Content uploaded by Michele Bevilacqua
Author content
All content in this area was uploaded by Michele Bevilacqua on Jan 07, 2021
Content may be subject to copyright.
One SPRING to Rule Them Both:
Symmetric AMR Semantic Parsing and Generation without a Complex Pipeline
Michele Bevilacqua Rexhina Blloshmi Roberto Navigli
Sapienza NLP Group
Department of Computer Science, Sapienza University of Rome
{bevilacqua,blloshmi,navigli}@di.uniroma1.it
Abstract
In Text-to-AMR parsing, current state-of-the-art semantic
parsers use cumbersome pipelines integrating several differ-
ent modules or components, and exploit graph recategoriza-
tion, i.e., a set of content-specific heuristics that are devel-
oped on the basis of the training set. However, the general-
izability of graph recategorization in an out-of-distribution
setting is unclear. In contrast, state-of-the-art AMR-to-Text
generation, which can be seen as the inverse to parsing, is
based on simpler seq2seq. In this paper, we cast Text-to-
AMR and AMR-to-Text as a symmetric transduction task and
show that by devising a careful graph linearization and ex-
tending a pretrained encoder-decoder model, it is possible
to obtain state-of-the-art performances in both tasks using
the very same seq2seq approach, i.e., SPRING (Symmetric
PaRsIng aNd Generation). Our model does not require com-
plex pipelines, nor heuristics built on heavy assumptions. In
fact, we drop the need for graph recategorization, showing
that this technique is actually harmful outside of the stan-
dard benchmark. Finally, we outperform the previous state
of the art on the English AMR 2.0 dataset by a large margin:
on Text-to-AMR we obtain an improvement of 3.6Smatch
points, while on AMR-to-Text we outperform the state of
the art by 11.2BLEU points. We release the software at
github.com/SapienzaNLP/spring.
1 Introduction
In recent years Abstract Meaning Representation (Ba-
narescu et al. 2013, AMR) has become an influential formal-
ism for capturing the meaning of a given utterance within
a semantic graph (parsing) and, vice versa, producing text
from such a graph (generation). AMR’s flexibility has re-
sulted in promising improvements in Machine Translation
(Song et al. 2019), Text Summarization (Hardy and Vlachos
2018; Liao, Lebanoff, and Liu 2018), Human-Robot Inter-
action (Bonial et al. 2020) and Information Extraction (Rao
et al. 2017). Moreover, AMR-to-Text and Text-to-AMR sys-
tems represent an effective bridge between natural language
and symbolic representations (which can be manipulated by
both humans and computer programs), thus providing a step
towards the decoupling of content planning – what to say –
from language competence – how to say it.
Copyright c
2021, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Recent state-of-the-art approaches to Text-to-AMR se-
mantic parsing feature very complex pre- and postpro-
cessing pipelines, in which the output of several different
components is integrated. Additionally, they employ fine-
grained, content-specific heuristics developed on the basis
of the training set that, as a consequence, can be very brit-
tle across domains and genres. The parsing performance of
simpler, full Sequence-to-Sequence (seq2seq) methods has
hitherto lagged behind, mainly because they are less data-
efficient than their alternatives.
When it comes to AMR-to-Text generation, which can
be seen as the inverse task to Text-to-AMR parsing, vanilla
seq2seq methods have, instead, achieved state-of-the-art re-
sults. This architectural asymmetry is not observed in other
bidirectional transduction tasks such as machine translation,
where the same architecture is used to handle the translation
from language Xto language Y, and vice versa. Motivated
by this, a key goal of this paper is to achieve symmetry in
AMR parsing and generation as well, by providing the same
architecture for both. Moreover, we reduce the complexity
of Text-to-AMR architectures by disposing of the need for
content-modifying pipelines and additional syntactic and se-
mantic features, which often depend on external components
and data-specific heuristics. We achieve this by linearizing
the AMR graph efficiently and by extending a pretrained
seq2seq model, i.e., BART (Lewis et al. 2020), to handle
both AMR-to-Text and Text-to-AMR. In fact, the only ex-
ternal resource consistently beneficial for our model is an
off-the-shelf system for Entity Linking – a task that is hard
to perform robustly with pure seq2seq models.
Our contributions are the following:
1. We extend a pretrained Transformer encoder-decoder ar-
chitecture to generate either an accurate linearization of
the AMR graph for a sentence or, vice versa, a sentence
for a linearization of the AMR graph.
2. Contrary to previous reports (Konstas et al. 2017), we find
that the choice between competing graph-isomorphic lin-
earizations does matter. Our proposed Depth-First Search
(DFS)-based linearization with special pointer tokens out-
performs both the PENMAN linearization and an analo-
gous Breadth-First Search (BFS)-based alternative, espe-
cially on AMR-to-Text.
3. We propose a novel Out-of-Distribution (OOD) setting for
estimating the ability of the Text-to-AMR and AMR-to-
Text approaches to generalize on open-world data.
4. We show that graph recategorization should be avoided
on open-world data because, although it slightly boosts
the performance in the standard benchmark, it is not able
to generalize in the OOD setting.
5. We outperform the previously best reported results in
AMR 2.0 by 11.2BLEU points for the generation task,
and by 3.6Smatch points for the parsing task.
2 Related Work
Our work is concerned with Text-to-AMR parsing, AMR-
to-Text generation, and with how to use pretrained seq2seq
models to handle both of these tasks.
2.1 Text-to-AMR Parsing
Pure seq2seq Seq2seq approaches model Text-to-AMR
parsing as a transduction of the sentence into a lineariza-
tion of the AMR graph. Due to their end-to-end nature,
such approaches are appealing for this task. However, since
seq2seq-based approaches are data-hungry, their perfor-
mances for AMR parsing have, until now, been rather un-
satisfactory, due to the relatively small amount of annotated
sentence-AMR pairs. To overcome data sparsity, various dif-
ferent techniques have been employed: self-training using
unlabeled English text (Konstas et al. 2017), character-level
networks (van Noord and Bos 2017), and concept recatego-
rization as a preprocessing step to reduce the open vocabu-
lary components, e.g., named entities and dates (Peng et al.
2017; van Noord and Bos 2017; Konstas et al. 2017). More-
over, seq2seq-based models often incorporate features such
as lemma, POS, or Named Entity Recognition (NER) tags,
as well as syntactic and semantic structures (Ge et al. 2019).
To counteract sparsity, we employ transfer learning by ex-
ploiting BART (Lewis et al. 2020) – a recently-released pre-
trained encoder-decoder – to generate a linearized graph in-
crementally with a single auto-regressive pass of a seq2seq
decoder. In fact, the base Transformer encoder-decoder of
BART is similar to that of Ge et al. (2019), which differs,
however, in that it trains the AMR parsing architecture from
scratch.
Hybrid approaches State-of-the-art results in Text-to-
AMR have been attained by approaches that use more com-
plex multi-modular architectures. These combine seq2seq
methods with graph-based algorithms in either two-stage
(Zhang et al. 2019a) or incremental one-stage (Zhang et al.
2019b; Cai and Lam 2020a) procedures. Moreover, they in-
tegrate similar processing pipelines and additional features
such as the above-mentioned seq2seq approaches (Kon-
stas et al. 2017), including fine-grained graph recatego-
rization (Zhang et al. 2019a,b; Zhou et al. 2020; Cai and
Lam 2020a), which all contribute significantly to the perfor-
mances achieved.
In contrast, our model relies almost exclusively on
seq2seq, does not need extra features, and employs a bare-
bone postprocessing pipeline only for ensuring graph valid-
ity. Nonetheless, we significantly outperform previous state-
of-the-art approaches. Additionally, we show that the ex-
tensive recategorization techniques, while boosting perfor-
mance on the traditional in-domain benchmarks, are harmful
in the OOD setting. Moreover, while other approaches have
employed pretrained encoders, such as BERT (Devlin et al.
2019), in order to have powerful features for a parsing archi-
tecture (Zhang et al. 2019a,b; Cai and Lam 2020a), we are
the first to show that pretrained decoders, too, are beneficial
for AMR parsing, even though the pretraining only involves
English, and does not include formal representations.
2.2 AMR-to-Text Generation
AMR-to-Text generation is currently performed with two
main approaches: explicitly encoding the graph structure in
a graph-to-text transduction fashion (Song et al. 2018; Beck,
Haffari, and Cohn 2018; Damonte and Cohen 2019; Zhu
et al. 2019; Cai and Lam 2020b; Yao, Wang, and Wan 2020),
or as a purely seq2seq task through AMR graph linearization
(Konstas et al. 2017; Mager et al. 2020). Recent graph-based
approaches rely on Transformers to encode AMR graphs
(Zhu et al. 2019; Cai and Lam 2020b; Wang, Wan, and Yao
2020; Song et al. 2020; Yao, Wang, and Wan 2020). The
model of Mager et al. (2020) is a pretrained Transformer-
based decoder-only model fine-tuned on a sequential repre-
sentation of the AMR graph. Instead, we use an encoder-
decoder architecture, which is more suitable for handling
conditional generation and casts AMR-to-Text as symmet-
ric to Text-to-AMR, therefore disposing of the need for a
task-specific model.
2.3 Linearization Information Loss
Previous approaches to Text-to-AMR parsing (Konstas et al.
2017; van Noord and Bos 2017; Peng et al. 2017; Ge et al.
2019) use seq2seq methods in conjunction with lossy lin-
earization techniques, which, in order to reduce complexity,
remove information such as variables from the graph. This
information is restored heuristically, making it harder to pro-
duce certain valid outputs. In contrast, we propose two lin-
earization techniques which are completely isomorphic to
the graph, and do not incur any information loss.
2.4 BART
BART is a Transformer-based encoder-decoder model
which is pretrained through a denoising self-supervised task,
i.e., reconstructing an English text which has been modified
through shuffling, sentence permutation, masking and other
kinds of corruption (Lewis et al. 2020). BART has shown
significant improvements in conditioned generation tasks
where the vocabulary of the input and output sequences
largely intersect, such as question answering and summa-
rization. Similarly, a large amount of AMR labels are drawn
from the English vocabulary – despite the fact that AMR
aims to abstract away from the sentence – and, therefore,
we hypothesize that BART’s denoising pretraining should be
suitable for AMR-to-Text and Text-to-AMR as well. More-
over, it is possible to see a parallel between BART’s pretrain-
ing task and AMR-to-Text generation, since the linearized
AMR graph can be seen as a reordered, partially corrupted
version of an English sentence, which the model has to re-
construct.
3 Method
We perform both Text-to-AMR parsing and AMR-to-Text
generation with the same architecture, i.e., SPRING, which
exploits the transfer learning capabilities of BART for the
two tasks. In SPRING AMR graphs are handled symmet-
rically: for Text-to-AMR parsing the encoder-decoder is
trained to predict a graph given a sentence; for AMR-to-
Text generation another specular encoder-decoder is trained
to predict a sentence given a graph.
In order to use the graphs within the seq2seq model,
we transform them into a sequence of symbols using vari-
ous different linearization techniques (Section 3.1). Further-
more, we modify the BART vocabulary in order to make it
suitable for AMR concepts, frames and relations (Section
3.2). Finally, we define lightweight, non content-modifying
heuristics to deal with the fact that, in parsing, seq2seq may
output strings which cannot be decoded into a graph (Sec-
tion 3.3).
3.1 Graph Linearizations
In this work we use linearization techniques which are fully
graph-isomorphic, i.e., it is possible to encode the graph
into a sequence of symbols and then decode it back into a
graph without losing adjacency information. We propose the
use of special tokens <R0>,<R1>,. . . ,<Rn>to represent
variables in the linearized graph and to handle co-referring
nodes. Just as happens with variable names in PENMAN,
i.e., the encoding that is used in the release files of AMR,
whenever such special tokens occur more than once it is
signaled in our encoding that a given node fulfills multiple
roles in the graph. By means of this modification we aim to
address the confusion arising from the use of seq2seq with
PENMAN (PM), which does not allow a clear distinction to
be made between constants and variables, as variable names
have no semantics. Our special tokens approach is used in
combination with two graph traversal techniques based on,
respectively, DFS and BFS; in addition, we also experiment
with PENMAN. In Figure 1 we show the linearizations of
the AMR graph for “You told me to wash the dog”.
DFS-based DFS, on which PENMAN is based, is very at-
tractive as it is quite closely related to the way natural lan-
guage syntactic trees are linearized: consider, e.g., the sen-
tence “the dog which ate the bone which my father found is
sleeping”, where the noun dog is far removed from its head
verb, is sleeping, because the dependents of dog are “ex-
plored” completely before the occurrence of the head verb.
Thus, we employ a DFS-based linearization with special
tokens to indicate variables and parentheses to mark visit
depth. Moreover, we dispose of the redundant slash token
(/). These features significantly reduce the length of the out-
put sequence compared to PENMAN, where variable names
are often split into multiple subtokens by the subword tok-
enizer. This is important for efficient seq2seq decoding with
Transformers, which are bottlenecked by the quadratic com-
plexity of attention mechanisms.
tell-01
you wash-01
I dog
:ARG0 :ARG1
:ARG2
:ARG0 :ARG1
PM ( t / tell-01 :ARG0 ( y / you ) :ARG1 (
w / wash-01 :ARG0 i :ARG1 ( d / dog ) )
:ARG2(i/i))
DFS ( <R0> tell-01 :ARG0 ( <R1> you ) :ARG1
( <R3> wash-01 :ARG0 <R2> :ARG1 ( <R4> dog
) ) :ARG2 ( <R2> i ) )
BFS <R0> tell-01 :ARG0 <R1> you :ARG1 <R3>
wash-01 :ARG2 <R2> i <stop> <R3> :ARG0
<R2> :ARG1 <R4> dog <stop>
Figure 1: The AMR graph for the sentence “You told me to
wash the dog.” with the three different linearizations.
BFS-based The use of BFS traversal is motivated by the
fact that it enforces a locality principle by which things be-
longing together are close to each other in the flat represen-
tation. Additionally, Cai and Lam (2019) suggest that BFS
is cognitively attractive because it corresponds to a core-
semantic principle which assumes that the most important
pieces of meaning are represented in the upper layers of
the graph. To this end, we present a BFS-based lineariza-
tion which, just like our DFS-based one, uses special tokens
to represent co-reference. We apply a BFS graph traversal
algorithm which starts from the graph root rand visits all
the children wconnected by an edge e, appending to the lin-
earization the pointer token to r,e, and then a pointer token
if wis a variable, or its value in case wis a constant. The
first time a pointer token is appended, we also append its
:instance attribute. At the end of the iteration at each
level, i.e., after visiting the children w, we append a special
<stop> token to signal the end node exploration. In Figure
1, the visit starts with tell-01, iterates over its children,
then, after the <stop>, goes on to wash-01.
Edge ordering All the above linearizations are decoded
into the same graph. However, in the PENMAN-linearized
gold annotations, an edge ordering can be extracted from
each AMR graph. There has been a suggestion (Konstas
et al. 2017) that annotators have used this possibility to en-
code information about argument ordering in the source sen-
tence. Our preliminary experiments confirmed that imposing
an edge ordering different from PENMAN has a big negative
effect on the evaluation measures of AMR-to-Text genera-
tion, due to their order-sensitive nature. To control this, we
have carefully designed the linearizations to preserve order
information.
3.2 Vocabulary
BART uses a subword vocabulary and its tokenization is
optimized to handle English, but it is not well-suited for
AMR symbols. To deal with this problem we expand the
tokenization vocabulary of BART by adding i) all the rela-
tions and frames occurring at least 5times in the training
corpus; ii) constituents of AMR tokens, such as :op; iii)
the special tokens that are needed for the various graph lin-
earizations. Moreover, we adjust the embedding matrices of
encoder and decoder to include the new symbols by adding a
vector which is initialized as the average of the subword con-
stituents. The addition of AMR-specific symbols in vocabu-
lary expansion avoids extensive subtoken splitting and thus
allows the encoding of AMRs as a more compact sequence
of symbols, cutting decoding space and time requirements.
Recategorization Recategorization is a popular technique
to shrink the vocabulary size for handling data sparsity. It
simplifies the graph by removing sense nodes, wiki links,
polarity attributes, and/or by anonymizing the named enti-
ties. To assess the contribution of recategorization, we exper-
iment with a commonly-used method in AMR parsing liter-
ature (Zhang et al. 2019a,b; Zhou et al. 2020; Cai and Lam
2020a). The method is based on string-matching heuris-
tics and mappings tailored to the training data, which also
regulate the restoration process at inference time. We di-
rect the reader to Zhang et al. (2019a) for further details.
We note that following common practice we use recatego-
rization techniques only in parsing, due to the considerably
higher information loss that could result in generation.
3.3 Postprocessing
In our approach we perform light postprocessing, mainly to
ensure the validity of the graph produced in parsing. To this
end, we restore parenthesis parity in PENMAN and DFS,
and also remove any token which is not a possible continua-
tion given the token that precedes it. For BFS, we recover a
valid set of triples between each subsequent pair of <stop>
tokens. Our approaches remove content limited to a few to-
kens, often repetitions or hallucinations. We notice that non-
recoverable graphs are very rare, roughly lower than 0.02%
in out-of-distribution data, with a negligible effect on over-
all performance. In addition, we integrate an external Entity
Linker to handle wikification, because it is difficult to han-
dle the edge cases with pure seq2seq. We use a simple string
matching approach to search for a mention in the input sen-
tence for each :wiki attribute that SPRING predicted in
the graph, then run the off-the-shelf BLINK Entity Linker
(Wu et al. 2020) and overwrite the prediction.
4 Experimental Setup
We now describe the setup of the experiments that we per-
form to evaluate SPRING in both Text-to-AMR parsing and
AMR-to-Text generation.
4.1 Datasets
In-Distribution We evaluate the strength of SPRING on
the standard evaluation benchmarks, which we refer to as
the In-Distribution (ID) setting. The data that we use in
this setting are the AMR 2.0 (LDC2017T10) and AMR
3.0 (LDC2020T02) corpora releases, which include, respec-
tively 39,260 and 59,255 manually-created sentence-AMR
pairs. AMR 3.0 is a superset of AMR 2.0. In both of them
the training, development and test sets are a random split
of a single dataset, therefore they are drawn from the same
distribution.
Out-of-Distribution While the ID setting enables a com-
parison against previous literature, it does not allow esti-
mates to be made about performances on open-world data,
which will likely come from a different distribution of that
of the training set. Motivated by common practice in related
semantic tasks, such as Semantic Role Labeling (Hajiˇ
c et al.
2009), we propose a novel OOD setting.
In this evaluation setting we assess the performance of
SPRING when trained on OOD data, contrasting it with
the ID results. We employ the AMR 2.0 training set, while
for testing we use three distinct Out-of-Distribution (OOD)
benchmarks, covering a variety of different genres: i) New3,
a set of 527 instances from AMR 3.0, whose original source
was the LORELEI DARPA project – not included in the
AMR 2.0 training set – consisting of excerpts from newswire
and online forums; ii) TLP, the full AMR-tagged children’s
novel The Little Prince (ver. 3.0), consisting of 1,562 pairs;
iii) Bio, i.e., the test set of the Bio-AMR corpus, consist-
ing of 500 instances, featuring biomedical texts (May and
Priyadarshi 2017).
Silver In order to determine whether silver-data augmen-
tation, another commonly used technique, is beneficial in
both ID and OOD, we follow Konstas et al. (2017) and create
pretraining data by running the SPRING parser using DFS
(trained on AMR 2.0) on a random sample of the Gigaword
(LDC2011T07) corpus consisting of 200,000 sentences.
4.2 Models
SPRING relies on BART with the augmented vocabulary,
as discussed in Section 3.2. We use the same model hy-
perparameters as BART Large (or Base, when specified), as
defined in Huggingface’s transformers library. Models
are trained for 30 epochs using cross-entropy with a batch
size of 500 graph linearization tokens, with RAdam (Liu
et al. 2020) optimizer and a learning rate of 1×10−5. Gra-
dient is accumulated for 10 batches. Dropout is set to 0.25.
Hyperparameter search We report in Table 1 the final
hyperparameters used to train and evaluate both the Text-to-
AMR and AMR-to-Text models. To pick these parameters,
we used random search with about 25 Text-to-AMR trials
in the search space indicated in the third column. Text-to-
AMR training requires about 22 and 30 hours on AMR 2.0
and AMR 3.0 using one 1080 Ti GPU, respectively; AMR-
to-Text requires 13 and 16.5hours on AMR 2.0 and AMR
3.0, respectively. At prediction time, we set beam size to
5following common practice in neural machine translation
(Yang, Huang, and Ma 2018).
SPRING variants We include models trained with the
three linearizations, indicated as SPRING[lin], where [lin]
is one of the linearizations: PENMAN (PM), DFS- (DFS)
or BFS-based (BFS). In addition, we include variants of
SPRINGDFS using i) BART Base (base); ii) graph recatego-
rization (+recat); iii) pretrained silver AMR data (+silver).
Parameter Pick Search Space
LR 5∗10−51/5/10/50 ∗10−5
Betas 0.9, 0.999 -
Dropout 0.25 0.1 to 0.25, (+0.05)
W. Decay 0.004 0.001 to 0.01, (+0.001)
LR sched. constant -
Grad. accum. 10 1/5/10/15/20
Beam size 5 [1,5]
Table 1: Hyperparameters and search space.
BART baseline We also report results on a vanilla BART
baseline which treats PENMAN as a string, uses no vocabu-
lary expansion and tokenizes the graph accordingly.
4.3 Comparison Systems
In-Distribution In the ID setting, we use the AMR 2.0
benchmark to compare SPRING variants against the best
models from the literature. To this end, we include the fol-
lowing Text-to-AMR parsers: i) Ge et al. (2019, Ge+), an
encoder-decoder model which encodes the dependency tree
and semantic role structure alongside the sentence; ii) Linde-
mann, Groschwitz, and Koller (2019, LindGK), a composi-
tional parser based on the Apply-Modify algebra; iii) Naseem
et al. (2019, Nas+), a transition-based parser trained with
a reinforcement-learning objective rewarding the Smatch
score; iv) Zhang et al. (2019b, Zhang+), a hybrid graph- and
transition-based approach incrementally predicting an AMR
graph; v) Zhou et al. (2020, Zhou+), an aligner-free parser
(Zhang et al. 2019a) enhanced with latent syntactic struc-
ture; vi) Cai and Lam (2020a, CaiL), a graph-based parser
iteratively refining an incrementally constructed graph.
For AMR-to-Text, instead, we include the following:
i) Zhu et al. (2019, Zhu+), a Transformer-based approach
enhanced with structure-aware self-attention; ii) Cai and
Lam (2020b, CaiL), a graph Transformer model which re-
lies on multi-head attention (Vaswani et al. 2017) to encode
an AMR graph in a set of node representations; iii) Wang,
Wan, and Yao (2020, Wang+), a Transformer-based model
generating sentences with an additional structure reconstruc-
tion objective; iv) Zhao et al. (2020, Zhao+), a graph at-
tention network which explicitly exploits relations by con-
structing a line graph; v) Yao, Wang, and Wan (2020, Yao+),
a graph Transformer-based model which encodes hetero-
geneous subgraph representations; vi) Mager et al. (2020,
Mag+), a fine-tuned GPT-2 model (Radford et al. 2019) pre-
dicting the PENMAN linearization of an AMR graph.
For AMR 3.0, which is a recent benchmark, there are no
previous systems to compare against. Thus, we train the pre-
vious state-of-the-art parsing model of Cai and Lam (2020a)
on AMR 3.0 and perform the corresponding evaluation.
Out-of-Distribution In the OOD setting we compare the
SPRINGDFS variants when trained on AMR 2.0 and test on
OOD data (New3, Bio and TLP) against the best of the same
variants trained on the corresponding ID training set when
available (i.e., New3 and Bio).
4.4 Evaluation
We evaluate on the Text-to-AMR parsing benchmarks by
using Smatch (Cai and Knight 2013) computed with the
tools released by Damonte, Cohen, and Satta (2017), which
also report fine-grained scores on different aspects of pars-
ing, such as wikification, concept identification, NER and
negations. As regards AMR-to-text, we follow previous ap-
proaches and evaluate using three common Natural Lan-
guage Generation (NLG) measures, i.e., BLEU (Papineni
et al. 2002, BL), chrF++ (Popovi´
c 2017, CH+), and ME-
TEOR (Banerjee and Lavie 2005, MET), tokenizing with
the script provided with JAMR (Flanigan et al. 2014). Ad-
ditionally, as AMR abstracts away from many lexical and
syntactic choices, we report the scores with untokenized
BLEURT (Sellam, Das, and Parikh 2020, BLRT), i.e., a re-
cent regression-based measure which has shown the highest
correlation with human judgements in machine translation.
5 Results
We now report the results of our experiments. First, we eval-
uate SPRING on AMR 2.0 parsing and generation; then, we
show, for the first time, the figures on the new AMR 3.0
benchmark. Finally, we tackle our proposed OOD setting.
5.1 AMR 2.0
Text-to-AMR The results on the AMR 2.0 benchmark
are reported in Table 2. Among the three different sim-
ple linearization models, i.e., SPRINGDFS, SPRINGBFS , and
SPRINGPM, the DFS-based one achieves the highest over-
all Smatch, obtaining slightly better results than the second-
best one, the PENMAN, and a wider margin over the BFS
one. All our configurations, however, outperform previous
approaches by a large margin, with SPRINGDFS outscoring
the recategorized model of Cai and Lam (2020a) by 3.6F1
points. The score gains are spread over most of the fine-
grained categories of Damonte, Cohen, and Satta (2017),
shown in the third column block in Table 2. The only no-
table exceptions are wikification and negations, where the
score of SPRINGDFS is lower than that of the previous state
of the art, i.e., Cai and Lam (2020a), which handles both
wiki links and negations heuristically. When we use recate-
gorization, i.e., in SPRINGDFS+recat, we obtain a significant
boost in performance, which is especially notable in the two
above-mentioned categories. Moreover, SPRINGDFS+recat
achieves the best reported overall performance so far, i.e.,
84.5Smatch F1 points. Regarding the other variants of
SPRINGDFS, we inspect the contribution of silver data pre-
traning, i.e., SPRINGDFS+silver, and notice a significant im-
provement over SPRINGDFS, suggesting that warm-starting
the learning is beneficial in this setting. Indeed, the model
of Ge et al. (2019), which does not exploit pretraining, per-
forms considerably worse. We note, however, that in addi-
tion to the powerful initialization of BART, our extensions
also provide a significant improvement over the BART base-
line, ranging from 0.5(SPRINGBFS) to 1.1(SPRINGDFS )
Smatch points. Finally, even when we limit the number of
parameters, and use BART Base instead, we outperform the
previous state of the art, obtaining 82.8 Smatch F1 points.
Model Recat. Smatch Unlab. NoWSD Conc. Wiki. NER Reent. Neg. SRL
Ge+ (2019) N 74.3 77.3 74.8 84.2 71.3 82.4 58.3 64.0 70.4
LindGK (2019)** N 75.3 - - - - - - - -
Nas+ (2019)** N 75.5 80.0 76.0 86.0 80.0 83.0 56.0 67.0 72.0
Zhang+ (2019b)** Y 77.0 80.0 78.0 86.0 86.0 79.0 61.0 77.0 71.0
Zhou+ (2020)* Y 77.5 80.4 78.2 85.9 86.5 78.8 61.1 76.1 71.0
CaiL (2020a)* N 78.7 81.5 79.2 88.1 81.3 87.1 63.8 66.1 74.5
CaiL (2020a)* Y 80.2 82.8 80.0 88.1 86.3 81.1 64.6 78.9 74.2
SPRINGDFS N 83.8 86.1 84.4 90.2 84.3 90.6 70.8 74.4 79.6
SPRINGBFS N 83.2 85.7 83.7 90.3 83.5 90.2 70.9 70.9 78.2
SPRINGPM N 83.6 86.1 84.1 90.1 83.1 90.2 71.4 72.7 79.4
BART baseline N 82.7 85.1 83.3 89.7 82.2 90.0 70.8 72.0 79.1
SPRINGDFS (base) N 82.8 85.3 83.3 89.6 83.5 89.9 70.2 71.5 79.0
SPRINGDFS +recat Y 84.5 86.7 84.9 89.6 87.3 83.7 72.3 79.9 79.7
SPRINGDFS +silver N 84.3 86.7 84.8 90.8 83.1 90.5 72.4 73.6 80.5
Table 2: Text-to-AMR parsing results (AMR 2.0). Row blocks: previous approaches; SPRING variants; baseline + other
SPRINGDFS. Columns: model; recategorization (Y/N); Smatch; Fine-grained scores. The best result per measure across the
table is shown in bold. The best result per measure within each row block is underlined. Models marked with */** rely on
BERT Base/Large.
BL CH+ MET BLRT
Zhu+ (2019) 31.8 64.1 36.4 -
CaiL (2020b) 29.8 59.4 35.1 -
Wang+ (2020) 32.1 64.0 36.1 -
Zhao+ (2020) 32.5 - 36.8 -
Mag+ (2020) 33.0 63.9 37.7 -
Yao+ (2020) 34.1 65.6 38.1 -
SPRINGDFS 45.3 73.5 41.0 56.5
SPRINGBFS 43.6 72.1 40.5 54.6
SPRINGPM 43.7 72.5 41.3 56.0
BART baseline 42.7 72.2 40.7 54.8
SPRINGDFS +silver 45.9 74.2 41.8 58.1
Table 3: AMR-to-Text generation results (AMR 2.0). Row
blocks: previous approaches; SPRING variants; baseline
+silver. Columns: measures. Bold/underline as in Table 2.
Finally, we compute the significance of performance dif-
ferences among SPRING variants using the non-parametric
approximate randomization test (Riezler and Maxwell
2005), which is very conservative and appropriate for
corpus-level measures. The improvement of SPRINGDFS
against SPRINGBFS and BART baseline is significant with
p < 0.005, while it is not significant when considering PEN-
MAN linearization.
AMR-to-Text We report in Table 3 the AMR 2.0 AMR-to-
Text results. SPRINGDFS achieves 45.3BLEU points, im-
proving the previous state of the art (Yao, Wang, and Wan
2020) by 11 points, and obtains very significant gains in
chrF++ and METEOR as well. As far as linearization is con-
cerned, SPRINGDFS proves to be significantly stronger than
both SPRINGPM and SPRINGBFS in 3 out of the 4 measures.
This could be due to the fact that DFS is closer to nat-
ural language than BFS, and is more compact and efficient
CaiL CaiL+r SDFS SDFS+s SDFS +r
Text-to-AMR
Smatch 78.0 76.783.0 83.0 80.2
Unlab. 81.9 80.685.4 85.4 83.1
NoWSD 78.5 77.283.5 83.5 80.7
Conc. 88.5 86.589.8 89.5 87.7
Wiki. 75.7 77.382.7 81.2 77.8
NER 83.7 74.787.2 87.1 79.8
Reent. 63.7 62.6 70.471.3 69.7
Neg. 68.9 72.6 73.0 71.775.1
SRL 73.2 72.2 78.979.1 78.1
AMR-to-Text
BL - - 44.946.5 -
CH+ - - 72.973.9 -
MET - - 40.641.7 -
BLRT - - 57.360.8 -
Table 4: Text-to-AMR and AMR-to-Text results on AMR
3.0. Best in bold. S[lin] = SPRING[lin]. +s/r = +silver/recat.
than PENMAN (see Section 3.1). Similarly to the Text-to-
AMR task results, the pretraining with silver data boosts the
performance, with SPRINGDFS+silver improving the base-
line by 0.6 BLEU points. Finally, there is a big gain against
the fine-tuned GPT-2 model of Mager et al. (2020), demon-
strating that using a pretrained decoder on its own is sub-
optimal. As in Text-to-AMR, we compute the significance
of results using the non-parametric approximate randomiza-
tion test. The performance gap between SPRINGDFS and the
alternatives in AMR-to-Text, i.e., SPRINGPM, SPRINGBFS,
and BART baseline, is significant with p < 0.001.
5.2 AMR 3.0
The results on AMR 3.0 (Table 4) confirm that SPRINGDFS
obtains the best performance. However, the important thing
New3 TLP Bio
Text-to-AMR
SPRINGDFS (ID) 78.6- 79.9
SPRINGDFS 73.7 77.359.7
SPRINGDFS+recat 63.8 76.249.5
SPRINGDFS+silver 71.877.5 59.5
AMR-to-Text
SPRINGDFS (ID) 61.5- 32.3
SPRINGDFS 51.7 41.5 5.2
SPRINGDFS+silver 50.2 40.45.9
Table 5: OOD evaluation on Text-to-AMR (Smatch) and
AMR-to-Text (BLEURT). Best in bold.
to note here is that graph recategorization, without signifi-
cant human effort in expanding the heuristics,1is not able
to scale on a more diverse benchmark such as AMR 3.0:
SPRINGDFS+recat achieves lower performances than the
non-recategorized counterpart, with the exception of nega-
tions, whose heuristics are probably more resilient to change
in data distribution. Note that the harmful impact of recate-
gorization outside of AMR 2.0 is noticeable even with the
pretrained model of Cai and Lam (2020a).
5.3 Out-of-Distribution
Finally, we show in Table 5 the results of the evaluation
on the OOD datasets. As can be seen, there is constantly
a big difference between the score achieved by the OOD
models and the best ID counterparts (see OOD paragraph
in Section 4.3), indicated as SPRINGDFS (ID). Interestingly
enough, not using recategorization results in consistently
higher performances than using it. This is especially notable
for Bio, which, in addition to being OOD with respect to
the AMR 2.0 training set, is also out-of-domain. On this
dataset SPRINGDFS (ID) model outperforms SPRINGDFS by
over 20 Smatch points, and SPRINGDFS+recat by over 30
points. On New3, which is not out-of-domain, the difference
with ID is noticeably narrower compared to SPRINGDFS
(4.9Smatch points), but considerably larger against the
SPRINGDFS+recat. Recategorization is not as harmful in
TLP, perhaps because the text of the underlying children’s
story is simpler. Differently from the results on AMR 2.0,
SPRINGDFS +silver does not show consistent improvements
over SPRINGDFS. We attribute this to the fact that the pre-
training corpus, i.e., Gigaword, is similar in distribution to
AMR 2.0, so that the boost in performance in AMR 2.0
benchmark comes due to overfitting on some genres and is
not general.
6 Case Study Analysis: Negation
Through the OOD and AMR 3.0 benchmark evaluation, we
demonstrated the harmful impact of recategorization rules
based on training sets. Interestingly, across experiments, the
1We use the heuristics designed by Zhang et al. (2019a) which
were optimized on the AMR 2.0 training set.
SPRINGDFS SPRINGDFS+recat
(1) I didn’t say he believes that.
(s / say-01
:polarity -
:ARG0 (i / i)
:ARG1 (b / believe-01
:ARG0 (h / he)
:ARG1 (t / that)))
(s / say-01
:polarity -
:ARG0 (i / i)
:ARG1 (b / believe-01
:ARG0 (h / he)
:ARG1 (t / that)))
(2) I didn’t say he said that.
(s / say-01
:polarity -
:ARG0 (i / i)
:ARG1 (s2 / say-01
:ARG0 (h / he)
:ARG1 (t / that)))
(s / say-01
:ARG0 (i / i)
:ARG1 (s2 / say-01
:polarity -
:ARG0 (h / he)
:ARG1 (t / that)))
(3) Don’t the pupils who have come last year love to study?
(l / love-01
:polarity -
:mode interrogative
:ARG0 (p / pupil
:ARG1-of (c / come-01
:time (y / year
:mod (l / last))))
:ARG1 (s / study-01
:ARG0 p))
(l / love-01
:mode interrogative
:ARG0 (p / pupil
:polarity -
:ARG1-of (c / come-01
:time (y / year
:mod (l / last))))
:ARG1 (s / study-01
:ARG0 p))
(4) Don’t eat or drink
(o / or
:op1 (e / eat-01
:mode imperative
:polarity -
:ARG0 (y / you))
:op2 (d / drink-01
:mode imperative
:polarity -
:ARG0 y))
(o / or
:op1 (e / eat-01
:mode imperative
:polarity -
:ARG0 (y / you))
:op2 (d / drink-01
:mode imperative
:ARG0 y))
Table 6: Negation examples.
breakdown scores (Damonte, Cohen, and Satta 2017) for
many aspects of meaning were consistently better without
recategorization, with the exception of negations. Negations
are handled by a commonly-used rule-based method (Zhang
et al. 2019a): :polarity attributes are discarded during
training – causing a loss of information – and are restored
by i) identifying the negated lemmas usually associated with
negative polarity words such as no, not and never; ii) align-
ing the lemma to the corresponding node in the graph by
string-matching heuristics; iii) adding the :polarity at-
tribute to the aligned node. Hand-crafted rules lead to high
precision due to the frequency of common patterns. How-
ever, there are many cases which the heuristics cannot han-
dle correctly, while fully-learned approaches are able to,
as they do not constrain the possible outputs they produce.
In Table 6 we contrast the predictions of SPRINGDFS with
SPRINGDFS +recat, trained on AMR 2.0, on several edge
cases which heuristics fail to handle. Example (1) shows
a standard negation with don’t + verb, which the designed
heuristics handle easily. However, simply changing a word,
as in example (2), makes the rule-based system crucially de-
pend on word-to-node alignment, which is non-trivial when
the same lemma (say) appears multiple times. Thus, in this
case, the heuristics misalign the negated occurrence of say,
and introduce :polarity at a lower level in the graph.
Additionally, syntax makes it such that assumptions based
on word order may easily fail: in example (3) heuristics
negate the closest lemma to the negation, i.e., pupil, instead
of the root of graph love-01, which corresponds to a word
occurring further along in the sentence. However, even if
the heuristics were rewritten to take syntax into account, it
would still be difficult to handle cases like example (4): the
negation don’t takes large scope over the conjunction, result-
ing in many :polarity edges in the AMR graph. Finally,
while due to space constraints the analysis here is limited to
negations, similar problems tend to appear whenever fine-
grained rules are applied to the input sentence, e.g., for enti-
ties, dates or politeness markers.
7 Conclusion
In this paper we presented a simple, symmetric approach
for performing state-of-the-art Text-to-AMR parsing and
AMR-to-Text generation with a single seq2seq architecture.
To achieve this, we extend a Transfomer encoder-decoder
model pretrained on English text denoising to also work with
AMR. Furthermore, we put forward a novel AMR graph
DFS-based linearization which, in addition to being more
compact than its alternatives, does not incur any informa-
tion loss. Most importantly, we drop most of the require-
ments of competing approaches: cumbersome pipelines,
heavy heuristics (often tailored to the training data), along
with most external components. Despite such cutting down
on complexity, we strongly outperform the previous state
of the art on both parsing and generation, reaching 83.8
Smatch and 45.3BLEU, respectively. We also propose an
Out-of-Distribution setting, which enables evaluation on dif-
ferent genres and domains from those of the training set.
Thanks to this setting, we are able to show that the inte-
gration of recategorization techniques or silver data – popu-
lar techniques for boosting performances – harm the perfor-
mances in both parsing and generation. Employing a sim-
pler approach like ours, based on lighter assumptions, al-
lows for more robust generalization. Here we show the gen-
eralizability of the models on different data distributions
and across domains, while leaving the extension across lan-
guages as in Blloshmi, Tripodi, and Navigli (2020) for fu-
ture work. Finally, we invite the community to use the OOD
evaluation to enable the development of more robust auto-
matic AMR approaches. Furthermore, we believe our con-
tributions will open up more directions towards the integra-
tion of parsing and generation. We release our software at
github.com/SapienzaNLP/spring.
Acknowledgments
The authors gratefully acknowledge the
support of the ERC Consolidator Grant
MOUSSE No. 726487 and the ELEXIS
project No. 731015 under the European
Union’s Horizon 2020 research and inno-
vation programme.
This work was partially supported by the MIUR under the
grant “Dipartimenti di eccellenza 2018-2022” of the Depart-
ment of Computer Science of Sapienza University.
References
Banarescu, L.; Bonial, C.; Cai, S.; Georgescu, M.; Griffitt,
K.; Hermjakob, U.; Knight, K.; Koehn, P.; Palmer, M.; and
Schneider, N. 2013. Abstract Meaning Representation for
Sembanking. In Proc. of the 7th Linguistic Annotation Work-
shop and Interoperability with Discourse, 178–186.
Banerjee, S.; and Lavie, A. 2005. METEOR: An Automatic
Metric for MT Evaluation with Improved Correlation with
Human Judgments. In Proc. of the ACL Workshop on Intrin-
sic and Extrinsic Evaluation Measures for Machine Transla-
tion and/or Summarization, 65–72. Ann Arbor, Michigan.
Beck, D.; Haffari, G.; and Cohn, T. 2018. Graph-to-Sequence
Learning using Gated Graph Neural Networks. In Proc. of
ACL 2018, 273–283. Melbourne, Australia.
Blloshmi, R.; Tripodi, R.; and Navigli, R. 2020. XL-AMR:
Enabling Cross-Lingual AMR Parsing with Transfer Learn-
ing Techniques. In Proc. of EMNLP 2020, 2487–2500. On-
line.
Bonial, C.; Donatelli, L.; Abrams, M.; Lukin, S. M.; Tratz,
S.; Marge, M.; Artstein, R.; Traum, D. R.; and Voss, C. R.
2020. Dialogue-AMR: Abstract Meaning Representation for
Dialogue. In Proc. of LREC 2020, 684–695.
Cai, D.; and Lam, W. 2019. Core Semantic First: A Top-
down Approach for AMR Parsing. In Proc. of EMNLP-
IJCNLP 2019, 3799–3809. Hong Kong, China.
Cai, D.; and Lam, W. 2020a. AMR Parsing via Graph-
Sequence Iterative Inference. In Proc. of ACL 2020, 1290–
1301. Online.
Cai, D.; and Lam, W. 2020b. Graph Transformer for Graph-
to-Sequence Learning. In Proc. of AAAI 2020, 7464–7471.
Cai, S.; and Knight, K. 2013. Smatch: an Evaluation Metric
for Semantic Feature Structures. In Proc. of ACL 2013, 748–
752.
Damonte, M.; and Cohen, S. B. 2019. Structural Neural
Encoders for AMR-to-text Generation. In Proc. of NAACL
2019: Human Language Technologies, 3649–3658. Min-
neapolis, Minnesota.
Damonte, M.; Cohen, S. B.; and Satta, G. 2017. An Incre-
mental Parser for Abstract Meaning Representation. In Proc.
of EACL 2017, 536–546. Valencia, Spain.
Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019.
BERT: Pre-training of Deep Bidirectional Transformers for
Language Understanding. In Proc. of NAACL 2019: Hu-
man Language Technologies, 4171–4186. Minneapolis, Min-
nesota. doi:10.18653/v1/N19-1423.
Flanigan, J.; Thomson, S.; Carbonell, J.; Dyer, C.; and Smith,
N. A. 2014. A Discriminative Graph-Based Parser for the
Abstract Meaning Representation. In Proc. of ACL 2014,
1426–1436. Baltimore, Maryland.
Ge, D.; Li, J.; Zhu, M.; and Li, S. 2019. Modeling Source
Syntax and Semantics for Neural AMR Parsing. In Proc. of
IJCAI 2019, 4975–4981.
Hajiˇ
c, J.; Ciaramita, M.; Johansson, R.; Kawahara, D.;
Mart´
ı, M. A.; M`
arquez, L.; Meyers, A.; Nivre, J.; Pad´
o, S.;
ˇ
Stˇ
ep´
anek, J.; Straˇ
n´
ak, P.; Surdeanu, M.; Xue, N.; and Zhang,
Y. 2009. The CoNLL-2009 Shared Task: Syntactic and Se-
mantic Dependencies in Multiple Languages. In Proc. of
CoNLL 2009: Shared Task, 1–18. Boulder, Colorado.
Hardy; and Vlachos, A. 2018. Guided Neural Language Gen-
eration for Abstractive Summarization using Abstract Mean-
ing Representation. In Proc. of EMNLP 2020, 768–773.
Konstas, I.; Iyer, S.; Yatskar, M.; Choi, Y.; and Zettlemoyer,
L. 2017. Neural AMR: Sequence-to-Sequence Models for
Parsing and Generation. In Proc. of ACL 2017, 146–157.
Vancouver, Canada.
Lewis, M.; Liu, Y.; Goyal, N.; Ghazvininejad, M.; Mohamed,
A.; Levy, O.; Stoyanov, V.; and Zettlemoyer, L. 2020. BART:
Denoising Sequence-to-Sequence Pre-training for Natural
Language Generation, Translation, and Comprehension. In
Proc. of ACL 2020, 7871–7880.
Liao, K.; Lebanoff, L.; and Liu, F. 2018. Abstract Mean-
ing Representation for Multi-Document Summarization. In
Proc. of the 27th International Conference on Computational
Linguistics, 1178–1190.
Lindemann, M.; Groschwitz, J.; and Koller, A. 2019. Com-
positional Semantic Parsing across Graphbanks. In Proc. of
ACL 2019, 4576–4585. Florence, Italy.
Liu, L.; Jiang, H.; He, P.; Chen, W.; Liu, X.; Gao, J.; and
Han, J. 2020. On the Variance of the Adaptive Learning Rate
and Beyond. In ICLR 2020.
Mager, M.; Fernandez Astudillo, R.; Naseem, T.; Sultan,
M. A.; Lee, Y.-S.; Florian, R.; and Roukos, S. 2020. GPT-
too: A Language-Model-First Approach for AMR-to-Text
Generation. In Proc. of ACL 2020, 1846–1852. Online.
May, J.; and Priyadarshi, J. 2017. SemEval-2017 Task 9: Ab-
stract Meaning Representation Parsing and Generation. In
Proc. of the 11th International Workshop on Semantic Eval-
uation, 536–545. Vancouver, Canada.
Naseem, T.; Shah, A.; Wan, H.; Florian, R.; Roukos, S.; and
Ballesteros, M. 2019. Rewarding Smatch: Transition-Based
AMR Parsing with Reinforcement Learning. In Proc. of ACL
2019, 4586–4592. Florence, Italy.
Papineni, K.; Roukos, S.; Ward, T.; and Zhu, W.-J. 2002.
Bleu: a Method for Automatic Evaluation of Machine Trans-
lation. In Proc. of ACL 2002, 311–318. Philadelphia, Penn-
sylvania, USA.
Peng, X.; Wang, C.; Gildea, D.; and Xue, N. 2017. Address-
ing the Data Sparsity Issue in Neural AMR Parsing. In Proc.
of EACL 2017, 366–375. Valencia, Spain.
Popovi´
c, M. 2017. chrF++: words helping character n-grams.
In Proc. of the 2nd Conference on Machine Translation, 612–
618. Copenhagen, Denmark.
Radford, A.; Wu, J.; Child, R.; Luan, D.; Amodei, D.; and
Sutskever, I. 2019. Language Models are Unsupervised Mul-
titask Learners. -.
Rao, S.; Marcu, D.; Knight, K.; and III, H. D. 2017. Biomedi-
cal Event Extraction using Abstract Meaning Representation.
In Proc. of BioNLP, 126–135. Vancouver, Canada.
Riezler, S.; and Maxwell, J. T. 2005. On Some Pitfalls in
Automatic Evaluation and Significance Testing for MT. In
Proc. of the ACL Workshop on Intrinsic and Extrinsic Evalu-
ation Measures for Machine Translation and/or Summariza-
tion, 57–64. Ann Arbor, Michigan.
Sellam, T.; Das, D.; and Parikh, A. 2020. BLEURT: Learning
Robust Metrics for Text Generation. In Proc. of ACL 2020,
7881–7892. Online.
Song, L.; Gildea, D.; Zhang, Y.; Wang, Z.; and Su, J. 2019.
Semantic Neural Machine Translation using AMR. TACL 7:
19–31.
Song, L.; Wang, A.; Su, J.; Zhang, Y.; Xu, K.; Ge, Y.; and
Yu, D. 2020. Structural Information Preserving for Graph-to-
Text Generation. In Proc. of ACL 2020, 7987–7998. Online.
Song, L.; Zhang, Y.; Wang, Z.; and Gildea, D. 2018. A
Graph-to-Sequence Model for AMR-to-Text Generation. In
Proc. of ACL 2018, 1616–1626. Melbourne, Australia.
van Noord, R.; and Bos, J. 2017. Neural Semantic Parsing
by Character-based Translation: Experiments with Abstract
Meaning Representations. CoRR abs/1705.09980.
Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones,
L.; Gomez, A. N.; Kaiser, u.; and Polosukhin, I. 2017. At-
tention is All You Need. In Proc. of NIPS 2017, NIPS’17,
6000–6010. Red Hook, NY, USA.
Wang, T.; Wan, X.; and Yao, S. 2020. Better AMR-To-Text
Generation with Graph Structure Reconstruction. In Proc. of
IJCAI 2020, 3919–3925.
Wu, L.; Petroni, F.; Josifoski, M.; Riedel, S.; and Zettle-
moyer, L. 2020. Scalable Zero-shot Entity Linking with
Dense Entity Retrieval. In Proc. of EMNLP 2020, 6397–
6407. Online.
Yang, Y.; Huang, L.; and Ma, M. 2018. Breaking the Beam
Search Curse: A Study of (Re-)Scoring Methods and Stop-
ping Criteria for Neural Machine Translation. In Proc. of
EMNLP 2018, 3054–3059. Brussels, Belgium.
Yao, S.; Wang, T.; and Wan, X. 2020. Heterogeneous Graph
Transformer for Graph-to-Sequence Learning. In Proc. of
ACL 2020, 7145–7154. Online.
Zhang, S.; Ma, X.; Duh, K.; and Van Durme, B. 2019a. AMR
Parsing as Sequence-to-Graph Transduction. In Proc. of ACL
2019, 80–94. Florence, Italy.
Zhang, S.; Ma, X.; Duh, K.; and Van Durme, B. 2019b.
Broad-Coverage Semantic Parsing as Transduction. In Proc.
of EMNLP-IJCNLP 2019, 3786–3798. Hong Kong, China.
Zhao, Y.; Chen, L.; Chen, Z.; Cao, R.; Zhu, S.; and Yu, K.
2020. Line Graph Enhanced AMR-to-Text Generation with
Mix-Order Graph Attention Networks. In Proc. of ACL 2020,
732–741. Online.
Zhou, Q.; Zhang, Y.; Ji, D.; and Tang, H. 2020. AMR Parsing
with Latent Structural Information. In Proc. of ACL 2020,
4306–4319. Online.
Zhu, J.; Li, J.; Zhu, M.; Qian, L.; Zhang, M.; and Zhou, G.
2019. Modeling Graph Structure in Transformer for Bet-
ter AMR-to-Text Generation. In Proc. of EMNLP-IJCNLP
2019, 5459–5468. Hong Kong, China.