Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
サブセット探索を用いた高速なkNNニューラル機械翻訳
Search
Hiroyuki Deguchi
March 22, 2024
Research
180
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
サブセット探索を用いた高速なkNNニューラル機械翻訳
第8回AAMTセミナー
AAMT若手翻訳研究会
最優秀賞
Hiroyuki Deguchi
March 22, 2024
More Decks by Hiroyuki Deguchi
See All by Hiroyuki Deguchi
20260427-EhimeDSCSeminar-MBR-CBDT-Decoding
de9uch1
1
12
260624_NLP-colloquium: Hubness
de9uch1
2
210
20250226 NLP colloquium: "SoftMatcha: 10億単語規模コーパス検索のための柔らかくも高速なパターンマッチャー"
de9uch1
1
830
20240820: Minimum Bayes Risk Decoding for High-Quality Text Generation Beyond High-Probability Text
de9uch1
0
370
20240226_AAMT-Japio
de9uch1
0
210
Searching for Needles in a Haystack: On the Role of Incidental Bilingualism in PaLM’s Translation Capability
de9uch1
0
180
Paper Reading: Sampling-Based Approximations to Minimum Bayes Risk Decoding for Neural Machine Translation
de9uch1
0
240
My Research Environmental Setup
de9uch1
0
350
Nearest Neighbor Machine Translation
de9uch1
0
300
Other Decks in Research
See All in Research
論文紹介: Understanding Epistemic Language with a Language-augmented Bayesian Theory of Mind
hisaokatsumi
0
150
LA-Bench 2025:実験指示から実行可能手順を生成するためのデータセット/LA-Bench 2025: A Dataset for Generating Executable Experimental Procedures from Experimental Instructions
stktu
0
150
AIエージェント時代のLLM-jpモデルのあるべき姿
k141303
0
610
XDPerf: A High-Performance Traffic Generator Built with WASM and eBPF
takehaya
1
280
[SNLP2026] Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
wataruuuuu
0
270
大規模言語モデルは誰を覚えているか / Who Do Large Language Models Memorize?
upura
0
150
PGDM: Physically Guided Diffusion Model for L Downscaling
satai
3
450
医療LLMの現在地〜最新研究から社会実装までを考える〜
kento1109
1
1.7k
シングルチャネルマルチトーカー音声認識の進展
ryomasumura
0
240
HAKARI-Bench - 実運用視点での情報検索モデル評価ベンチマーク
hotchpotch
1
710
[BlackHatAsia2026] Hidden Telemetry: Uncovering TraceLogging ETW Providers You're Not Using (Yet)
asuna_jp
1
690
2025年度秋葉原ウォーカブルプロジェクト調査報告 「アキバらしいウォーカブル」とは何か
izumiyama_lab
1
200
Featured
See All Featured
Fashionably flexible responsive web design (full day workshop)
malarkey
408
67k
Visual Storytelling: How to be a Superhuman Communicator
reverentgeek
2
630
Applied NLP in the Age of Generative AI
inesmontani
PRO
4
2.4k
Bootstrapping a Software Product
garrettdimon
PRO
306
120k
Measuring & Analyzing Core Web Vitals
bluesmoon
9
990
The Cult of Friendly URLs
andyhume
79
7k
The Curious Case for Waylosing
cassininazir
1
490
"I'm Feeling Lucky" - Building Great Search Experiences for Today's Users (#IAC19)
danielanewman
230
23k
Evolution of real-time – Irina Nazarova, EuRuKo, 2024
irinanazarova
9
1.5k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
367
27k
A Guide to Academic Writing Using Generative AI - A Workshop
ks91
PRO
1
400
Reality Check: Gamification 10 Years Later
codingconduct
0
2.3k
Transcript
𝒌
◼ ⚫ ⚫ ◼ ⚫ (Zhang+, NAACL2018; Gu+, AAAI2018; Khandelwal+,
ICLR2021) ▶ (Nagao, 1984) ▶ ⚫ 𝑘 (Khandelwal+, ICLR2021) ▶ ▶ ▶ Guiding Neural Machine Translation with Retrieved Translation Pieces (Zhang+, NAACL2018) Search Engine Guided Neural Machine Translation (Gu+, AAAI2018) Nearest Neighbor Machine Translation (Khandelwal+, ICLR2021) A framework for a mechanical translation between Japanese and English by analogy principle (Nagao, 1984)
◼ ◼ ⚫ ⚫
𝒌 (Khandelwal+, ICLR2021) ◼ ⚫ ⚫ ⚫ ◼ ⚫ ▶
⚫ ▶ ≈ Nearest Neighbor Machine Translation (Khandelwal+, ICLR2021) 𝒙 𝒚
𝒌 (Khandelwal+, ICLR2021) 𝒌𝑖 ∈ ℝ𝐷 𝑓 𝒙, 𝒚<𝑡 ∈
ℝ𝐷 Nearest Neighbor Machine Translation (Khandelwal+, ICLR2021) ◼ 𝑘 ◼ ⚫ ⚫ 𝑝𝑘NN 𝑦𝑡 𝒙, 𝒚<𝑡 ∝ 𝑖=1 𝑘 𝟙𝑦𝑡=𝑣𝑖 exp − 𝒌𝑖 − 𝑓 𝒙, 𝒚<𝑡 2 2 𝜏 ◼ 𝑘
𝒌 ◼ (Martins+, EMNLP2022) ◼ (Meng+, ACLFindings2022) ⚫ 𝑘 𝑘
𝜆 = 0.5 𝑘 = 16 Chunk-based Nearest Neighbor Machine Translation (Martins+, EMNLP2022) Fast Nearest Neighbor Machine Translation (Meng+, ACL Findings2022)
𝒌 ◼ 𝑘 ◼ ⚫ 𝑘 (Matsui+, ACMMM2018) ⚫ 𝑘
𝑘 𝑘 Reconfigurable Inverted Index (Matsui+, ACMMM2018) 𝒌
◼ ⚫ 𝑘 ⚫ 𝑘 ◼ ◼ 𝑘
𝑛 𝑘 1 1 1 1 1 1 1 1
1
𝑛 𝑘 1 1 1 1 1 1 1 1
1
𝑛 𝑘 1 1 1 1 1 1 1 1
1
⚫ ⚫ ⚫ ⚫ ⚫ 𝑘 𝜆 = 0.5 𝑘
= 16 𝑛 = 56
𝑘 𝑘 ◼ 𝑘 ⚫ ▶ ⚫ ▶
◼ 𝑘 𝒌 𝒌
◼ ⚫ 𝑘
𝒌 𝒌 ◼ ⚫ ⚫ ◼ 𝑘 ⚫ ⚫ ◼
⚫
⚫ ⚫ ▶ ⚫ ▶