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
Open-Retrieval Conversational Question Answering
Search
Scatter Lab Inc.
July 24, 2020
Research
0
2.3k
Open-Retrieval Conversational Question Answering
Scatter Lab Inc.
July 24, 2020
Tweet
Share
More Decks by Scatter Lab Inc.
See All by Scatter Lab Inc.
zeta introduction
scatterlab
0
1.7k
SimCLR: A Simple Framework for Contrastive Learning of Visual Representations
scatterlab
0
4k
Adversarial Filters of Dataset Biases
scatterlab
0
2.2k
Sparse, Dense, and Attentional Representations for Text Retrieval
scatterlab
0
2.3k
Weight Poisoning Attacks on Pre-trained Models
scatterlab
0
2.2k
Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval
scatterlab
0
2.5k
Beyond Accuracy: Behavioral Testing of NLP Models with CheckList
scatterlab
0
2.3k
What Can Neural Networks Reason About?
scatterlab
0
2.2k
Exploring the Limits of Transfer Learning with Unified Text-to-Text Transformer
scatterlab
0
2.2k
Other Decks in Research
See All in Research
SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery
satai
3
230
Cross-Media Information Spaces and Architectures
signer
PRO
0
230
心理言語学の視点から再考する言語モデルの学習過程
chemical_tree
2
470
NLP2025参加報告会 LT資料
hargon24
1
330
When Submarine Cables Go Dark: Examining the Web Services Resilience Amid Global Internet Disruptions
irvin
0
250
電通総研の生成AI・エージェントの取り組みエンジニアリング業務向けAI活用事例紹介
isidaitc
1
720
CSP: Self-Supervised Contrastive Spatial Pre-Training for Geospatial-Visual Representations
satai
3
220
利用シーンを意識した推薦システム〜SpotifyとAmazonの事例から〜
kuri8ive
1
220
Delta Airlines® Customer Care in the U.S.: How to Reach Them Now
bookingcomcustomersupportusa
PRO
0
100
電力システム最適化入門
mickey_kubo
1
740
[輪講] SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features
nk35jk
2
650
A scalable, annual aboveground biomass product for monitoring carbon impacts of ecosystem restoration projects
satai
3
110
Featured
See All Featured
Into the Great Unknown - MozCon
thekraken
40
1.9k
Statistics for Hackers
jakevdp
799
220k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
282
13k
ピンチをチャンスに:未来をつくるプロダクトロードマップ #pmconf2020
aki_iinuma
126
53k
Chrome DevTools: State of the Union 2024 - Debugging React & Beyond
addyosmani
7
750
Code Review Best Practice
trishagee
69
19k
Responsive Adventures: Dirty Tricks From The Dark Corners of Front-End
smashingmag
251
21k
Keith and Marios Guide to Fast Websites
keithpitt
411
22k
Product Roadmaps are Hard
iamctodd
PRO
54
11k
Making Projects Easy
brettharned
116
6.3k
The Art of Delivering Value - GDevCon NA Keynote
reverentgeek
15
1.6k
Why Our Code Smells
bkeepers
PRO
337
57k
Transcript
Open-Retrieval Conversational Question Answering ࢲ࢚ (ܻࢲ ࢎ౭झ, ೝಯ)
ѐਃ Open-Retrieval Conversational Question Answering
ѐਃ ѐਃ • SIGIR 20 • Chen Qu, Liu Yang,
Cen Chen, Minghui Qiu, W. Bruce Croft, Mohit Iyyer • University of Massachusetts Amherst, Ant Financial, Alibaba Group • Conversational searchਸ ਤ೧ ConvQAܳ open retrieval settingਵ۽ ഛೞח Ѫ ਃ োҳ ਃ
ѐਃ ѐਃ • Conversational search information retrieval Ҿӓੋ ݾী ೞա
• ୭Ӕ োҳٜ conversational searchܳ response rankingҗ conversational question answering۽ ೧Ѿ • ױࣽ ߸ਸ য candidate setীࢲ ҊܰѢա য passageীࢲ spanਸ ࢶఖ • ח conversational searchীࢲ retrieval ӝୡੋ ഝਸ ޖदೞח ߑध • ࠄ ֤ޙ open-retrieval conversational question answering(ORConvQA) settingਸ ઁউೞৈ ޙઁܳ ೧Ѿ
ѐਃ ѐਃ • ORConvQAী ೠ োҳܳ ਤ೧ OR-QuAC ؘఠ ࣇਸ
ٜ݅ਵݴ ORConvQAܳ ਤೠ end-to-end दझమਸ ҳ୷ೞݴ ےझನݠ ӝ߈ retriever, reranker ৬ reader ١ਸ ನೣ • OR-QuACܳ ࢚ਵ۽ ೠ ֤ޙ प learnable retriever ਃࢿਸ ૐݺ • ژೠ ݽٚ दझమ ҳࢿ ਃࣗ(retriever, reranker ৬ reader)ীࢲ history modelingਸ ࢎਊೞݶ दझమ ѱ ѐࢶ ؼ ࣻ ਸ ࠁ
Dataset Open-Retrieval Conversational Question Answering
ORConvQA? Dataset • conversational search systemsਸ ҳ୷ೞӝ ਤೠ ୶о ױ҅۽ࢲ
߸ਸ Ҋܰ ӝ ী retrieve evidenceܳ large collection۽ ࠗఠ Ѩ࢝ 1. ࠁܳ ҳೞח ചܳ ઁҕ(information seeker৬ information provider)৬ ೞח QuAC dataset 2. QuAC ޙਸ context-independentೞѱ द ࢿೠ CANARD dataset 3. Wikipedia passage
Dataset
CANARD? Dataset • QuAC dialogsח self-containedೞ ঋח ড חؘ ח
ࠛ৮ೠ ୡӝ ޙਵ۽ ੋ೧ ߊࢤ • ܳ ٜয seekerীѱ a Chinese polymathic scientistੋ Zhang Hengী ೧ ߓۄҊ ೮חؘ ޙ "җҗ ӝࣿҗ যڃ ҙ ҅о णפө?” • ۞ೠ ࠛౠೞҊ ݽഐೠ ୡӝ ޙ ചܳ ೧ࢳೞӝ য۵ѱ ೞӝ ٸޙী ҕѐ Ѩ࢝ ജ҃ীࢲ ޙઁܳ ঠӝ • CANARD ؘఠ ࣁীࢲ ઁҕೞח context-independent rewritesਵ۽ ೞৈ ޙઁܳ ೧Ѿ, Ӓۢ "Zhang Heng җ ӝ ࣿҗ যڃ ҙ҅о णפө?"۽ ޙ
CANARD? Dataset • ߣ૩ ޙী ೧ࢲ݅ Үܳ ࣻ೯ೞݶ ച
ղীࢲ history dependenciesਸ Ӓ۽ ਬೞݶࢲ ചо self-contained • QuAC test set ҕѐغয ঋӝ ٸޙী QuAC dev setਸ ਊೞৈ CANARD test setਸ ݅ٞ • ژೠ QuAC train set 10%ܳ dev۽ ഝਊ. • CANARDী হח QuAC ޙ ತӝ೮ਵݴ ܳ ਊೠ ࢤ ؘఠ ੋ OR-QuAC ؘఠ ా҅ח җ э.
Model Open-Retrieval Conversational Question Answering
ݽ؛ Retriever, Reranker, Reader۽ ա Model
ݽ؛ Retriever, Reranker, Reader۽ ա Model
Passage Retriever Dataset • Passage Encoder • Question Encoder •
Retrieval Score
Retrieval score ӝળਵ۽ ࢚ਤ top-Kѐ ޙࢲܳ rerank৬ reader۽ ׳ Model
ݽ؛ Retriever, Reranker, Reader۽ ա Model
Reranker& Reader Encoding Dataset • Input • Contextualized Representations •
sequence representation
Reranker& Reader Dataset • Sequence Representation • Reranker (W_rr is
vector) • Reader (span prediction)
Training Open-Retrieval Conversational Question Answering
Retriever pretraining Training • retrieval scores for the batch •
to maximize the probability of the gold passage for each question • Pretraining loss Pretraning റী passage encoderח offlineਵ۽ ك. Faissܳ ࢎਊ೧ࢲ Ѿҗܳ ࡳই১.
Concurrent Learning Training • Retriever loss • Reranker loss •
Reader loss
Inference Training • Retrieval Ѿҗ Top-K ޙࢲܳ ݽف ੋಌ۠झ ೞৈ
п ޙࢲ߹ spanਸ ஏ • Retriever loss + Reranker loss + Reader lossо ઁੌ ޙࢲ spanਸ ୭ઙ ਵ۽ ஏ
RESULTS Open-Retrieval Conversational Question Answering
Competing Method RESULTS • DrQA : TF-IDF + RNN based
reader • BERTserini : BM25 + BERT reader • ORConvQA without history : our method + window size 0 • ORConvQA : our method • Evaluation Metric : word level F1, human equivalence score (HEQ), Mean Reciprocal Rank(MRR), Recall
DrQA < BERTserini < Ours w/o hist < Ours RESULTS
Ablation study RESULTS
History windows size ઑ RESULTS
хࢎפ✌ ୶о ޙ ژח ҾӘೠ ݶ ઁٚ ইې োۅ۽
োۅ ࣁਃ! ࢲ࢚ (ܻࢲ ࢎ౭झ, ೝಯ)
[email protected]
Linked in. @pingpong