HephaestusProject / pytorch-ReCoSa

MIT License
1 stars 0 forks source link

[PAPER] ReCoSa: Detecting the Relevant Contexts with Self-Attention for Multi-turn Dialogue Generation #1

Open soeque1 opened 4 years ago

soeque1 commented 4 years ago

πŸ“ Introduction

[NLP] ReCoSa: Detecting the Relevant Contexts with Self-Attention for Multi-turn Dialogue Generation

Why?

λŠ” 훼이크고 https://github.com/modulabs/beyondBERT λ°œν‘œ μ˜ˆμ •μž…λ‹ˆλ‹€ γ… 

Issue card

Issues

  1. PPL μ„±λŠ₯이 논문에 μ œμ‹œλœ 것 만큼 λ‚˜μ˜€μ§€ μ•ŠμŠ΅λ‹ˆλ‹€. μ•„λ§ˆλ„ tokenizer κ°€ λ‹€λ₯Έ μ΄μŠˆκ°€ μžˆμ„ 것 κ°™μŠ΅λ‹ˆλ‹€. ν˜„μž¬ κ΅¬ν˜„μ½”λ“œλŠ” ν•™μŠ΅ μ½”νΌμŠ€μ˜ vocab이 μ•„λ‹ˆλΌ, GPT2의 vocab을 μ‚¬μš©ν•˜κ³  μžˆμŠ΅λ‹ˆλ‹€ (vocab의 μˆ«μžκ°€ 더 λ§ŽμŠ΅λ‹ˆλ‹€)

Reference

https://arxiv.org/abs/1907.05339

ssaru commented 4 years ago

@soeque1 ν”„λ‘œμ νŠΈλ₯Ό Automated Kanban으둜 μ„€μ •ν–ˆμŠ΅λ‹ˆλ‹€! γ…Žγ…Ž