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
20250226 NLP colloquium: "SoftMatcha: 10億単語規模コー...
Search
Hiroyuki Deguchi
February 26, 2025
Research
840
1
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
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"
Hiroyuki Deguchi
February 26, 2025
More Decks by Hiroyuki Deguchi
See All by Hiroyuki Deguchi
20260427 Ehime DS Seminar: MBR-CBDT Decoding
de9uch1
1
21
20260624 NLP colloquium: 単一のhubテキストがCLIPを壊す:hubnessによる埋め込みの脆弱性特定
de9uch1
2
230
20240820: Minimum Bayes Risk Decoding for High-Quality Text Generation Beyond High-Probability Text
de9uch1
0
370
サブセット探索を用いた高速なkNNニューラル機械翻訳
de9uch1
0
180
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
某助成金プロジェクト採択に向けて企業研究所のアウトリーチ専任者がやったこと
afroscript
0
180
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
160
コーディングエージェントとABNを再考
hf149
2
860
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
720
The story of RefactoringMiner. Slow research, long-term impact
tsantalis
0
140
Cross-Media Human-Information Interaction
signer
PRO
0
220
COMETAを用いたデータ民主化運動の歴史
sazimai
0
240
2026年版中小企業白書・小規模企業白書の概要
ozekinote
0
210
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
[Fishers] DIVER OSINT CTF 2026 特化AIエージェントハーネスで挑戦するOSINT CTF
analokmaus
0
550
Using our influence and power for patient safety
helenbevan
0
410
Ghost in the 7‑Zip: The Shadow of Residential Proxies Creeping into Your Life
nttcom
0
2k
Featured
See All Featured
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
141
35k
Why You Should Never Use an ORM
jnunemaker
PRO
61
10k
Technical Leadership for Architectural Decision Making
baasie
3
550
We Are The Robots
honzajavorek
0
330
A Modern Web Designer's Workflow
chriscoyier
698
190k
Fantastic passwords and where to find them - at NoRuKo
philnash
52
3.8k
Ruling the World: When Life Gets Gamed
codingconduct
0
320
Art, The Web, and Tiny UX
lynnandtonic
304
22k
Rails Girls Zürich Keynote
gr2m
96
14k
Crafting Experiences
bethany
1
280
Visualization
eitanlees
152
17k
Stop Working from a Prison Cell
hatefulcrawdad
274
21k
Transcript
None
◼ ⚫ ⚫ ⚫ ⚫ ⚫ ◼ ⚫ ⚫ ⚫
◼ ◼ ◼ ◼
◼ ◼ ◼ ◼
◼ ◼ ◼ ◼
𝑤 ◼ 𝑤 ◼ ⚫ 𝑤 ⚫ 𝑤 ◼ ⚫
◼ ⚫ ◼ ⚫ ◼ ⚫ (Radovanovic+, JMLR2010) ⚫ ▶
Wang+, (arxiv) 2022, “Text Embeddings by Weakly-Supervised Contrastive Pre-training”. Radovanovic+, JMLR 2010, “Hubs in space: Popular nearest neighbors in high-dimensional data”.
◼ ◼ ◼ ◼
◼ ⚫ 𝐩 = 𝑝1 , … , 𝑝𝑀 ∈
Σ∗ ⚫ 𝐭 = 𝑡1 , … , 𝑡𝑁 ∈ Σ∗ ▶ Σ∗ ◼ ⚫ ⚫ ◼ ◼
◼ ⚫ 𝑤 ∈ 𝒱 𝐷 ⚫ 𝐯𝑤 ∈ ℝ𝐷
≔ 𝑤 ⚫ ▶ cos 𝐯person , 𝐯people > cos 𝐯person , 𝐯bird ▶
◼ ◼ ⚫ ※
◼ ⚫ 𝑡𝑖 = 𝑝𝑗 cos 𝐯𝑡𝑖 , 𝐯𝑝𝑗 ≥
𝛼 ▶ 𝛼 = 1.0 ※ 𝛼 = 0.7 ◼ ⚫ ⚫
𝑡1 𝑡2 𝑡3 𝑡4 𝑡5 𝑡6 𝑡7 𝑡8 𝑡9 𝑡10
𝑡11 𝑡12 𝑡13 ◼ ◼ ◼ ⚫
◼ 𝒮𝑤 ≔ 𝑣 ∈ 𝒱 cos 𝐯𝑣 ⊤𝐯𝑤 ≥
𝛼 ⚫ 𝑤 𝒮we 𝒮talk 𝒮about
⇔ 𝒮we , 𝒮talk , 𝒮about ⇔ 𝒮we , 𝒮talk
, 𝒮about 𝑖, 𝑖 + 1, 𝑖 + 2 𝑖 𝒮we 𝒮talk 𝒮about
𝒮we 𝒮talk 𝒮about 𝒮we ℳ ℳ ← 1,10,6 𝒮talk ℳ
ℳ′ ← 2 − 1, 11 − 1,7 − 1 = 1,10,6 ℳ ← ℳ ∩ ℳ′ = 1,10,6 𝒮about ℳ ℳ′ ← 8 − 2, 12 − 2 = 6,10 ℳ ← ℳ ∩ ℳ′ = 6,10 ℳ
𝒮we 𝒮talk 𝒮about 𝑡1 𝑡2 𝑡3 𝑡4 𝑡5 𝑡6 𝑡7
𝑡8 𝑡9 𝑡10 𝑡11 𝑡12 𝑡13
𝒮𝑝1 𝐼𝒮𝑝1 𝒮𝑝𝑀 𝐼𝒮𝑝𝑀 𝒮𝑝1 𝐼𝒮𝑝1 ℳ ℳ ← 𝐼𝒮𝑝1
𝑘 = 2, … , 𝑀 ℳ′ ← 𝑖 − 𝑘 + 1 𝑖 ∈ 𝐼𝒮𝑝𝑘 ℳ ← ℳ ∩ ℳ′ ℳ 𝐩 = 𝑝1 , … , 𝑝𝑀
◼ ⚫ ▶ ▶ ⚫
◼ ◼ ⚫ (Wang+, 2024) ⚫ (Douze+, 2024) (Malkov &
Yashunin, IEEE TPAMI, 2018) ◼ ⚫ ▶ 𝛼 = 0.55 (Pennington+, EMNLP2014) ▶ 𝛼 = 0.50 (Grave+, arXiv:1802.06893) Wang+, arXiv:2402.05672, “Multilingual E5 Text Embeddings: A Technical Report”. Douze+, arXiv:2401.08281, “The Faiss library”. Malkov & Yashunin, IEEE TPAMI, 2018, “Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs”. Pennington+, EMNLP2014, “GloVe: Global Vectors for Word Representation”. Grave+, arXiv:1802.06893, “Learning Word Vectors for 157 Languages”.
◼ ⚫ ▶ ⚫
◼ ⚫ ⚫ ◼ ⚫
◼ ⚫ (Crane, IJDL 2023) ⚫ (Bothwell+, EMNLP2023) ◼ Crane,
IJDL 2023, “The Perseus Digital Library and the future of libraries.”. Bothwell+, EMNLP2023, “Introducing Rhetorical Parallelism Detection: A New Task with Datasets, Metrics, and Baselines”.
◼ ⚫ ⚫ ⚫ ◼ ◼
◼ ⚫ ◼ ⚫ ▶ ◼ ⚫ ▶ ▶ ⚫
⚫
◼ ⚫ 𝐼𝒮𝑝𝑘 ▶ ⚫ ◼ ⚫
◼ ◼ ⚫ ⚫ ⚫ 𝑂 1 ▶ ◼ ⚫
⚫ ⚫ 𝑂 log |𝐵| ▶
◼ ⚫ ⚫ ◼ ⚫ ⚫ ▶ ⚫