Learning Relation Entailment with Structured and Textual Information - - PowerPoint PPT Presentation

▶
learning relation entailment with structured and textual
SMART_READER_LITE
LIVE PREVIEW

Learning Relation Entailment with Structured and Textual Information - - PowerPoint PPT Presentation

Jun 19, 2023 562 likes •856 views

Learning Relation Entailment with Structured and Textual Information Zhengbao Jiang 1 , Jun Araki 2 , Donghan Yu 1 , Ruohong Zhang 1 , Wei Xu 3 , Yiming Yang 1 , Graham Neubig 1 Carnegie Mellon University 1 , Bosch Research North America 2 , Ohio

slide-1
SLIDE 1

Learning Relation Entailment with Structured and Textual Information

Zhengbao Jiang1, Jun Araki2, Donghan Yu1, Ruohong Zhang1, Wei Xu3, Yiming Yang1, Graham Neubig1

Carnegie Mellon University1, Bosch Research North America2, Ohio State University3

zhengbaj@cs.cmu.edu

1

slide-2
SLIDE 2

Motivation

  • Relations among entities play a fundamental role in knowledge graphs.

2

slide-3
SLIDE 3

Motivation

  • However, relations are treated as independent.

KG embedding: each relation is treated as an atomic unit with separate parameters.

3

slide-4
SLIDE 4

Motivation

  • However, relations are treated as independent.

KG embedding: each relation is treated as an atomic unit with separate parameters. Relation extraction: each relation is an independent class.

3

slide-5
SLIDE 5

Meta-relation: Relations Between Relations

  • Relation entailment: existence of one relation can entail the existence of another

relation.

4

slide-6
SLIDE 6

Applications of Relation Entailment

  • Knowledge graph representation learning.

5

slide-7
SLIDE 7

Applications of Relation Entailment

  • Knowledge graph representation learning.
  • Relation extraction.

6

slide-8
SLIDE 8

Applications of Relation Entailment

  • Knowledge graph representation learning.
  • Relation extraction.
  • KG-based question answering.

Q & A

7

slide-9
SLIDE 9

Relation Entailment Task Definition

  • Notations
  • Head and tail entities ℎ, 𝑢 ∈ ℰ.
  • Relations 𝑠 ∈ ℛ.
  • Instances of a relation 𝐷! =

ℎ, 𝑠, 𝑢

" ".

  • Relation entailment
  • 𝑠 ⊨ 𝑠# if and only if 𝐷! ⊆ 𝐷!!.
  • Task of predicting relation entailment
  • Given a relation 𝑠, choose its (direct) parent 𝑠# ∈ ℒ.
  • A |ℒ|-way multi-class classification problem.

8

slide-10
SLIDE 10

RelEnt Dataset

9

slide-11
SLIDE 11

RelEnt Dataset

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) (Cisco, Bosack) …

  • 1. Instances collection

9

slide-12
SLIDE 12

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) …

RelEnt Dataset

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) (Cisco, Bosack) …

  • 1. Instances collection
  • 2. Downsampling

9

slide-13
SLIDE 13

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) …

RelEnt Dataset

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) (Cisco, Bosack) …

  • 1. Instances collection
  • 2. Downsampling
  • 3. Relation expansion

9

slide-14
SLIDE 14

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) …

RelEnt Dataset

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) (Cisco, Bosack) …

  • 1. Instances collection
  • 2. Downsampling
  • 3. Relation expansion

parent Sub-relations parent organization <laboratory, university>, <airline, airline>, <record label, record label>, … architectural style <railway station, architectural style>, <church, architectural style>, ... award received <film, Academy Awards>, <human, campaign medal>, <human, scholarship>, ...

9

slide-15
SLIDE 15

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) …

RelEnt Dataset

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) (Cisco, Bosack) …

  • 1. Instances collection
  • 2. Downsampling
  • 3. Relation expansion
  • 4. Entity linking

parent Sub-relations parent organization <laboratory, university>, <airline, airline>, <record label, record label>, … architectural style <railway station, architectural style>, <church, architectural style>, ... award received <film, Academy Awards>, <human, campaign medal>, <human, scholarship>, ...

9

slide-16
SLIDE 16

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) …

RelEnt Dataset

(Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) (Cisco, Bosack) …

  • 1. Instances collection
  • 2. Downsampling
  • 3. Relation expansion
  • 5. Train/dev/test split

Train founded by ⊨ creator illustrator ⊨ creator author ⊨ creator Dev developer ⊨ creator Test designed by ⊨ creator

  • 4. Entity linking

parent Sub-relations parent organization <laboratory, university>, <airline, airline>, <record label, record label>, … architectural style <railway station, architectural style>, <church, architectural style>, ... award received <film, Academy Awards>, <human, campaign medal>, <human, scholarship>, ...

9

slide-17
SLIDE 17

Relation Representation

  • With structured information

Apple Microsoft Facebook Cisco Jobs Gates Zuckerberg Bosack founded by

  • 1. relation embedding
  • 2. entity embedding aggregation

10

slide-18
SLIDE 18

Relation Representation

  • With textual information

Apple Microsoft Facebook Cisco Jobs Gates Zuckerberg Bosack is founded by is created by founded is the founder of …

nsubjpass ←

founded agent

→ by pobj → nsubj ← is attr → the CEO prep → of pobj →

…

  • 1. words in the middle
  • 2. dependency path

11

slide-19
SLIDE 19

Relation Representation

  • Distribution-based

founded by (Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) (Cisco, Bosack) … author (Pride and Prejudice, Jane Austen) (The Lord of the Rings, J. R. R. Tolkien) (Anna Karenina, Leo Tolstoy) …

12

slide-20
SLIDE 20

Relation Representation

  • Distribution-based

founded by (Apple, Jobs) (Microsoft, Gates) (Facebook, Zuckerberg) (Cisco, Bosack) … author (Pride and Prejudice, Jane Austen) (The Lord of the Rings, J. R. R. Tolkien) (Anna Karenina, Leo Tolstoy) … Kernel density estimation with a Gaussian kernel

12

slide-21
SLIDE 21

Relation Entailment Prediction

cos( , ) 𝑠 𝑠# euc( , ) KL( , )

Unsupervised methods Supervised methods

13

slide-22
SLIDE 22

Experimental Settings

  • RelEnt Dataset
  • #Train, #Dev., #Test relations: 2055, 804, 692
  • #Classes: 498
  • Evaluation Metrics
  • Accuracy@1, Accuracy@3, and mean reciprocal rank (MRR)
  • Implementation Details
  • KG embedding methods: TransE, DistMult, ComplEx.
  • 50-dimensional GloVe embeddings.
  • BiLSTM with 64 hidden units, CNN with window size of 3 and 64 filters.

14

slide-23
SLIDE 23

Unsupervised Methods’ Results

0.327 0.506 0.491 0.572 0.1 0.2 0.3 0.4 0.5 0.6 relation emb cos entity emb dist entity emb euc entity emb cos

Acc@1 of different unsupervised methods with TransE.

15

slide-24
SLIDE 24

Unsupervised Methods’ Results

0.327 0.506 0.491 0.572 0.1 0.2 0.3 0.4 0.5 0.6 relation emb cos entity emb dist entity emb euc entity emb cos

Acc@1 of different unsupervised methods with TransE.

0.572 0.486 0.501 0.1 0.2 0.3 0.4 0.5 0.6 TransE DistMult ComplEx

Acc@1 of entity embedding with cosine using different KG representations.

15

slide-25
SLIDE 25

Supervised Methods’ Results

  • Supervised > unsupervised.
  • Textual information is complementary to structured information.

0.681 0.706 0.712 0.863 0.861 0.872 0.779 0.791 0.798 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 struct.

  • struct. + word
  • struct. + dep.

Acc@1 Acc@3 MRR 16

slide-26
SLIDE 26

Error cases

Parent Child (train) Child (test) follows has cause replaces instance of taxon rank legal form participant performer participating team (2010 Wimbledon Championships, Roger Federer) (First Continental Congress, George Washington) (Hambach Festival, Ludwig Börne) (Runaway, Linkin Park) (The Freewheelin’ Bob Dylan, Bob Dylan) (1977 UEFA Cup Final, Juventus FC) (2016–17 Premier League, Watford F.C.) (1956 Wrestling World Cup, Iran)

17

slide-27
SLIDE 27

Paper: https://openreview.net/pdf?id=ToTf_MX7Vn Code: https://github.com/jzbjyb/RelEnt

Take away

  • 1. Both structured and textual information contribute to relation entailment

prediction.

  • 2. Relation entailment prediction requires high-level abstraction.

18