完全なアノテーションが得られない状況下での固有表現抽出

Transcript

1. ׬શͳΞϊςʔγϣϯ͕

   ಘΒΕͳ͍ঢ়گԼͰͷݻ༗දݱநग़
   খྛ ᕣՏ @kajyuuen
   

   View Slide

2. ࣗݾ঺հ
   
   
   খྛ ᕣՏ (Koga Kobayashi)
   ID: @kajyuuen
   ஜ೾େֶେֶӃ म࢜1೥
   NERϒϩάΛॻ͍ͯ·͢
   

   View Slide

3. NERͷֶशʹ࢖͑Δσʔλ
   Ξϊςʔγϣϯίʔύε
   • NERλεΫͰҰൠతʹར༻͞ΕΔσʔληοτ
   • ୯ޠʹରͯ͠ɼ֤ʑରԠ͢Δϥϕϧ͕෇༩͞Ε͍ͯΔ
   σϝϦοτ
   • υϝΠϯຖʹ࡞Γ௚͢ඞཁ͕͋Δ
   • Ξϊςʔγϣϯʹߴ͍ίετ͕͔͔Δ
   Donald
   B-PER
   John
   I-PER
   Trump
   E-PER
   is
   O
   president
   O
   of
   O
   the
   O
   US
   S-LOC
   the
   O
   ΞϊςʔγϣϯίʔύεΛ࡞ΔίετΛݮΒ͢ or ແ͍ͨ͘͠
   
   
   

   View Slide

4. NERͷֶशʹ࢖͑Δσʔλ
   ෦෼తΞϊςʔγϣϯίʔύε
   • Ұ෦ͷ୯ޠʹ͚ͩΞϊςʔγϣϯ͕෇༩͞Ε͍ͯΔίʔύε
   • Ξϊςʔλ͸ࣗ৴ͷͳ͍୯ޠʹΞϊςʔγϣϯ͠ͳͯ͘ࡁΉ
   σϝϦοτ
   • ֶशʹগ͠޻෉͕ඞཁ(ҰൠతͳCRFͰ͸ֶशग़དྷͳ͍)
   
   
   Donald
   -
   John
   B-PER
   Trump
   E-PER
   is
   -
   president
   -
   of
   O
   the
   -
   US
   -
   the
   O
   ෦෼తΞϊςʔγϣϯίʔύε
   - Eraldo R Fernandes and Ulf Brefeld. 2011. Learning from partially annotated sequences. In Proceedings of ECML-KDD.

   - Andrew Carlson, Scott Gaffney, and Flavian Vasile. 2009. Learning a named entity tagger from gazetteers with the partial perceptron.
   In Proceedings of AAAI Spring Symposium: Learning by Reading and Learning to Read.

   - Jie Zhanming, Xie Pengjun, Lu Wei, Ding Ruixue and Li Linlin. 2019. Better Modeling of Incomplete Annotations for Named Entity
   Recognition. In Proceedings of NAACL.
   

   View Slide

5. NERͷֶशʹ࢖͑Δσʔλ
   
   
   ࣙॻ
   • ͍ΘΏΔࣙॻɼͨͩݴ༿ͷ಺༰ɼҙຯ·Ͱ࢖͏͜ͱ͸গͳ͍
   • ༷ʑͳ෼໺ʹஔ͍ͯݩ͔Βଘࡏ͍ͯ͠Δ͜ͱ͕ଟ͍
   • ਓ໊ࣙయ, ༀֶ༻ޠࣙయ, ෺࣭ɾࡐྉσʔλϕʔε
   ࣙॻϚονʹΑΔݻ༗දݱநग़Ͱ͸
   • ࣙॻʹଘࡏ͠ͳ͍୯ޠΛݕग़Ͱ͖ͳ͍
   • ݻ༗දݱʹؒҧͬͨϥϕϧ͕෇༩͞ΕΔ
   ͱ͍ͬͨ໰୊͕ੜ͡Δ
   Donald
   -
   John
   S-PER
   Trump
   -
   is
   -
   president
   -
   of
   -
   the
   -
   US
   -
   the
   -
   Person Dictionary
   John Michael
   ࣙॻϚον
   ࣙॻ + ੜίʔύεΛ༻͍ͨDistantly Supervisedͱ͍͏ख๏͕ग़ݱ
   

   View Slide

6. ࠓճͷ෼໺ʹؔ܎͋Δ࿦จ
   ෦෼తΞϊςʔγϣϯίʔύε
   • Better Modeling of Incomplete Annotation for Named Entity
   Recognition
   Distantly Supervised NER
   • Distantly Supervised Named Entity Recognition using Positive-Unlabeled
   Learning
   • Distantly Supervised NER with Partial Annotation Learning and
   Reinforcement Learning
   • Learning Named Entity Tagger using Domain-Specific Dictionary
   • Distant supervision for relation extraction without labeled data.
   • Distant supervision for relation extraction via piecewise convolutional
   neural networks.
   
   
   

   View Slide

7. Better Modeling of Incomplete Annotations

   for Named Entity Recognition
   ෆ׬શͳΞϊςʔγϣϯ͕ߦΘΕ͍ͯΔσʔλ͔ΒNERΛߦ͏ϞσϧͷఏҊ
   
   
   Zhanming Jie, Pengjun Xie, Wei Lu, Ruixue Ding, Linlin Li
   NAACL 2019
   എܠ
   • ෆ׬શͳΞϊςʔγϣϯΛԾఆ͢Δࡍʹɼ

   ୯ޠϨϕϧͰϥϕϧΛऔΓআ͘ͷ͸ݱ࣮తͰ͸ͳ͍ (A.1)
   • ໌ࣔతʹOλάΛΞϊςʔγϣϯ͢Δ͜ͱ͸ແ͍ (A.1, A.2)
   ࣮ӡ༻Ͱى͖͏Δϥϕϧܽଛ͸

   A.3ʹͳΔͱஶऀ͸ओு
   https://github.com/kajyuuen/Incomplete-NER-Methods
   ࠶ݱ࣮૷͠·ͨ͠:
   

   View Slide

8. Better Modeling of Incomplete Annotations

   for Named Entity Recognition
   
   
   Zhanming Jie, Pengjun Xie, Wei Lu, Ruixue Ding, Linlin Li
   NAACL 2019
   ௨ৗͷCRF
   ఏҊख๏ͷCRF
   ఏҊख๏
   CRFΛ֦ு͠ɼऔΓ͏Δϥϕϧͷ૊Έ߹ΘͤΛߟྀ͢ΔϞσϧ
   ֬཰෼෍ ͷਪఆ
   Hard: ࠷΋Մೳੑͷߴ͍ϥϕϧܥྻʹ֬཰1ΛׂΓ౰ͯΔ
   Soft: ͋Γ͏Δશͯͷϥϕϧܥྻʹରͯ֬͠཰ΛׂΓ౰ͯΔ
   ͜ΕΒͷ֬཰෼෍͸k෼ׂަࠩݕূʹΑͬͯਪఆΛߦ͏
   q
   q = 1
   CRF-PA
   ఏҊख๏
   

   View Slide

9. Better Modeling of Incomplete Annotations

   for Named Entity Recognition
   
   
   Zhanming Jie, Pengjun Xie, Wei Lu, Ruixue Ding, Linlin Li
   NAACL 2019
   ݁Ռ
   ৚݅ઃఆ: ޒׂͷϥϕϧܽଛ+શͯͷOλάΛ࡟আ
   • ׬શͳΞϊςʔγϣϯ͕෇͍ͨͱ͖ʹൺ΂ͯ΋·ͣ·ͣͳੑೳΛࣔ͢
   • ϥϕϧ͕෇͍͍ͯͳ͍৔ॴΛOͱͯ͠Έͳ͢Simpleʹ͸େউར
   

   View Slide

10. Distantly Supervised Named Entity Recognition 

    using Positive-Unlabeled Learning
    
    
    Minlong Peng, Xiaoyu Xing, Qi Zhang, Jinlan Fu, Xuanjing Huang
    ACL 2019
    PUֶशΛར༻ͯࣙ͠ॻͱੜςΩετ͚ͩΛ༻͍ͯNERΛߦ͏
    ఏҊख๏
    • ࠷௕Ұக๏Λ༻͍ͯɼࣙॻ͔Βੜίʔύεʹରͯ͠ΞϊςʔγϣϯΛߦ͏
    • ϥϕϧ෇͚͕ߦΘΕͨσʔλΛPositive, ͦΕҎ֎ΛUnlabeledͱֶͯ͠श
    • BIOɼBIOESͱ͍ͬͨϥϕϦϯάεΩʔϚΛར༻͠ͳ͍͜ͱͰɼ

    ࣙॻʹΑΔޡΞϊςʔγϣϯΛݮΒ͢͜ͱ͕ग़དྷΔ
    • ֤ΫϥεຖʹPU෼ྨػΛ࡞੒ɼ༧ଌ֬཰͕࠷΋ߴ͍ΫϥεΛ࠾༻
    

    View Slide

11. Distantly Supervised Named Entity Recognition 

    using Positive-Unlabeled Learning
    
    
    Minlong Peng, Xiaoyu Xing, Qi Zhang, Jinlan Fu, Xuanjing Huang
    ACL 2019
    ݁Ռ
    Ξϊςʔγϣϯίʔύε ࣙॻͱੜίʔύε
    ࣙॻϚονʹൺ΂ͯɼେ͖͘ੑೳΛ޲্ͤͨ͞
    

    View Slide

12. Distantly Supervised NER with
    Partial Annotation Learning and Reinforcement Learning
    
    
    Yaosheng Yang, Wenliang Chen, Zhenghua Li, Zhengqiu He, Min Zhang
    COLING 2018
    Distant SupervisionͱڧԽֶशͷ૊Έ߹ΘͤʹΑΔNER
    ఏҊख๏
    • ڧԽֶशΛ༻͍ͯnoisyͳϥϕϧ෇͚Λ࡟আ
    • CRF-PAΛ༻͍Δ͜ͱͰ

    imcompleteͳܥྻͰ΋ֶशΛՄೳʹ͢Δ
    എܠ
    • ࣙॻϚονʹΑͬͯ࡞੒ͨ͠ڭࢣσʔλ(Distant Supervision)ʹ͸

    imcomplete, noisyͳϥϕϧ෇͚͕ߦΘΕΔͱ͍͏໰୊͕͋Δ
    

    View Slide

13. Distantly Supervised NER with
    Partial Annotation Learning and Reinforcement Learning
    
    
    Yaosheng Yang, Wenliang Chen, Zhenghua Li, Zhengqiu He, Min Zhang
    COLING 2018
    ݁Ռ
    ࣙॻϚον΍LSTM-CRF-PAͷΈͷ৔߹ʹൺ΂ͯɼߴ͍நग़ੑೳΛ࣋ͭ
    ͕খ͞ΊͷΞϊςʔγϣϯίʔύε
    
    ͕Distantly SupervisedʹΑͬͯ࡞ΒΕͨڭࢣσʔλ
    ℋ
    
    

    View Slide

14. αʔϕΠ ײ૝
    • ෦෼తΞϊςʔγϣϯίʔύεΛֶश͢ΔϞσϧ͸සൟʹݟΒΕͨ
    • Distantly Supervised NERͷݚڀͰ͸͔ͳΓͷ֬཰Ͱར༻ɼҾ༻͞ΕΔ
    • σʔληοτͲ͏͢Δͷ໰୊
    • ࣗ࡞͢Δύλʔϯ͕͔ͳΓଟ͍
    • CoNLL2003͔ΒϥϯμϜܽଛͤ͞Δύλʔϯ΋ݟΔ
    • ͲͪΒʹ͠Ζ࠶ݱੑͷ͋ΔϥϕϧܽଛΛߦ͏͜ͱ͕೉͍͠
    • ྲྀߦ͖͍ͬͯͯΔײ͡͸͢Δ
    • Nested NER΍DS NERͷΑ͏ͳෳࡶͳλεΫઃఆͷ࿦จଟ͍
    • ࣙॻ͚ͩͰNERΛߦ͍͍ͨؾ࣋ͪͷਓͨͪ͸୔ࢁ͍ͦ͏
    • ී௨ͷܥྻϥϕϦϯάλεΫʹݶք͕དྷ͍ͯΔʁ
    
    
    

    View Slide