DeBERTA : Decoding-Enhanced BERT with Disentangled Attention
- 1. DEBERTA Decoding-Enhanced BERT with Disentangled Attention Arxiv 2020.06.05 || ICLR 2021 🤗 딥러닝 논문 읽기 모임 4기 NLP팀 진명훈 김수빈 신문종 아이린 최상우
- 2. TABLE OF CONTENTS 2 Introduction Background 3 Contributions (2 Q&A) Experiment Conclusion
- 3. Introduction 3 • He et al., 2020 에서 제안된 모델 • DeBERTa: Decoding-enhanced BERT with Disentangled Attention • Google의 BERT(2018)과 Facebook(현재 Meta)의 RoBERTa(2019) 기반 • RoBERTa + disentangled attention + enhanced mask decoder • With half of the data used in RoBERTa (80GB) • Scale Invariant Fine-Tuning 도입 • #5929 PR로 🤗transformers에 Merge됨 • Outperform RoBERTa an a majority of NLU tasks • e.g., SQuAD, MNLI and RACE
- 4. Background 4 • Positional Information • Masked Language Model • Adversarial Training
- 5. Background: Positional Information 5 • The standard self-attention mechanism lacks a natural way to encode word position information • Add Positional Bias (ref: 딥논읽 Rotary Embedding 발표) • Absolute Position Embedding • Relative Position Embedding
- 6. Background: Masked Language Model 6 • Large-scale Transformer-based PLMs are typically pre-trained on large amounts of text to learn contextual word representations using a self- supervision objective, known as Masked Language Model (MLM) max 𝜃 log 𝑝𝜃(𝑋| ෨ 𝑋) ≈ max 𝜃 𝑖∈𝐶 log 𝑝𝜃 𝑥𝑖 = 𝑥𝑖 ෨ 𝑋
- 7. Background: Adversarial Training 7 • 정상 데이터를 모델에 학습 • 정상 데이터 + adversarial sample을 같이 학습 • 일반화 성능 향상
- 8. 3 Contributions 8 • Disentangled attention • Transformer-xl처럼 additive하게 attention을 분해 • Shaw, Transformer-xl과 다르게 position-to-content term을 살림 • query token의 위치가 달라지는 부분도 반영 • Position-to-position term은 RPE에서 불필요하기 때문에 제거 • Enhanced Mask Decoder • A new store opened beside the new mall • Absolute position information 또한 중요하다! • Scale Invariant Fine-Tuning • Adversarial Training은 모델의 일반화에 도움을 준다 • NLP에서 embedding vector norm의 분산은 모델바이모델, 단어바이단어 • Word embedding을 normalize해주고 Perturbation을 추가하자!
- 9. Disentangled Attention 9 • Disentangled Attention: A two vector approach to content and position embedding • 논문에서 아래와 같은 수식을 제안하며 token repr을 content와 position에 대한 두 벡터로 decomposition 수행 • 이렇게 쪼개는 것은 사실 Transformer-XL에서 제안되었어요! 𝐴𝑖,𝑗 = 𝐻𝑖, 𝑃𝑖|𝑗 × 𝐻𝑗, 𝑃𝑗|𝑖 𝑇 = 𝐻𝑖𝐻𝑗 𝑇 + 𝐻𝑖𝑃𝑗|𝑖 𝑇 + 𝑃𝑖|𝑗𝐻𝑗 𝑇 + 𝑃𝑖|𝑗𝑃𝑗|𝑖 𝑇
- 10. History: Relative Position Embedding 10 17.06.12 18.03.06 19.10.23 Transformer Shaw RPE T5 18.09.12 Music Transformer 19.01.09 Transformer-XL 19.06.19 XLNet 20.06.05 DeBERTa
- 11. History: Relative Position Embedding 11 Transformer upgrade! Layer에 직접 위치 정보를 주입하자! Self-Attention with Relative Position Representations Music Transformer
- 12. History: Relative Position Embedding 12 Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context XLNet: Generalized Autoregressive Pretraining for Language Understanding 𝐴𝑖,𝑗 𝑎𝑏𝑠 = 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝐸𝑥𝑗 + 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑈𝑗 + 𝑈𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝐸𝑥𝑗 + 𝑈𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑈𝑗 (a) (b) (c) (d) Content-to-Content Content-to-Position Position-to-Content Position-to-Position
- 13. History: Relative Position Embedding 13 Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context XLNet: Generalized Autoregressive Pretraining for Language Understanding 𝐴𝑖,𝑗 𝑎𝑏𝑠 = 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝐸𝑥𝑗 + 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑈𝑗 + 𝑈𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝐸𝑥𝑗 + 𝑈𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑈𝑗 (a) (b) (c) (d) Content-to-Content Content-to-Position Position-to-Content Position-to-Position 𝐴𝑖,𝑗 = 𝐻𝑖, 𝑃𝑖|𝑗 × 𝐻𝑗, 𝑃𝑗|𝑖 𝑇 = 𝐻𝑖𝐻𝑗 𝑇 + 𝐻𝑖𝑃𝑗|𝑖 𝑇 + 𝑃𝑖|𝑗𝐻𝑗 𝑇 + 𝑃𝑖|𝑗𝑃𝑗|𝑖 𝑇
- 14. History: Relative Position Embedding 14 𝐴𝑖,𝑗 𝑎𝑏𝑠 = 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝐸𝑥𝑗 + 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑈𝑗 + 𝑈𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝐸𝑥𝑗 + 𝑈𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑈𝑗 𝐿𝑒𝑡 𝑄 = 𝑊 𝑞𝐸𝑥𝑖 𝑈𝑞 = 𝑊 𝑞𝑈𝑖 𝐾 = 𝑊𝑘𝐸𝑥𝑗 𝑈𝑘 = 𝑊𝑘𝑈𝑗 𝐴𝑖,𝑗 𝑎𝑏𝑠 = 𝑄𝑇𝐾 + 𝑄𝑇𝑈𝑘 + 𝑈𝑞 𝑇𝐾 + 𝑈𝑞 𝑇𝑈𝑘 이 수식을 분석해봅시다! 이렇게 미리 정해둘게요 그러면 수식이 이렇게 정리됩니다! 이걸 묶어서 정리하면? 𝐴𝑖,𝑗 𝑎𝑏𝑠 = 𝑄 + 𝑈𝑞 𝑇 (𝐾 + 𝑈𝑘) 즉, 이렇게 정리할 수 있겠군요! 𝐴𝑖,𝑗 𝑎𝑏𝑠 = 𝑊 𝑞 𝐸𝑥𝑖 + 𝑈𝑖 𝑇 𝑊𝑘 𝐸𝑥𝑗 + 𝑈𝑗 𝑇 0.09 0.05 0.86 0. 0.05 0.91 0.04 0. 0.50 0. 10 0.40 0. 0.45 0.40 0.15 0. 𝑄𝑊 𝑖 𝑄 ∈ ℝ4×3 𝑞1 𝑞2 𝑞3 𝑞4 </s> 𝑡1 𝑡2 𝑡3 Embedding + Projection 𝑘1 𝑘2 𝑘3 𝑘4 𝐾𝑊𝑖 𝐾 ∈ ℝ3×4 𝐴𝑡𝑡 ∈ ℝ4×4 𝑎11 𝑎12 𝑎13 𝑎14
- 15. 𝐴𝑖,𝑗 𝑎𝑏𝑠 = 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝐸𝑥𝑗 + 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑈𝑗 + 𝑈𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝐸𝑥𝑗 + 𝑈𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑈𝑗 𝐴𝑖,𝑗 𝑅𝑃𝐸 = 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘,𝐸𝐸𝑥𝑗 + 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇𝑊𝑘,𝑅𝑅𝑖−𝑗 + 𝑢𝑇𝑊𝑘,𝐸𝐸𝑥𝑗 + 𝑣𝑇𝑊𝑘,𝑅𝑅𝑖−𝑗 Learned position 𝐴𝑖,𝑗 𝑎𝑏𝑠 = 𝑊 𝑞 𝐸𝑥𝑖 + 𝑈𝑖 𝑇 𝑊𝑘 𝐸𝑥𝑗 + 𝑈𝑗 𝑇 𝐴𝑖,𝑗 𝑆ℎ𝑎𝑤_𝑅𝑃𝐸 = 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇 𝑊𝑘𝐸𝑥𝑗 + 𝐸𝑥𝑖 𝑇 𝑊 𝑞 𝑇 𝑊𝑘𝑅𝑖−𝑗 Dai(transformer-xl의 주저자)의 연구는? Sinusoid position Shaw의 연구는? History: Relative Position Embedding 15
- 16. Shaw’s RPE Transformer-XL T5 q𝑚 𝑇 𝑘𝑛 = 𝑥𝑚 𝑇 𝑊 𝑞 𝑇 𝑊𝑘𝑥𝑛 + 𝑥𝑚 𝑇 𝑊 𝑞 𝑇 ෩ 𝑊𝑘 𝑝𝑚−𝑛 + 𝑢𝑇 𝑊𝑘𝑥𝑛 + 𝑣𝑇 ෩ 𝑊𝑘 𝑝𝑚−𝑛 q𝑚 𝑇 𝑘𝑛 = 𝑥𝑚 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑥𝑛 + 𝑥𝑚 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑝𝑚−𝑛 q𝑚 𝑇 𝑘𝑛 = 𝑥𝑚 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑥𝑛 + 𝑏𝑚,𝑛 History: Relative Position Embedding
- 17. Shaw’s RPE Transformer-XL T5 q𝑚 𝑇 𝑘𝑛 = 𝑥𝑚 𝑇 𝑊 𝑞 𝑇 𝑊𝑘𝑥𝑛 + 𝑥𝑚 𝑇 𝑊 𝑞 𝑇 ෩ 𝑊𝑘 𝑝𝑚−𝑛 + 𝑢𝑇 𝑊𝑘𝑥𝑛 + 𝑣𝑇 ෩ 𝑊𝑘 𝑝𝑚−𝑛 q𝑚 𝑇 𝑘𝑛 = 𝑥𝑚 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑥𝑛 + 𝑥𝑚 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑝𝑚−𝑛 q𝑚 𝑇 𝑘𝑛 = 𝑥𝑚 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑥𝑛 + 𝑏𝑚,𝑛 History: Relative Position Embedding DeBERTa q𝑚 𝑇 𝑘𝑛 = 𝑥𝑚 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑥𝑛 + 𝑥𝑚 𝑇 𝑊 𝑞 𝑇𝑊𝑘 𝑝𝑚−𝑛 + 𝑝𝑚−𝑛 𝑇 𝑊 𝑞 𝑇𝑊𝑘𝑥𝑛
- 18. Disentangled Attention 18 • Shaw 연구진 등의 기존 RPE 접근 방법은 content-to-content (a) term과 content-to-position (b) term을 사용하여 attention weights를 계산 • Attention weight는 어느 한 쪽 방향으로만 모델링할 수 없다. • Position-to-content term (c) term 또한 중요하다! • Relative position embedding에서 (d) term은 이미 고려하고 있음
- 19. Disentangled Attention 19 • k: maximum relative distance • 𝛿 𝑖, 𝑗 ∈ 0,2𝑘 • 𝛿 𝑖, 𝑗 = ቐ 0 2𝑘 − 1 𝑖 − 𝑗 + 1 𝑖 − 𝑗 ≤ −𝑘 𝑖 − 𝑗 ≥ 𝑘 for for o. w
- 21. Enhanced Mask Decoder 21 • DeBERTa는 MLM으로 pre-trained • MLM을 위해 context words의 content와 position information을 활용 • 하지만 absolute positions을 고려하지 않음 • e.g., • A new store opened beside the new mall • BERT는 absolute positions을 input layer에 주입 • DeBERTa는 Transformer layer를 전부 거치고 Masked token prediction을 위 해 softmax layer를 통과시키기 전에 absolute positions을 주입
- 24. Enhanced Mask Decoder 24 https://github.com/microsoft/DeBERTa/blob/master/DeBERTa/apps/run.py Seed 고정 Tokenizer, task object 반환 Eval, test data load Load train data and get model
- 28. BertForMaskedLM 28 Transformer Layer with disentangled attention Transformer Layer with disentangled attention Transformer Layer with disentangled attention Transformer Layer with disentangled attention Transformer Layer with disentangled attention … Sub-word embedding Token type embedding + Encoder output Encoder output Encoder output Encoder output Encoder output Absolute position embedding Token type embedding … CLS 딥 ##러 MASK 논 MASK 모임 SEP lm_head Token type embedding … CLS 딥 ##러 ##닝 논 ##문 모임 SEP lm_logits, lm_loss BERT Module
- 29. DeBERTaForMaskedLM with EDM 29 Transformer Layer with disentangled attention Transformer Layer with disentangled attention Transformer Layer with disentangled attention Transformer Layer with disentangled attention Transformer Layer with disentangled attention … Sub-word embedding Token type embedding + DeBERTa Module Encoder output Encoder output Encoder output Encoder output (H) Encoder output Absolute position embedding • (1)은 position_biased_input 옵션이 True인 경우에만 더해줌 • 분홍색 Transformer Layer는 shared • lm_head는 word embedding matrix와 shared • 저자에 의하면 EDM이 누락되도 PLM의 수렴에 영향을 끼치지 않는다고 함 • MLM training의 perplexity에 약간의 영향을 미치는 부분 Token type embedding … CLS 딥 ##러 MASK 논 MASK 모임 SEP Query state (I) + Transformer Layer with disentangled attention Transformer Layer with disentangled attention Query state (I) EDM Module (n=2) (1) Encoder output Token type embedding … CLS 딥 ##러 ##닝 논 ##문 모임 SEP lm_head lm_logits, lm_loss
- 30. Scale Invariant Fine-Tuning 30 • Virtual adversarial training은 regularization method • 모델의 일반화 성능을 강화 • Input에 small perturbation(noise)를 줘서 adversarial attack에도 동일한 output prediction을 만드는 것이 목적 • NLP task에서 perturbation은 word embedding에 주어짐 • 그러나 model by model, word by word로 emb vector의 norm은 상이함 • Bigger model일수록 분산은 커지고 adversarial training의 불안정성을 키움 • Layer norm에서 영감을 받아 normalized word embeddings에 perturbation을 추가하여 Adversarial Fine-Tuning • 1.5B 모델에만 적용했고 comprehensive study는 향후에 진행할 예정
- 33. Scale Invariant Fine-Tuning 33 Embedding Module DeBERTa Module Token type embedding … CLS 딥 ##러 ##닝 논 ##문 모임 SEP Task-specific Layer (SuperGLUE) Token type embedding … O B-XX I-XX I-XX B-XX I-XX I-XX O ℒ(𝑙𝑜𝑔𝑖𝑡𝑠, 𝑔𝑜𝑙𝑑𝑒𝑛_𝑡𝑟𝑢𝑡ℎ) Sub-word embedding Token type embedding Absolute position embedding + If position_biased_input LayerNorm Sift hook!
- 34. Scale Invariant Fine-Tuning 34 Embedding Module DeBERTa Module Token type embedding … CLS 딥 ##러 ##닝 논 ##문 모임 SEP Task-specific Layer (SuperGLUE) Token type embedding … O B-XX I-XX I-XX B-XX I-XX I-XX O ℒ(𝑙𝑜𝑔𝑖𝑡𝑠, 𝑔𝑜𝑙𝑑𝑒𝑛_𝑡𝑟𝑢𝑡ℎ) Sub-word embedding Token type embedding Absolute position embedding + If position_biased_input LayerNorm LayerNorm(inputs) https://github.com/microsoft/DeBERTa/blob/master/DeBERTa/sift/sift.py#L29
- 35. Scale Invariant Fine-Tuning 35 Embedding Module DeBERTa Module Token type embedding … CLS 딥 ##러 ##닝 논 ##문 모임 SEP Task-specific Layer (SuperGLUE) Token type embedding … O B-XX I-XX I-XX B-XX I-XX I-XX O ℒ(𝑙𝑜𝑔𝑖𝑡𝑠, 𝑔𝑜𝑙𝑑𝑒𝑛_𝑡𝑟𝑢𝑡ℎ) Sub-word embedding Token type embedding Absolute position embedding + If position_biased_input LayerNorm LayerNorm(inputs) https://github.com/microsoft/DeBERTa/blob/master/DeBERTa/sift/sift.py#L29 + perturbation delta ~𝑁 0, 0.02 If 𝛿 ≥ 0.04 𝑜𝑟 𝛿 ≤ −0.04, clamp 𝑎𝑑𝑣
- 37. Experiment: Pre-training 37 • RoBERTa처럼 dynamic data batching 적용 • SpanBERT처럼 span masking 적용
- 48. Conclusion 48 • Disentangled attention과 enhanced mask decoder로 RoBERTa 개선 • Downstream task에서 모델 일반화를 개선하기 위해 SIFT 제안 • Macro score 측면에서 SuperGLUE 벤치마크에서 인간의 성능을 상회 • 아직 인간 수준의 지능까진 도달하지 못함 • 최근에 V3가 나왔습니다! → 다음 차례에 wrap-up하며 발표하도록 하겠습니다.