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
文献紹介: Confidence Modeling for Neural Semantic P...
Search
Yumeto Inaoka
October 24, 2018
Research
3
230
文献紹介: Confidence Modeling for Neural Semantic Parsing
2018/10/24の文献紹介で発表
Yumeto Inaoka
October 24, 2018
Tweet
Share
More Decks by Yumeto Inaoka
See All by Yumeto Inaoka
文献紹介: Quantity doesn’t buy quality syntax with neural language models
yumeto
1
190
文献紹介: Open Domain Web Keyphrase Extraction Beyond Language Modeling
yumeto
0
240
文献紹介: Self-Supervised_Neural_Machine_Translation
yumeto
0
160
文献紹介: Comparing and Developing Tools to Measure the Readability of Domain-Specific Texts
yumeto
0
170
文献紹介: PAWS: Paraphrase Adversaries from Word Scrambling
yumeto
0
160
文献紹介: Beyond BLEU: Training Neural Machine Translation with Semantic Similarity
yumeto
0
280
文献紹介: EditNTS: An Neural Programmer-Interpreter Model for Sentence Simplification through Explicit Editing
yumeto
0
350
文献紹介: Decomposable Neural Paraphrase Generation
yumeto
0
230
文献紹介: Analyzing the Limitations of Cross-lingual Word Embedding Mappings
yumeto
0
230
Other Decks in Research
See All in Research
Towards a More Efficient Reasoning LLM: AIMO2 Solution Summary and Introduction to Fast-Math Models
analokmaus
2
880
MetaEarth: A Generative Foundation Model for Global-Scale Remote Sensing Image Generation
satai
4
300
長期・短期メモリを活用したエージェントの個別最適化
isidaitc
0
160
EcoWikiRS: Learning Ecological Representation of Satellite Images from Weak Supervision with Species Observation and Wikipedia
satai
3
240
Mechanistic Interpretability:解釈可能性研究の新たな潮流
koshiro_aoki
1
440
2021年度-基盤研究B-研究計画調書
trycycle
PRO
0
330
20250624_熊本経済同友会6月例会講演
trafficbrain
1
660
Creation and environmental applications of 15-year daily inundation and vegetation maps for Siberia by integrating satellite and meteorological datasets
satai
3
380
日本語新聞記事を用いた大規模言語モデルの暗記定量化 / LLMC2025
upura
0
240
20250725-bet-ai-day
cipepser
2
470
PhD Defense 2025: Visual Understanding of Human Hands in Interactions
tkhkaeio
1
240
問いを起点に、社会と共鳴する知を育む場へ
matsumoto_r
PRO
0
650
Featured
See All Featured
Into the Great Unknown - MozCon
thekraken
40
2.1k
Documentation Writing (for coders)
carmenintech
75
5k
Bash Introduction
62gerente
615
210k
We Have a Design System, Now What?
morganepeng
53
7.8k
Optimizing for Happiness
mojombo
379
70k
Typedesign – Prime Four
hannesfritz
42
2.8k
Making the Leap to Tech Lead
cromwellryan
135
9.5k
The Straight Up "How To Draw Better" Workshop
denniskardys
237
140k
Sharpening the Axe: The Primacy of Toolmaking
bcantrill
45
2.5k
Building Applications with DynamoDB
mza
96
6.6k
The Power of CSS Pseudo Elements
geoffreycrofte
79
6k
Mobile First: as difficult as doing things right
swwweet
224
10k
Transcript
Confidence Modeling for Neural Semantic Parsing จݙհɹ Ԭٕज़Պֶେֶɹࣗવݴޠॲཧݚڀࣨ ҴԬɹເਓ
Literature Confidence Modeling for Neural Semantic Parsing Li Dong† and
Chris Quirk‡ and Mirella Lapata† †School of Informatics, University of Edinburgh ‡Microsoft Research, Redmond Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Long Papers), pages 743–753, 2018. !2
Abstract • Neural Semantic Parsing (seq2seq) ʹ͓͚Δ֬৴ϞσϦϯά • ೖྗͷͲ͕͜ෆ͔֬͞ͷཁҼʹͳ͍ͬͯΔ͔Λࣝผ •
ࣄޙ֬ɺΞςϯγϣϯʹґଘ͢Δख๏ΑΓ༏ल !3
Introduction • Neural Semantic ParsingظͰ͖Δ݁ՌΛग़͢ҰํͰ ग़ྗͷݪҼ͕ղऍͮ͠Β͍ϒϥοΫϘοΫεͱͯ͠ಈ࡞ • Ϟσϧͷ༧ଌʹର͢Δ֬৴ͷਪఆʹΑͬͯ༗ҙٛͳ ϑΟʔυόοΫ͕ՄೳʹͳΔͷͰͳ͍͔ •
֬৴ͷείΞϦϯάख๏ࣄޙ֬ p(y|x) ͕Α͘༻͞ΕΔ → ઢܗϞσϧͰ༗ޮ͕ͩχϡʔϥϧϞσϧͰྑ͘ͳ͍ !4
Neural Semantic Parsing • In: Natural Language Out: Logical form
• Seq2seq with LSTM • Attention mechanism • Maximize the likelihood • Beam Search !5 !5
Confidence Estimation • ೖྗqͱ༧ଌͨ͠ҙຯදݱa͔Β֬৴s(q, a) ∈ (0, 1)Λ༧ଌ • ֬৴ͷஅʹʮԿΛΒͳ͍͔ʯΛਪఆ͢Δඞཁ͕͋Δ
• Ϟσϧͷෆ͔֬͞ɺσʔλͷෆ͔֬͞ɺೖྗͷෆ͔֬͞Λجʹ ࡞ΒΕΔࢦඪ͔Β֬৴ΛճؼϞσϧʹΑͬͯٻΊΔ !6
Model Uncertainty • ϞσϧͷύϥϝʔλߏʹΑΔෆ͔֬͞Ͱ֬৴͕Լ ← ྫ͑܇࿅σʔλʹؚ·ΕΔϊΠζ֬తֶशΞϧΰϦζϜ • Dropout Perturbation, Gaussian
Noise, Posterior Probability͔Β ࢦඪΛ࡞͠ɺෆ͔֬͞Λ༧ଌ !7
Dropout Perturbation • DropoutΛςετ࣌ʹ༻ (ਤதͷi, ii, iii, ivͷՕॴ) • จϨϕϧͰͷࢦඪɿ
• τʔΫϯϨϕϧͰͷࢦඪɿ • ɹɹઁಈͤ͞Δύϥϝʔλɹ݁ՌΛूΊͯࢄΛܭࢉ !8
Gaussian Noise • Gaussian NoiseΛϕΫτϧՃ͑ͯDropoutͱಉ༷ʹࢄΛܭࢉ ← DropoutϕϧψʔΠɺ͜ΕΨεʹै͏ϊΠζ • ϊΠζͷՃ͑ํҎԼͷ2ͭ (vݩͷϕΫτϧ,
gGaussian Noise) !9
Posterior Probability • ࣄޙ֬ p(a | q)ΛจϨϕϧͰͷࢦඪʹ༻ • τʔΫϯϨϕϧͰҎԼͷ2ͭΛࢦඪʹ༻ •
ɹɹɹɹɹɹɹɹɹɹɹɹɿ࠷ෆ͔֬ͳ୯ޠʹண • ɹɹɹɹɹɹɹɹɹɹɹɹɹɹɿτʔΫϯຖͷperplexity !10
Data Uncertainty • ܇࿅σʔλͷΧόϨοδෆ͔֬͞ʹӨڹΛ༩͑Δ • ܇࿅σʔλͰݴޠϞσϧΛֶशͤ͞ɺೖྗͷݴޠϞσϧ֬Λ ࢦඪʹ༻͍Δ • ೖྗͷະޠτʔΫϯΛࢦඪʹ༻͍Δ !11
Input Uncertainty • Ϟσϧ͕ᘳͰೖྗ͕ᐆດͩͱෆ͔֬͞ൃੜ͢Δ (e.g. 9 o’clock -> flight_time(9am) or
flight_time(9pm) ) • ্Ґީิͷ֬ͷࢄΛ༻͍Δ • ΤϯτϩϐʔΛ༻͍Δ ← a’αϯϓϦϯάۙࣅ !12
Confidence Storing • ͜ΕΒͷ༷ʑͳࢦඪΛ༻͍ͯ֬৴ͷείΞϦϯάΛߦ͏ • ޯϒʔεςΟϯάϞσϧʹ֤ࢦඪΛ༩ֶ͑ͯशͤ͞Δ ग़ྗ͕0ʙ1ʹͳΔΑ͏ϩδεςΟοΫؔͰϥοϓ • ޯϒʔεςΟϯάϞσϧҎԼͷղઆهࣄ͕͔Γ͍͢ (ʮGradient
Boosting ͱ XGBoostʯ: ɹ https://zaburo-ch.github.io/post/xgboost/ ) !13
Uncertainty Interpretation • Ͳͷೖྗ͕ෆ͔֬͞ʹ࡞༻͍ͯ͠Δ͔Λಛఆ → ͦͷೖྗΛಛผͳέʔεͱͯ͠ѻ͏͕ग़དྷΔ • ༧ଌ͔ΒೖྗτʔΫϯؒ·ͰΛٯൖ → ֤τʔΫϯͷෆ͔֬͞ͷد༩͕Θ͔Δ
!14
Experiments (Datasets) • IFTTT σʔληοτ (train-dev-test : 77,495 - 5,171
- 4,294) • DJANGO σʔληοτ (train-dev-test : 16,000 - 1,000 - 1,805) !15
Experiments (Settings) • Dropout Perturbation Dropout rate0.1ɺ30ճ࣮ߦͯ͠ࢄΛܭࢉ • Gaussian Noise
ඪ४ภࠩΛ0.05ʹઃఆ • Probability of Input ݴޠϞσϧͱͯ͠KenLMΛ༻ • Input Uncertainty 10-best ͷީิ͔ΒࢄΛܭࢉ !16
Experiments (Results) • Model Uncertainty͕࠷ޮՌత • Data UncertaintyӨڹ͕খ͍͞ → In-domainͰ͋ΔͨΊ
!17
Experiments (Results) !18
Experiments (Results) • Model Uncertaintyͷ ࢦඪ͕ॏཁ • ಛʹIFTTT#UNKͱ Var͕ॏཁ !19
Experiments (Results) !20
Experiments (Results) • ϊΠζΛՃ͑ͨτʔΫϯྻͱ ٯൖͰಘͨτʔΫϯྻͷ ΦʔόʔϥοϓͰධՁ • Attentionͱൺֱͯ͠ߴ͍ • K=4ʹ͓͍ͯ80%͕Ұக
!21
Experiments (Results) !22
Conclusions • Neural Semantic ParsingͷͨΊͷ֬৴ਪఆϞσϧΛఏࣔ • ෆ͔֬͞ΛೖྗτʔΫϯϨϕϧͰղऍ͢Δํ๏Λఏࣔ • IFTTT, DJANGOσʔληοτʹ͓͍ͯ༗ޮੑΛ֬ೝ
• ఏҊϞσϧSeq2seqΛ࠾༻͢Δ༷ʑͳλεΫͰద༻Մೳ • Neural Semantic ParsingͷActive Learningʹ͓͍ͯར༻Ͱ͖Δ !23