LAW-MWE-CxG 2018 Shared task poster boosters 1. DEEP-BGT AT PARSEME - - PowerPoint PPT Presentation

▶

Dec 01, 2022 400 likes •556 views

LAW-MWE-CxG 2018 Shared task poster boosters 1. DEEP-BGT AT PARSEME SHARED TASK 2018: BIDIRECTIONAL LSTM- CRF MODEL FOR VERBAL MULTIWORD EXPRESSION IDENTIFICATION, Gzde Berk, Berna Erden and Tunga Gungor 2. GBD-NER AT PARSEME SHARED TASK

SLIDE 1

LAW-MWE-CxG 2018

Shared task poster boosters

SLIDE 2

1. DEEP-BGT AT PARSEME SHARED TASK 2018: BIDIRECTIONAL LSTM-

CRF MODEL FOR VERBAL MULTIWORD EXPRESSION IDENTIFICATION, Gözde Berk, Berna Erden and Tunga Gungor

2. GBD-NER AT PARSEME SHARED TASK 2018: MULTI-WORD

EXPRESSION DETECTION USING BIDIRECTIONAL LONG-SHORT-TERM MEMORY NETWORKS AND GRAPH-BASED DECODING, Tiberiu Boroș and Ruxandra Burtica

3. MUMPITZ AT PARSEME SHARED TASK 2018: A BIDIRECTIONAL LSTM

FOR THE IDENTIFICATION OF VERBAL MULTIWORD EXPRESSIONS, Rafael Ehren, Timm Lichte and Younes Samih

4. CRF-SEQ and CRF-DEPTREE AT PARSEME SHARED TASK 2018::

DETECTING VERBAL MWEs USING SEQUENTIAL AND DEPENDENCY- BASED APPROACHES, Erwan Moreau, Ashjan Alsulaimani, Alfredo Maldonado and Carl Vogel

5. VARIDE AT PARSEME SHARED TASK 2018: ARE VARIANTS REALLY AS

ALIKE AS TWO PEAS IN A POD?, Caroline Pasquer, Agata Savary, Carlos Ramisch and Jean-Yves Antoine

6. VEYN AT PARSEME SHARED TASK 2018: RECURRENT NEURAL

NETWORKS FOR VMWE IDENTIFICATION, Nicolas Zampieri, Manon Scholivet, Carlos Ramisch and Benoit Favre

SLIDE 3

Deep-BGT at PARSEME Shared Task 2018: Bidirectional LSTM-CRF Model for Verbal Multiword Expression Identification

Gözde Berk, Berna Erden and Tunga Güngör

Department of Computer Engineering, Boğaziçi University, Istanbul, Turkey

SLIDE 4

Bidirectional LSTM-CRF model

Gappy 1-level tagging

Language- independent system

Sequence tagging approach

2nd in the

pen track

First Bidirectional LSTM-CRF model in VMWE identification

SLIDE 5

Multi-word Expression Detection using Bidirectional Long-Short-Term Memory Networks and graph-based decoding

SLIDE 6

SLIDE 7

Mumpitz at PARSEME Shared Task 2018: A Bidirectional LSTM for the Identification of Verbal Multiword Expressions

Rafael Ehren, Timm Lichte, Younes Samih

University of Düsseldorf, Germany

August 25, 2018

SFB 991

SLIDE 8

A BRNN with LSTMs for VMWE identification

The assumption: Bidirectional recurrent neural nets (BRNNs) are a good fit for this task, because they perform well on simi- lar sequence tagging tasks like POS tagging and named-entity recognition and a BRNN was employed in the last year’s edi- tion of the shared task. The system: Mumpitz consists of a BRNN with long short- term memory (LSTM) units and a heuristic that leverages the dependency information provided in the PARSEME corpus data to differentiate VMWEs in a sentence. The results: We submitted results for seven languages in the closed track (BG, DE, EL, ES, FR, PL, PT) and for one language in the open track (DE). For the open track we used pretrained instead of randomly initialized word embeddings to improve the performance of the system.

Ehren, Lichte, Samih (HHU) 2

SLIDE 9

A BRNN with LSTMs for VMWE identification

Come L LP D LSTM LSTM Out * in LSTM LSTM Out * and LSTM LSTM Out * have LSTM LSTM Out VMWE some LSTM LSTM Out * breakfast LSTM LSTM Out VMWE

Figure 1: Unrolled BRNN for VMWE tagging.

Ehren, Lichte, Samih (HHU) 3

SLIDE 10

CRF-Seq and CRF-DepTree at PARSEME Shared Task 2018:

Detecting Verbal MWEs Using Sequential and Dependency-Based Approaches

Erwan Moreau, Ashjan Alsulaimani, Alfredo Maldonado and Carl Vogel

Adapt Centre & Trinity College Dublin

SLIDE 11

Dependency Tree Approach

(the fancy one) ▶ Based on data analysis ▶ Target: discontinuous

expressions

▶ Uses information from the

dependency tree

▶ Complex CRF dependencies ▶ Feature selection ▶ ...

Sequential Approach

(the basic one) ▶ Sequential CRF ▶ Lemma and POS tag only ▶ Brute-force feature selection

→ 1 month work → Decent perf

▶ F-score 50.6%

→ 6 hours work (on the last day!) → Frustratingly high perf

▶ F-score 54.6%

SLIDE 12

VarIDE at PARSEME Shared Task 2018: Are variants really as alike as two peas in a pod?

Caroline Pasquer, Carlos Ramisch, Agata Savary and Jean-Yves Antoine

Problem

Identify previously seen VMWEs and their variants. Omit literal readings and incidental co-occurrences. train:

We build a bridge between countries.

test:

The bridges he built between cultures. Tower Bridge was built in 1886. The house was built near the bridge .

Method

Extract (relevant) co-occurrences of the lemmas in seen VMWEs. Classify the candidates based on their variability profile.

Naive Bayes classifier absolute features (POS, morphology, dependencies) relative features (comparison of a candidates with seens VMWEs)

Results 5th (out of 13) in the closed track 2nd for discontinuous VMWEs, 6th for variant-of-train VMWEs

2 2 1

SLIDE 13

What does this have to do with?

Veyn: recurrent neural networks for VMWE identification

Nicolas Zampieri, Carlos Ramisch, Manon Scholivet, Benoit Favre

Sequence tagger based on recurrent neural networks
Variant of begin-inside-outside + VMWE category

Jean prend de longues douches

douche NOUN

~ ~ ~ ~ ~

O B-LVC G G I-LVC

Sentence Label

long ADJ de DET prendre VERB Jean PROPN

Lemma POS Concatenation 2 recurrents layers biGRU Softmax

SLIDE 14

Why should you check out our poster?

Know the gory details about the system implementation
Study of alternative encoding schemes on dev corpus
Discuss strengths and weaknesses of the system
19 languages, ranked 9th (MWE-based) and 8th (Tok-based)
Improvements and extensions since submission

The icing on the cake

Veyn is freely available: https://github.com/zamp13/Veyn