[PPT] - TUPA at MRP 2019 A Multi-Task Baseline System CoNLL Shared Task 3 PowerPoint Presentation

SLIDE 1

TUPA at MRP 2019

A Multi-Task Baseline System Daniel Hershcovich1, Ofjr Arviv2

1University of Copenhagen 2Hebrew University of Jerusalem

CoNLL Shared Task 3 November 2019

1 / 9

SLIDE 2

Transition-Based UCCA Parser

After L the F break C P H , U you A should D look P at R my E poster C C A H A (Hershcovich et al., 2017)

2 / 9

SLIDE 3

Scarcity of Training Data

DM AMR UCCA

3 / 9

SLIDE 4

Multi-Task Parser

{

After L the F break C P H , U you A should D look P at R my E poster C C A H A

look-01 after break-02

p1

t i m e you ARG1 poster i poss ARG2 ARG0

After the break , you should look at my poster

ARG2 BV ARG0 ARG1 ARG1 top ARG2 ARG1 ARG2 poss}

Multi-task improves UCCA parsing (Hershcovich et al., 2018).

Private Shared After break my poster … Classifjer

4 / 9

SLIDE 5

Multi-Task Parser

{

After L the F break C P H , U you A should D look P at R my E poster C C A H A

look-01 after break-02

p1

t i m e you ARG1 poster i poss ARG2 ARG0

After the break , you should look at my poster

ARG2 BV ARG0 ARG1 ARG1 top ARG2 ARG1 ARG2 poss}

Multi-task improves UCCA parsing (Hershcovich et al., 2018).

Private Shared After break my poster … Classifjer

4 / 9

SLIDE 6

MRP

Intermediate graph representation, extended transition system.

move-01 after graduate-01

p1

time person name

p1=John

name ARG0 city name

p1=New
p2=York
p3=City

name ARG2 ARG0

⇔

ℓ-01 after ℓ-01

p1

time person name

p=ℓ

name ARG0 city name

p=ℓ

name ARG2 TOP ARG0 After graduation , John moved at New York City A n c h

r

Anchor A n c h

r

Anchor A n c h

r

Anchor A n c h

r

5 / 9

SLIDE 7

Transition Classifjer

BiLSTM + BERT (Devlin et al., 2019).

S gazed B at the… G The fox S gazed B at the… G The fox S gazed B at the… G The fox S gazed B at the… G The fox S gazed B at the… G fox

Shared BiLSTM ⊕ Private BiLSTM ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ MLP

transition

MLP

transition

MLP

transition

MLP

transition

MLP

transition

6 / 9

SLIDE 8

Transition Classifjer

BiLSTM + BERT (Devlin et al., 2019).

S gazed B at the… G The fox S gazed B at the… G The fox S gazed B at the… G The fox S gazed B at the… G The fox S gazed B at the… G fox

Shared BiLSTM ⊕ Private BiLSTM ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ ⊕ MLP

transition

MLP

transition

MLP

transition

MLP

transition

MLP

transition

6 / 9

SLIDE 9

Results

Baseline: single-task + multi-task.

Overall DM PSD EDS UCCA AMR

75.7 75.6 70.9 84.9 77.7 53.9 66 62.2 66 79.4 64.1 39 86.2 95.5 91.3 91.9 81.7 73.4 HIT-SCIR SJTU Saarland SUDA–Alibaba HIT-SCIR Amazon Full Evaluation MRP F-score (%). TUPA scores are post-evaluation. TUPA Single-Task TUPA Multi-Task Best System 7 / 9

SLIDE 10

Results

Baseline: single-task + multi-task.

Overall DM PSD EDS UCCA AMR

75.7 75.6 70.9 84.9 77.7 53.9 66 62.2 66 79.4 64.1 39 86.2 95.5 91.3 91.9 81.7 73.4 HIT-SCIR SJTU Saarland SUDA–Alibaba HIT-SCIR Amazon Full Evaluation MRP F-score (%). TUPA scores are post-evaluation. TUPA Single-Task TUPA Multi-Task Best System 7 / 9

SLIDE 11

TUPA at MRP 2019

A Multi-Task Baseline System Daniel Hershcovich1, Ofjr Arviv2

1University of Copenhagen 2Hebrew University of Jerusalem

CoNLL Shared Task 3 November 2019

8 / 9

SLIDE 12

References

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina

Toutanova. 2019. BERT: Pre-training of deep

bidirectional transformers for language understanding. In

Proc. of NAACL, pages 4171–4186.

Daniel Hershcovich, Omri Abend, and Ari Rappoport. 2017. A transition-based directed acyclic graph parser for UCCA. In Proc. of ACL, pages 1127–1138. Daniel Hershcovich, Omri Abend, and Ari Rappoport. 2018. Multitask parsing across semantic representations. In

Proc. of ACL, pages 373–385.

9 / 9