ͩഉ೘͕ͳ͍ɻ

ҩࢣࠃՈࢼݧ ୈ115ճ A໰୊ 26໰໨ΑΓ

2िલ 5೔લ ࠓ
ࡢ೔ 10࣌ؒ
࣌ܥྻ
Ұࡢ೔

View Slide

11
• ༰қʹར༻Ͱ͖Δ΋ͷ͕౰࣌ݶΒΕ͍ͯͨ

• ϧʔϧϕʔε࣮૷͸੒ᖒΒ[4] ͷ࿦จ࣮૷Ͱ͋ΔnormalizeNumexp ͷΈ

• C++࣮૷ͰɺPythonόΠϯσΟϯά͸͋Δ΋ͷͷखݩͰಈ͔͢͜ͱ͕Ͱ͖ͳ͔ͬͨ

• ͳ͓ɺݱࡏͰ͸ pynormalizenumexp ͱ͍͏PythonͰͷ࠶࣮૷͕ଘࡏ͢Δ

• spaCyͷݻ༗දݱநग़Ϟσϧ͸༰қʹར༻Ͱ͖ͨ

• DATEͱTIMEʹରԠ͍͕ͯͨ͠ɺ೔෇දݱʹऑ͍ͳͲਫ਼౓ͷ՝୊Λײͨ͡

• ࣌ؒ৘ใදݱΛநग़͢ΔͷΈͰɺ͔ͦ͜ΒPythonͷdatetime౳΁ͷม׵͸Ͱ͖ͳ͍

• σʔληοτ͕ͳ͔ͬͨ

• TIMEX3ܗࣜ౳Ͱ࣌ؒ৘ใදݱ͕Ξϊςʔγϣϯ͞Ε͍ͯΔσʔληοτ͕ར༻Ͱ͖ͳ͔ͬͨ

• CRF/ܥྻϞσϧ/BERT౳Ͱɺؾܰʹڭࢣ͋ΓֶशͷNERΛߏஙͰ͖ͳ͔ͬͨ
ja-timex࡞੒ͷϞνϕʔγϣϯ - ࣮૷ฤ

View Slide

12
ja-timexͷ࣮૷

View Slide

13
ja-timexͷ࢖͍ํ

View Slide

14
ͲͷΑ͏ʹϧʔϧΛهड़͍ͯ͠Δ͔ʁ
1. ࣌ؒ৘ใදݱͷ਺஋෦෼ΛάϧʔϓԽ

• ਖ਼نදݱͰͷ਺஋දݱͷऔಘͱՄಡੑ
୲อͷͨΊʹɺ໊લ෇͖άϧʔϓΛ
࡞͓ͬͯ͘

2. ࣌ؒ৘ใදݱͷύλʔϯΛਖ਼نදݱͰߏங

• จࣈྻ্Ͱදݱ͞Ε͏Δ࣌ؒ৘ใදݱ
Λਖ਼نදݱʹམͱ͠ࠐΉ

• ରԠ͢Δղੳ༻ͷؔ਺ͱඥ෇͚Δ

3. நग़͞Εͨ࣌ؒ৘ใදݱΛղੳ

• நग़ͨ࣌ؒ͠৘ใදݱͷจࣈྻΛ
TIMEX3λάͷ࢓༷ʹ߹Θͤͯղੳ͠ɹ
ग़ྗ͢Δ

View Slide

15

࡞͓ͬͯ͘





ग़ྗ͢Δ

View Slide

16

࡞͓ͬͯ͘





ग़ྗ͢Δ

View Slide

17
࣌ؒ৘ใදݱΛѻ͏্Ͱͷͦͷଞͷػೳ
• ΞϥϏΞ਺ࣈ/׽਺ࣈɺ੢ྐྵ/࿨ྐྵͳͲͷଟ࠼ͳϑΥʔϚοτʹରԠ

• ׽਺ࣈˠΞϥϏΞ਺ࣈ΁ͷม׵͢Δ͜ͱͰ࣮ݱ

• ࿨ྐྵͷදݱ΋ରԠ

• Pythonͷ೔෇ܕ/࣌ؒܕ΁ͷม׵

• datetime΍timedeltaܗࣜʹม׵͠ɺϓϩάϥϜ͔Βར༻Ͱ͖ΔΑ͏ʹ͢Δ

• ج४࣌Λઃఆ͠ɺ૬ରతͳ࣌ؒ৘ใදݱͰͷิ׬͕Մೳ

• ࠓ೔Λج४೔ͱͨ͠ͱ͖ɺʮ4/8ʯˠ ʮ2022-04-08ʯͷΑ͏ʹ೥͕ิ׬͞ΕΔ

View Slide

18
։ൃͷొΓํ (1)
• طଘݚڀΛಡΈࠐΉ

• ͱʹ͔͘࢓༷͕ෳࡶͰཧղ͕೉͍͠

• ͜ΕΛҰ͔ΒखಈΞϊςʔγϣϯ͠ΖͱݴΘΕΔͱɺ͔ͳΓ೉͍͠Μ͡Όͳ͍͔ͱࢥ͏

• ແݶʹٙ໰͕ग़ͯ͘ΔͷͰɺͳΜͱ͔໌จԽͰ͖Δܗʹམͱ͠ࠐΜͰ͍͘

• e.g. DATEͱTIMEͬͯԿ͕ҧ͏ͷʁ → 1೔ͱ͍͏ɺ஍ٿ্ͷपظతͳ࠷খ୯ҐͰ۠ผ

• e.g.ʮே8࣌ʯͷʮேʯ͸ؚΊΔ΂͖ʁ → طଘݚڀͩͱؚΊ͍ͯΔͷͰϤγʂ

• ·ͣ͸ߟ͑ΒΕΔදݱΛ۪௚ʹ࣮૷

• ςετۦಈ։ൃͰɺ࣌ؒ৘ใදݱͱ͋Δ΂͖ղੳ݁ՌΛେྔʹྻڍ͢Δ

• ͦͷςετʹ௨ΔΑ͏ʹͻͨ͢Βຊମͷ࣮૷ΛਐΊΔ

View Slide

19
։ൃͷొΓํ (2)
• Livedoor χϡʔείʔύεશจʹର࣮ͯ͠ߦͯ͠False
PositiveΛ௵͍ͯ͘͠ɺΛ܁Γฦ͢

• ͜ΕͰ͸False Negative͸վળ͕೉͍͕͠ɺϧʔϧ
ϕʔεͷ࣮૷Ͱ͸͋·Γى͜Βͳ͍ͱ൑அ

• ੨ۭจݿͰ׽਺ࣈͷදݱΛ୳ͯ͠ɺςετʹ௥Ճ͍ͯ͘͠

• ׽਺ࣈˠΞϥϏΞ਺ࣈม׵ࣗମͷ࣮૷΋ฒߦ

• ༧૝͠ͳ͍਺ࣈදݱ͕େྔʹݟ͔ͭΓࠔ࿭

• ׽਺ࣈ͔ΒΞϥϏΞ਺ࣈͷม׵ʹࠔΔྫ

• ʮҰɺʓʓ࢛ઍࣣޒʓʯ

• ʮࡾೋʓઍʓʓʓʯ
ʮޒϱ೥ܭըͱιϰΣτಉໍͷจԽతඈ༂ʯ ٶຊඦ߹ࢠ ΑΓ

View Slide

20
ja-timexͷݶքͱ՝୊
• จ຺Λߟྀͨ͠೔෇දݱͷநग़͕Ͱ͖ͳ͍

• False Positive: ʮੴͷ্ʹ΋ࡾ೥ʯʮҙؾࠐΈ͸े෼Ͱ͢ʯʮ12/5 ͸ 2.4Ͱ͢ʯ

• False Negative: DD/MM/YYYYͳͲͷඇ೔ຊޠݍͷ೔෇දه

• TIMEX3࢓্༷ͷݶք

• ͋Δಛఆͷ1೔Λද͍ͯ͠Δͷʹ΋ؔΘΒͣɺෳ਺ʹ෼ׂ͞Εͯ͠·͏

• ʮ4݄8೔༵ۚ೔ʯˠ

• ᐆດ͕͋͞Δ೔෇දݱΛදݱ͖͠Εͳ͍

• ʮઌ݄ͷ೔༵೔ʯʮ4݄7,8೔ʯ

• ਫ਼౓ΛͪΌΜͱܭଌͰ͖͍ͯͳ͍

• ධՁσʔλΛ࡞Δͷ͕໘౗ष͗͢Δͱ͍͏ͷͰɺ͜Ε͸ࢲͷଵຫͰ͢͝ΊΜͳ͍͞

View Slide

21
ϧʔϧϕʔεղੳثͱ࣌ؒ৘ใදݱநग़ͷ͜Ε͔Β

View Slide

22
࣌ؒ৘ใදݱͷநग़͸ࠓޙͲ͏ͳ͍ͬͯ͘ͷ͔ʁ
• લड़ͷ௨Γɺจ຺Λߟྀ͠ͳ͍ͱؒҧ͑Δྫ͕͋Δͷ͸͔֬

• ԿΒ͔จ຺Λߟྀͨࣗ͠વݴޠॲཧͷΞϓϩʔν͕ඞཁ

• ৭ʑͳߟ͑ํ͕͋Γ͏Δ

• ݫີͳऔಘʹ͸ਖ਼نදݱͰͷܾఆతͳॲཧ͕ඞਢͩΑ೿

• ϧʔϧʹج͍࣮ͮͯ֬ʹಈ͘ͱ͍͏҆৺ײɺϛεΛଈ࠲ʹमਖ਼Ͱ͖ΔରԠྗ

• TransformerͳͲͷDNNʹΑΔݻ༗දݱநग़͕༏੎ʹͳΔΑ೿

• จ຺Λߟྀͨ͠ݻ༗දݱநग़ͱͯ͠ɺநग़ثͷ෦෼Λ୅ସ͢Δ

• ͦ΋ͦ΋End2EndͳγεςϜͰ͸࣌ؒ৘ใදݱΛऔಘ͢ΔͳΜͯ͜ͱ͸͠ͳ͍Α೿

• μΠϨΫτʹܭࢉػ͕ղऍՄೳͳ࣌ؒΛग़ྗ͢ΔɺQuestion Answeringతʹղ͘ɺͳͲ

View Slide

23
࣌ؒ৘ใදݱͷۙ೥ͷݚڀ
• BERTΛ༻͍ͨݻ༗දݱநग़Ξϓϩʔν (2019) [5]

• σʔληοτͷछྨʹΑͬͯɺBERTͱϧʔϧϕʔεͷ༏ྼ͕มΘΔ݁Ռ

• seq2seqλεΫͱͯ͠ɺTIMEX3λάΛؚΉςΩετΛ௚઀ੜ੒͢ΔΞϓϩʔν (2021) [6]

• ϧʔϧϕʔεͷ݁ՌΛWeak SupervisionʹΑΓFine-tuningͷσʔλʹར༻

• BERT౳ͷNER (Token Classification) ͱൺֱͯ͠ɺଟ͘ͷσʔληοτͰߴਫ਼౓

• Huggingfaceʹͯར༻ՄೳͳϞσϧ͕഑෍͞Ε͍ͯΔ

• https://github.com/satya77/Transformer_Temporal_Tagger

View Slide

24
ͦ΋ͦ΋࣌ؒ৘ใදݱ͸ࣗવݴޠॲཧʹͱͬͯԿͳͷ͔ʁ
• ਺஋తͳଆ໘

• Ճࢉݮࢉͱ͍ͬͨૢ࡞͕Մೳͳɺ఺΍ͦͷൣғͱͯ͠ͷ਺஋දݱ

• 1࣌ؒ + 30෼ = 1.5࣌ؒ = 90෼

• ࡢ೔ͷ໌ޙ೔ = ໌ޙ೔ͷࡢ೔ = ໌೔

• ݴޠతͳଆ໘

• ਓؒੜ׆Λओ࣠ͱͨ͠ײ֮తදݱ

• ேɾனɾ൩ɿओ؍తͳ1೔ͷ۠෼ͷදݱ

• ઌ೔ɿ͋ΔҰఆظؒ಺ͷաڈͷ1೔Λࢦ͢දݱ

• 25࣌ɿཌ1࣌ΛಉҰͷ೔ͷԆ௕ઢ্Ͱ͋Δଊ͑ͨͱ͖ͷදݱ

• Ұສ೥ͱೋઍ೥લɿ͸Δ͔ੲͰ͋Δ͜ͱΛද͢ތுදݱ

View Slide

25
࣌ؒ৘ใදݱ͸ ৗࣝ Commonsense ͱߟ͑ΒΕΔ
• Numerical Commonsense [7]

• ࣌ؒ৘ใදݱ͸ɺ਺஋ʹؔ͢ΔৗࣝͷҰ෦

• ۩ମతͳλεΫ

• ؚҙؔ܎ೝࣝ (Recognizing Textual Entailment: RTE)

• ৗࣝਪ࿦ (Commonsense Reasoning)

• ϚεΫ͞Εͨ୯ޠͷ༧ଌ (Masked-word-prediction)

• σʔληοτ

• NumerSense [8]

• Ұ෦ʹӈਤͷΑ͏ͳ࣌ؒʹؔ͢Δ࣭໰͕͋Δ ※ ja-timexͰ͸೥ྸ͸࣌ؒ৘ใදݱͱͯ͠
͸ѻ͍ͬͯͳ͍ͷͰɺݫີʹ͸͜Ε͸ɹ
࣌ؒ৘ใදݱʹؔ͢ΔλεΫͰ͸ͳ͍
https://inklab.usc.edu/NumerSense/

View Slide

26
ݱࡏͷେن໛ݴޠϞσϧʹΑΔจੜ੒͸࣌ؒΛͲ͏ଊ͍͑ͯΔ͔ʁ
• Tnrasformerϕʔεͷେن໛ݴޠϞσϧʹΑͬͯɺੜ੒͞ΕΔจॻͷࣗવ͞͸֨ஈʹ্͕ͬͨ

• ҰํͰɺจͱͯ͠ͷهड़಺༰ͷҰ؏ੑ΍ৗࣝͷѻ͍͸·ͩ·ͩ

• ࢼ͠ʹGPT-3Ͱ࣌ؒʹؔ͢ΔςΩετΛੜ੒ͯ͠ΈΔ

• ʮ9࣌ͷ1࣌ؒޙ͸10࣌Ͱ͢ɻ9࣌ͷ2࣌ؒޙ͸ʯʹଓ͘จΛੜ੒ͤ͞Δ

• Ϟσϧ͸rinna/japanese-gpt-1bɺύϥϝʔλ͸do_sample=True, top_p=0.95
9࣌ͷ1࣌ؒޙ͸10࣌Ͱ͢ɻ9࣌ͷ2࣌ؒޙ͸1࣌Ͱ͢ɻ͜͜Ͱݴ͏1࣌ͱ͸9࣌00෼ͷ͜ͱΛࢦ͠·͢ɻͦͯ͠

9࣌ͷ1࣌ؒޙ͸10࣌Ͱ͢ɻ9࣌ͷ2࣌ؒޙ͸ɺ10͔࣌ͩΒɺԿ΋ແ͠Ͱྑ͍ɻԿ΋ແ͍͔Βग़དྷΔΜͰ͢ɻͦΕʹࠓɺ

9࣌ͷ1࣌ؒޙ͸10࣌Ͱ͢ɻ9࣌ͷ2࣌ؒޙ͸ɺ͓னͷ12࣌Ͱ͢ɻ2͔࣌Β6࣌·Ͱ͕ɺࢲֶ͕ͨͪߍʹ͍Δ࣌ؒͰ͢

9࣌ͷ1࣌ؒޙ͸10࣌Ͱ͢ɻ9࣌ͷ2࣌ؒޙ͸11࣌Ͱ͢ɻ18࣌ͷ1࣌ؒޙ͸20࣌Ͱ͢ɻ͜ΕΒͷ1೔ͷதʹ͸

9࣌ͷ1࣌ؒޙ͸10࣌Ͱ͢ɻ9࣌ͷ2࣌ؒޙ͸18࣌Ͱ͢ɻͳͷͰɺͦΜͳʹ͋ͬͱ͍͏ؒʹ͸ಈ͖·ͤΜɻ͔ͩΒ஗͔Ε

9࣌ͷ1࣌ؒޙ͸10࣌Ͱ͢ɻ9࣌ͷ2࣌ؒޙ͸10࣌ͷ5෼ޙͰ͢ɻͭ·Γ10࣌͸ʮ໿5࣌50෼ʯ

9࣌ͷ1࣌ؒޙ͸10࣌Ͱ͢ɻ9࣌ͷ2࣌ؒޙ͸11࣌ɺ12࣌͸͓னɻ25࣌24෼͸2࣌ؒޙɺͦͯ͠ɺ౔༵೔ͷ

9࣌ͷ1࣌ؒޙ͸10࣌Ͱ͢ɻ9࣌ͷ2࣌ؒޙ͸11࣌Ͱ͢ɻ9͔࣌Β11࣌·Ͱͷ2࣌ؒͷฏۉͷ࣌ܭͷ਑ͱͷ࣌ؒ

View Slide

27
͋ΒͨΊͯ࣌ؒ৘ใදݱநग़ͷ͜Ε͔Β
• ϧʔϧϕʔεղੳث͸ࠓޙ΋ར༻͞Εଓ͚ΔͩΖ͏

• ࢖͍΍͢͞΍ϝϯςφϯεͷ༰қ͞ͱ͍͏ར༻࣌ͷ؍఺͔Β

• Weak SupervisionͰͷֶशσʔλ࡞੒ͱ͍͏ɺΑΓߴਫ਼౓ͳख๏ʹܨ͕Δେ͖ͳ໾໨

• จ຺ΛߟྀͰ͖ΔTransformerϕʔεͷݻ༗දݱநग़͸ݱ࣮త

• ӳޠΛର৅ͱͨ͠ݚڀͰ͸ϧʔϧϕʔεͱൺֱͯ͠΋ߴਫ਼౓

• ja-timexͷ݁ՌΛ࢖ͬͨWeak SupervisionతΞϓϩʔνͰja-timexͷڧԽ൛Λ࡞ͬͯΈ͍ͨ

• ධՁ༻ͷ೔ຊޠσʔληοτ΋࡞ΒͶ͹……

• େن໛ݴޠϞσϧࣗମ͕࣌ؒ৘ใදݱΛਖ਼͘͠ѻ͑Δͷ͸·ͩ·ͩઌ͔΋

• ςΩετੜ੒΍ਪ࿦ϨϕϧͰ͖ͪΜͱͨ݁͠Ռ͕ฦͬͯ͘Δͷ͸ظ଴Ͱ͖ͳ͍ (?)

• End2Endతʹ࣌ؒ৘ใΛ͍͍ײ͡ʹѻͬͯ͘ΕΔͷ͸ɺ·ͩ·ͩઌ͔΋

View Slide

28
·ͱΊ
• ja-timexͷ঺հ

• ࣌ؒ৘ใදݱΛ؆୯ʹநग़/ن֨ԽͰ͖Δϧʔϧϕʔεͷղੳث

• େྔͷਖ਼نදݱΛهड़͢Δ͜ͱͰ࣮ݱ͍ͯ͠Δ

• จ຺Λߟྀͨ࣌ؒ͠৘ใදݱͷநग़ʹ͸ɺϧʔϧϕʔεͰ͸ݶք͕͋Δ

• ϧʔϧϕʔεղੳثͷ͜Ε͔Β

• େن໛ݴޠϞσϧʹΑΔվྑ͕ൃද͞Ε࢝Ίͨ

• ϧʔϧϕʔεղੳثͷ݁ՌΛֶशσʔλʹར༻͢ΔྲྀΕ

• ࠓޙTransformerϕʔεͷݻ༗දݱநग़ثʹஔ͖׵Θ͍͖ͬͯͦ͏

• ࣌ؒ৘ใදݱ͸ɺৗࣝΛѻ͏λεΫͷҰ෦ͱଊ͑Δ͜ͱ͕Ͱ͖Δ

• େن໛ݴޠϞσϧ͕࣌ؒ৘ใදݱΛݩʹਖ਼͘͠ਪ࿦͢Δͷ͸·ͩ·ͩ೉ͦ͠͏

View Slide

29
[1] খ੢ޫ, ઙݪਖ਼޾, & લ઒تٱ༤. (2013). ʰݱ୅೔ຊޠॻ͖ݴ༿ۉߧίʔύεʱ ʹର͢Δ࣌ؒ৘ใΞϊςʔγϣϯ. ࣗવݴޠॲཧ, 20(2),
201-221.

[2] Sun, W., Rumshisky, A., & Uzuner, O. (2013). Temporal reasoning over clinical text: the state of the art. Journal of the American
Medical Informatics Association, 20(5), 814-819.

[3] Alfattni, G., Peek, N., & Nenadic, G. (2020). Extraction of temporal relations from clinical free text: A systematic review of current
approaches. Journal of Biomedical Informatics, 108, 103488.

[4] ੒ᖒࠀຑ (2014)ʮࣗવݴޠॲཧʹ͓͚Δ਺ྔදݱͷऔΓѻ͍ʯ౦๺େֶେֶӃ म࢜࿦จ

[5] Chen, S., Wang, G., & Karlsson, B. (2019). Exploring word representations on time expression recognition. Technical report,
Microsoft Research Asia.

[6] Almasian, S., Aumiller, D., & Gertz, M. (2021). BERT got a Date: Introducing Transformers to Temporal Tagging. arXiv preprint
arXiv:2109.14927.

[7] Narisawa, K., Watanabe, Y., Mizuno, J., Okazaki, N., & Inui, K. (2013, August). Is a 204 cm man tall or small? acquisition of
numerical common sense from the web. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics
(Volume 1: Long Papers) (pp. 382-391).

[8] Lin, B. Y., Lee, S., Khanna, R., & Ren, X. (2020). Birds have four legs?! numersense: Probing numerical commonsense knowledge of
pre-trained language models. Proceedings of EMNLP
Reference

View Slide

時間情報表現抽出とルールベース解析器のこれから / Temporal Expression Analysis in Japanese and Future of Rule-based Approach

時間情報表現抽出とルールベース解析器のこれから / Temporal Expression Analysis in Japanese and Future of Rule-based Approach

yag_ays

More Decks by yag_ays

Other Decks in Research

Featured

Transcript