Te Text Generation from Kn Knowledge Graphs with Gr Graph - - PowerPoint PPT Presentation

▶

Feb 28, 2023 293 likes •538 views

Te Text Generation from Kn Knowledge Graphs with Gr Graph Transforme rmers NAACL19 Rik Koncel-Kedziorski , Dhanush Bekal , Yi Luan , Mirella Lapata , and Hannaneh Hajishirzi University of Washington University of Edinburgh Allen

SLIDE 1

Te Text Generation from Kn Knowledge Graphs with Gr Graph Transforme rmers

NAACL19 Rik Koncel-Kedziorski , Dhanush Bekal , Yi Luan , Mirella Lapata , and Hannaneh Hajishirzi University of Washington University of Edinburgh Allen Institute for Artificial Intelligence

Reporter：Xiachong Feng https://www.youtube.com/watch?v=BiRyvB2NmCM

SLIDE 2

Outline

Author
Motivation
Task
Dataset
Model
Experiments
Conclusion

SLIDE 3

Author

Rik Koncel-Kedziorski
Lives on a sailboat
University of Washington Ph.D. Winter 2019

SLIDE 4

Knowledge

SLIDE 5

Knowledge

SLIDE 6

Task

Input
Title of a scientific article;
Knowledge graph constructed by an automatic

information extraction system;

Output
Abstract (text);

Graph T i t l e

SLIDE 7

Dataset

Abstract GENeration DAtaset (AGENDA) Dataset
12 top AI conferences
SciIE system : a state-of-the-art science domain

information extraction system.

NER、Co-Reference、Relations

SLIDE 8

Dataset

SLIDE 9

Model-GraphWriter

Encoder Decoder

SLIDE 10

Graph Preparation

disconnected labeled graph connected unlabeled graph

SLIDE 11

Embedding Vertices, Encoding Title

Relation：forward- and backward-looking, two

embeddings per relation

Entities correspond to scientific terms which are
ften multi-word expressions.
Bidirectional RNN run over embeddings of each

word

The title input is also a short string, and so we

encode it with another BiRNN

SLIDE 12

Graph Transformer

SLIDE 13

GAT

SLIDE 14

Graph Attention

concat

SLIDE 15

Block networks

global contextualization

SLIDE 16

Decoder

At each decoding timestep t we use decoder

hidden state ht to compute context vectors cg and cs for the graph and title sequence

SLIDE 17

Copy

entities

SLIDE 18

Experiments

Evaluation Metrics
Human evaluation
Grammar
Fluency
Coherence
Informativeness
Automatic metrics
BLEU
METEOR

SLIDE 19

Baselines

GAT : PReLU activations stacked between 6 self-

attention layers.

EntityWriter : uses only entities and title (no graph)
Rewriter : uses only the document title

SLIDE 20

Does Knowledge Help?

SLIDE 21

Conclusion

Propose a new graph transformer encoder that

applies the successful sequence transformer to graph structured inputs.

Provide a large dataset of knowledge graphs paired

with scientific texts for further study.

SLIDE 22

Te Text Generation from Kn Knowledge Graphs with Gr Graph Transforme rmers

Thanks!