Tweet
Tweet

Slide 1

Slide 1 text

MultiWOZ – A Large-Scale Multi-Domain Wizard-of-Oz Dataset for Task-Oriented Dialogue Modeling Tosho Hirasawa

Slide 2

Slide 2 text

0. Overview • -6<+?E$> • 4L3I/%) H2 • :@Multi-Domain Wizard-of-Oz (MultiWOZ) • KJ • ("*72 GA/9F8 #!= 5 ,1 • 0.BD&* ' &(*;C

Slide 3

Slide 3 text

1. Introduction • Conversational Artificial Intelligence • human-level *)&($ • #%' ! • Seneff and Polifroni, 2000 • "Raux et al., 2005 • Amazon AlexaRam et al., 2018

Slide 4

Slide 4 text

1. Introduction • \T@F [C0*%0# RA • 2DKU • =W:J • ?6) 8V • OXN3A • PH517 E2E ,"/LI • <];Z17MYB(>E • &!-0Q • " 9 • [C$+0_4D • GS5'.-0^

Slide 5

Slide 5 text

1. Introduction , , 2017

Slide 6

Slide 6 text

2. Related Works • >K&.(%3/9 ! • Machine-to-Machine • *5/4+"O6K"R • HLJ-$) T DM6K\E ]X • Human-to-Machine • 7:=@^Y'(*0UZ9";I • GOE! :B • HLJ^Y'(*0 YS?,1$5&.(NI • Human-to-Human • G

Slide 7

Slide 7 text

3. Data Collection Set-up • Wizard-of-Oz E4 • Dialogue Task: • *,-@ ontology random sampling !'#%"8(6 • User Side: • (6=197CF.;A • System (Wizard) Side: • $ 2: 97/D • Wizard/User (6>, (6JG+ • (6)I30< • (6H5&?B)I30

Slide 8

Slide 8 text

3. Data Collection Set-up • Annotation of Dialogue Acts • Dialogue Act = intent + slot-value pairs • intent: inform / request • slot-value: domain, price, … • Amazon Mechanical Turk +!" &$ dialogue acts .) • !" &$'-/( • % ,*0.8843#0

Slide 9

Slide 9 text

4. MultiWOZ Dialogue Corpus • : domain

Slide 10

Slide 10 text

4. MultiWOZ Dialogue Corpus : expensive : domain

Slide 11

Slide 11 text

4. MultiWOZ Dialogue Corpus • (turns in a dialogue) • 8.93 (single-domain), 15.39 (multi-domain) • 115,434 turns • >70% 10 turns • (sentence length) • 11.75 (user), 15.12 (wizard)

Slide 12

Slide 12 text

4. MultiWOZ Dialogue Corpus • Dialogue Acts • 60% turns action • %# • "$ • %# !"$

Slide 13

Slide 13 text

4. MultiWOZ Dialogue Corpus • • • Multi-Domain, Dialogue Act

Slide 14

Slide 14 text

5. MultiWOZ as a New Benchmark • Dialogue modelling task • Dialogue State Tracking • (,# '/ • &,.5-0)1 ontology • Dialogue-Context-to-Text Generation • (,Dialogue State, # '/ • &,!16 • Cam676/MultiWOZ 28 • % $"+* • RNN 473 • Cam676: GRU • MultiWOZ: LSTM

Slide 15

Slide 15 text

5. MultiWOZ as a New Benchmark • Dialogue-Act-to-Text Generation • Structured meaning representation (Dialogue Act?) • • Semantically Conditioned LSTM (Wen+, 2015) • SFX MultiWOZ restaurant • SER = (missing slots + redundant slots) / total slots Wen+, 2015

Slide 16

Slide 16 text

6. Conclusion • )1"&7* 8 E2E #$20 • Modular-based (+%' • MultiWOZ 3 46 • !-53. github /,