文献紹介：MoverScore: Text Generation Evaluating with Contextualized Embeddings and Earth Mover Distance

1. จݙ঺հʢʣ
   MoverScore: Text Generation Evaluating with Contextualized Embeddings and Earth Mover Distance Wei Zhao† , Maxime Peyrard† , Fei Liu‡ , Yang Gao† , Christian M. Meyer† , Steffen Eger† EMNLP2019 ௕Ԭٕज़Պֶେֶ
   ࣗવݴޠॲཧݚڀࣨɹ
   ૬ాɹଠҰ
2. Abstract • ੜ੒ͷλεΫʹ͓͍ͯɺؤڧͳධՁई౓Λௐࠪ • จ຺Λߟྀͨ͠୯ޠ෼ࢄදݱ ͱ Word Mover’s Distance ͷ૊Έ߹Θ͕ͤ࠷΋ྑ͔ͬͨ • ιʔείʔυΛެ։ɹɹɹɹɹɹɹɹɹɹɹɹɹɹɹ https://github.com/AIPHES/emnlp19-moverscore 2
3. Related work • ৭ʑͳධՁख๏ʢ1ʣ • ཁ໿ɿROUGE(Lin 2004) • ػց຋༁ɿBLEU(Papinemi 2002), RUSE(Shimanaka 2018) • Image CaptioningɿBLEU, CIDEr(Vedantam 2015), SPICE(Anderson 2016) 3 #-&6
   ͸޲͔ͳ͍
4. Related work • ৭ʑͳධՁख๏ʢ2ʣ • ҙຯతྨࣅ౓ɿ “BERTScore”(Zhang 2019) • ຋༁ɿڭࢣ͋Γɾڭࢣͳ͠ BERT ෼ࢄදݱ(Mathur 2019) • ཁ໿ɺΤοηΠ࠾఺ɿELMo + Sentence Mover’s Simirality(Clark 2019) 4 จ຺Λߟྀͨ͠෼ࢄදݱ $POUFYUVBMJ[FE
   SFQSFTFOUBUJPO
   Λ༻͍Δख๏͕૿͖͑ͯͨ ࣮ݧͷ
   #BTFMJOF
   ʹग़͖ͯ·͢
5. Method • ༷ʑͳੜ੒λεΫΛධՁͰ͖Δࢦඪ(MoverScore)Λௐࠪ • ੜ੒จͱࢀরจͷྨࣅ౓ʢʁʣΛଌΔ • จ຺Λߟྀͨ͠෼ࢄදݱɿBERT, ELMo • ग़ྗจͱࢀরจͷҙຯతڑ཭ɿWord Mover's Distance 5
6. Method • MoverScore Variations • Granularityɿn-gram (n=1, 2, size-of-sentence) • Embeddingɿword2vec, BERT, ELMo • Fine-tuningɿMultiNLI, QANLI, QQP • Aggregationɿpower means, routing mechanism 6 /-* 1BSBQISBTJOH #&35
   &-.P #&35
7. Method • MoverScore Variations • Granularityɿn-gram (n=1, 2, size-of-sentence) • Embeddingɿword2vec, BERT, ELMo • Fine-tuningɿMultiNLI, QANLI, QQP • Aggregationɿpower means, routing mechanism 7 #&35
   &-.P
8. Method • Aggregation ʢ౷߹ํ๏ʣ • จ຺Λߟྀͨ͠෼ࢄදݱɿBERT, ELMo • ֤୯ޠ͸֤૚͔ΒͦΕͧΕҟͳΔϕΫτϧ͕౉͞ΕΔ • Power MeansɿฏۉΛऔΓ ৐( )ɺconcat • Routing Mechanismɿৄ͘͠͸(Zhang 2018) p p = 1, ± ∞ 8
9. Method • ग़ྗจͱࢀরจͷҙຯతڑ཭ • Word Mover's Distance (WMD) • Sentence Mover's Distance (SMD) • ઌ΄Ͳͷ૊Έ߹ΘͤΛɺWMD, SMD ͦΕͧΕͰݕূ͢Δ 9
10. Experiment • Tasks • ػց຋༁ • ཁ໿ • ର࿩ʢλεΫࢤ޲ʣ • Image Captioning 10 ʢࢀরจɺෳ਺ͷγεςϜʹΑΔग़ྗจʣͷϖΞ γεςϜͷग़ྗจʹ͸ਓखධՁ͕͞Ε͍ͯΔ ʲධՁࢦඪɺMoverScore Ͱ΍Δ͜ͱʳ ɾγεςϜͷग़ྗจΛධՁ ɾਓखධՁͱͷ૬ؔΛݟΔ
11. Experiment • ػց຋༁ • DatasetɿWMT2017 • ࢀՃγεςϜͷग़ྗจʹɺ࠷௿Ͱ΋15ਓͷਓखධՁ • BaselinesɿSentBLEU, METEOR++, RUSE, BERTScore(Zhang 2019) 11
12. Result • WMD+BERT+MNLI+PMeans ͕ Baseline Λ্ճΔ 12
13. Result • Sentence Representation Ͱ͸৘ใ͕ࣦΘΕΔʁ 13
14. Experiment • ཁ໿ • DatasetɿTAC-2008, TAC-2009 • Responsivenessɿ಺༰ʴจ๏తͳ඼࣭ • Pyramidɿࢀরจʹؚ·ΕΔॏཁͳ಺༰͕ͲΕ͚ͩଟ͘Χόʔ͞ Ε͍ͯΔ͔ • BaselinesɿROUGE-1, ROUGE-2, (Peyrard 2017), BERTScore(Zhang 2019) S3 best 14 ڭࢣ͋ΓͷධՁࢦඪ
15. Result • WMD+BERT+MNLI+PMeans Ͱ Baselines Λ্ճΔ 15
16. Experiment • ର࿩ʢλεΫࢤ޲ʣ • DatasetɿBAGEL, SFHOTEL • Informativeness (Inf)ɿఏڙ͢Δ৘ใྔ • Naturalness (Nat)ɿਓͷԠ౴΁ͷۙ͞ • Quality (Qual)ɿྲྀெੑɾจ๏ • BaselinesɿBLEU, METEOR, BERTScore(Zhang 2019) 16
17. Result • શମతʹ૬͕ؔ௿͍͕ɺఏҊख๏͸ͦͷதͰ΋ߴ͍ํ 17
18. Experiment • Image Captioning • DatasetɿMSCOCO • M1 ~ M5 ͷධՁ͕͋Δ • ࠓճ͸ɺશମͷ඼࣭ʹؔ͢ΔM1, M2 Λ࠾༻ • BaselinesɿCIDEr, SPICE, METEOR, LEIC(Cui 2018), BERTScore(Zhang 2019) 18 ڭࢣ͋ΓͷධՁࢦඪ
19. Result • Baseline ͷ LEIC ʹྼΔ͕ɺͦΕͰ΋ߴ͍૬ؔΛࣔ͢ 19 M: BERT fine-tuning ʹ MultiNLI Λ࢖༻ P: ELMo / BERT ͷ౷߹ (Aggregation) ʹ Power Means Λ࢖༻
20. Discussion • ࣮ݧͷ Baseline ͱͯ͠ग़͖ͯͨ BERTScore ͱͷൺֱ 20
21. Discussion • ࣮ݧͷ Baseline ͱͯ͠ग़͖ͯͨ BERTScore ͱͷൺֱ 21 One-to-one ͷڧ͍ alignment Many-to-one ͷऑ͍ alignment WMD Ͱద੾ͳڑ཭ ͕औΕ͍ͯΔ
22. Discussion • ػց຋༁ͰਓखධՁͷߴ͍΋ͷ(good)ͱ௿͍΋ͷ(bad)ͷɹ 2ͭʹ෼͚ɺ෼෍Λௐࠪ • ൺֱର৅ • Baseline: SentBLEU • Proposal: MoverScore(WMD+BERT) 22
23. Discussion • SentBLEU ͸ਓखධՁ͕ྑͯ͘΋தఔ౓ͷՕॴʹଟ͘෼෍ • MoverScore ͸៉ྷʹ2ͭͷۃΛදݱͰ͖͍ͯΔ 23
24. Conclusion • ੜ੒λεΫͷڭࢣͳ͠ධՁࢦඪΛఏҊ • 4ͭͷੜ੒λεΫͰ Baselines Λ ௒͑Δ/ഭΔ ݁Ռʹ • ιʔείʔυΛެ։ɹɹɹɹɹɹɹɹɹɹɹɹɹɹɹ https://github.com/AIPHES/emnlp19-moverscore 24