Text Generation from Knowledge Graphs with Graph Transformers
Rik Koncel-Kedziorski
1
, Dhanush Bekal
1
, Yi Luan
1
, Mirella Lapata
2
, and Hannaneh Hajishirzi
1,3
1
University of Washington
{kedzior,dhanush,luanyi,hannaneh}@uw.edu
2
University of Edinburgh
mlap@inf.ed.ac.uk
3
Allen Institute for Artificial Intelligence
Abstract
Generating texts which express complex ideas
spanning multiple sentences requires a struc-
tured representation of their content (docu-
ment plan), but these representations are pro-
hibitively expensive to manually produce. In
this work, we address the problem of gener-
ating coherent multi-sentence texts from the
output of an information extraction system,
and in particular a knowledge graph. Graph-
ical knowledge representations are ubiquitous
in computing, but pose a significant challenge
for text generation techniques due to their
non-hierarchical nature, collapsing of long-
distance dependencies, and structural variety.
We introduce a novel graph transforming en-
coder which can leverage the relational struc-
ture of such knowledge graphs without impos-
ing linearization or hierarchical constraints.
Incorporated into an encoder-decoder setup,
we provide an end-to-end trainable system
for graph-to-text generation that we apply to
the domain of scientific text. Automatic and
human evaluations show that our technique
produces more informative texts which ex-
hibit better document structure than competi-
tive encoder-decoder methods.
1
1 Introduction
Increases in computing power and model capac-
ity have made it possible to generate mostly-
grammatical sentence-length strings of natural
language text. However, generating several sen-
tences related to a topic and which display over-
all coherence and discourse-relatedness is an open
challenge. The difficulties are compounded in do-
mains of interest such as scientific writing. Here
the variety of possible topics is great (e.g. top-
ics as diverse as driving, writing poetry, and pick-
ing stocks are all referenced in one subfield of
1
Data and code available at https://github.com/
rikdz/GraphWriter
Our Model outperforms
HMM models by 15% on
this data.
used-for
comparison
We present a CRF Model
for Event Detection.
CRF Model
Event Detection
SemEval 2011
Task 11
used
-
for
We evaluate this model
on SemEval 2010 Task 11
evaluate-for
evaluate-for
evaluate
-
for
evaluate
-
for
HMM Models
comparison
Title: Event Detection with Conditional Random Fields
Abstract
Graph
Figure 1: A scientific text showing the annotations of
an information extraction system and the correspond-
ing graphical representation. Coreference annotations
shown in color. Our model learns to generate texts from
automatically extracted knowledge using a graph en-
coder decoder setup.
one scientific discipline). Additionally, there are
strong constraints on document structure, as sci-
entific communication requires carefully ordered
explanations of processes and phenomena.
Many researchers have sought to address these
issues by working with structured inputs. Data-to-
text generation models (Konstas and Lapata, 2013;
Lebret et al., 2016; Wiseman et al., 2017; Pudup-
pully et al., 2019) condition text generation on
table-structured inputs. Tabular input representa-
tions provide more guidance for producing longer
texts, but are only available for limited domains
as they are assembled at great expense by manual
annotation processes.
The current work explores the possibility of us-
ing information extraction (IE) systems to auto-
matically provide context for generating longer
texts (Figure 1). Robust IE systems are avail-
able and have support over a large variety of tex-
tual domains, and often provide rich annotations
of relationships that extend beyond the scope of
arXiv:1904.02342v2 [cs.CL] 18 May 2019