Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Read: "DeepType 2: Superhuman Entity Linking, A...

Read: "DeepType 2: Superhuman Entity Linking, All You Need Is Type Interactions"

AAAI-22読み会で下記の論文を読みました。

Jonathan Raiman, "DeepType 2: Superhuman Entity Linking, All You Need Is Type Interactions", AAAI-22

Avatar for Tatsuya Shirakawa

Tatsuya Shirakawa

April 06, 2022
Tweet

More Decks by Tatsuya Shirakawa

Other Decks in Research

Transcript

  1. UCPhrase: Unsupervised Context-aware Quality Phrase Tagging DeepType 2: Superhuman Entity

    Linking, All You Need Is Type Interactions 
 Jonathan Raiman (NVIDIA) 2022-04-06 Reader: Tatsuya Shirakawa LINE Developer Meetup: AAAI-22 ࿦จಡΈձ
  2. 2 Beatrust – vision ୭΋͕࠷ߴͷࣗ෼Λ ࣮ݱͰ͖ΔੈքΛͭ͘Δ We are hiring! Machine

    Learning Engineer | ػցֶशΤϯδχΞ Data Scientist | σʔλαΠΤϯςΟετ Analytics Engineer | ΞφϦςΟΫεΤϯδχΞ https://speakerdeck.com/beatrust/we-are-hiring
  3. 3 Tatsuya Shirakawa @s_tat1204 Machine Learning Lead at Beatrust 


    ← Researcher at ABEJA ← Researcher at NTT-Data Mathematical Systems Inc. Selected Posts • ϩʔϯνલͷ Tag Suggestion ػೳΛػցֶशͰ࡞Δ • ೔ຊޠࣙॻ͋ΓΩʔϫʔυநग़ث͔Βͷଟݴޠࣙॻͳ͠Ωʔ ϫʔυநग़ثͷ Distillation • ػ͸ख़ͨ͠ʂάϥϑߏ଄ʹର͢ΔDeep LearningɺGraph Convolutionͷ͝঺հ • ҟۭؒ΁ͷຒΊࠐΈʂPoincare Embeddings͕୓͘දݱֶशͷ ৽ల։ • Retail Face Analysis Inside-Out ML Data Science NLP Mathematical Optimization Interview https://note.com/beatrust/n/ne63297b9c546 CV
  4. 5 Alias TableΛ΋͍ͪͨޮ཰తͳEntity Linking https://qiita.com/izuna385/items/9d658620b9b96b0b4ec9 Λ΋ͱʹ࡞੒ Step 0. Alias TableΛԿΒ͔ͷํ๏Ͱ࡞͓ͬͯ͘ʢWikipediaͷ಺෦ϦϯΫΛऩूͨ͠Γʣ

    Step 1. จதͷϝϯγϣϯʹର͠ɺީิͱͳΔΤϯςΟςΟͷϦετΛAlias Table͔Βऔಘ Step 2. ީิΛϥϯΩϯά͠ɺτοϓީิΛϝϯγϣϯͷ༧ଌΤϯςΟςΟͱͯ͠ग़ྗ Alias Table
  5. 9 Paper Summary DeepTypeΛվળͨ͠Entity Linkingͷ৽ख๏DeepType 2ΛఏҊ 1. ਓؒͷύϑΥʔϚϯεͱͷൺֱΛߦ͏ͨΊɺ໖ີͳΞϊςʔ γϣϯΛߦͬͨσʔληοτΛ࡞੒ɾެ։ɻͦͷ݁Ռɺ SoTA͸ਓؒͷύϑΥʔϚϯεʹୡ͍ͯ͠ͳ͍͜ͱΛ

    
 ໌Β͔ʹͨ͠ɻ 
 2. Typeʢϝϯγϣϯ/ΤϯςΟςΟͷΧςΰϦ৘ใͳͲʣͷਪ ఆΛجૅͱͨ͠DeepTypeΛൃలͤͨ͞৽ख๏DeepType 2Λ ఏҊɻֶशࡁΈݴޠϞσϧΛ͔ͭΘͣʹɺॳΊͯਓؒΛ྇ ͙ਫ਼౓Λୡ੒ɻ Single authorͳͷʹ͔ͳΓؤு͍ͬͯΔײͷ͋Δ࿦จͩͬͨɻ
  6. 11 DeepType 2ͷશମ૾ ೖྗςΩετ͔ΒͷBi-LSTMϕʔεͷ 
 ϝϯγϣϯͷಛ௃நग़ ᶃ Type neighborhoods ᶄ

    Latent type interactions ᶅ Discrete type interactions LSTMʹΑΔஞ࣍༧ଌ ݁ ߹ Alias Tableதͷ֤ΤϯςΟςΟʹର͢Δಛ௃நग़ 
 ᶃ Type neighborhoods … KB্Ͱͷؔ࿈৘ใ༝དྷͷಛ௃நग़ 
 ᶄ Latent type interactions … LSTMͷঢ়ଶͱType neighborhood༝དྷͷಛ௃நग़ 
 ᶅ Discrete type interactions … ϧʔϧϕʔεʢطఆͷΫΤϦʔʹର͢Δ݁Ռʣ༝དྷͷಛ௃நग़
  7. 16 ਓؒͷύϑΥʔϚϯεͷଌఆ Amazon Mechanical TurkͰ༷ʑͳ޻෉Λ͠ɺߴ͍Ұக౓ͷਖ਼֬ͳΞϊςʔγϣϯΛಘͨɻ 1. AMT's Master ͷশ߸Λ͍࣋ͬͯΔਓ͚ͩʹݶఆͨ͠ 2.

    ਖ਼ղͨ͠ΒϘʔφεΛ෷ͬͨ 3. આ໌ΛಡΜ͔ͩɺςετηοτͰ࠷௿ݶͷਫ਼౓͕ग़͍ͯΔ͔Ͱ଍੾Γͨ͠ 4. ֤mentionʢϑϨʔζʣʹ3ਓׂ౰ɺ࠷΋Ұகͨ͠΋ͷΛ౴͑ͱͨ͠ʢOracleʣ
  8. 20 Ablation Study – Entity Vectorͷߏ੒๏ͷൺֱ طଘSoTAख๏Ͱ࢖ΘΕ͍ͯΔಛ௃ϕΫλʢunique entity vectorɺී௨ͷembedding?ʣʹtype interactionsΛՃ͑Δͱେ෯ʹਫ਼౓

    ޲্ɻ·ͨɺunique entity vectorͷ͔ΘΓʹtype neighborhoodsΛ࢖͏ͱ͞Βʹਫ਼౓޲্ʢύϥϝʔλ਺΋type neighborhoods ͸unique entity vectorͷ1/6Ͱ͢Ήʣɻ
  9. 23 ·ͱΊ - Wikidata/Wikipediaͷ৘ใΛϑϧ׆༻ͯ͠Entity LinkingΛߦ͏DeepType 2ΛఏҊ - ҆௚ʹPretrained Language ModelϞσϧΛ࢖Θͳͯ͘΋ɺType৘ใΛ͏·͘࢖͏͜ͱͰਫ਼౓Α͘

    Entity Linking͕Ͱ͖Δ͜ͱ͕Α͘Θ͔ͬͨ - ਓؒͷਫ਼౓Λਪఆ͢ΔͨΊʹAMTͰਫ਼៛ͳΞϊςʔγϣϯΛಘΔͨΊͷ਺ʑͷ޻෉Λߦ͓ͬͯΓɺ ΞϊςʔγϣϯͷઃܭΛ͢ΔࡍͷࢀߟʹͳΓͦ͏ 
 ؾʹͳͬͨͱ͜Ζ 
 - ద੾ͳTypeͷબͼํ͸೉ͦ͠͏ 
 - Wikidata/Wikipediaʹด͡ͳ͍EntityΛѻ͓͏ͱͨ͠ͱ͖ʹͲ͏ͳΔͷ͔ 
 - Pretrained modelΛ࢖༻ͨ͠ͱ͖ʹਫ਼౓޲্͕͋Δͷ͔ 

  10. 24 Beatrust – vision ୭΋͕࠷ߴͷࣗ෼Λ ࣮ݱͰ͖ΔੈքΛͭ͘Δ We are hiring! Machine

    Learning Engineer | ػցֶशΤϯδχΞ Data Scientist | σʔλαΠΤϯςΟετ Analytics Engineer | ΞφϦςΟΫεΤϯδχΞ https://speakerdeck.com/beatrust/we-are-hiring Thanks!