Understanding and Detecting Hallucinations in
Neural Machine Translation via Model Introspection
Weijia Xu
Microsoft Research,
Redmond, USA
weijiaxu@microsoft.com
Sweta Agrawal
University of Maryland, USA
sweagraw@cs.umd.edu
Eleftheria Briakou
University of Maryland, USA
ebriakou@cs.umd.edu
Marianna J. Martindale
University of Maryland, USA
mmartind@umd.edu
Marine Carpuat
University of Maryland, USA
marine@cs.umd.edu
Abstract
Neural sequence generation models are known
to ‘‘hallucinate’’, by producing outputs that are
unrelated to the source text. These hallucina-
tions are potentially harmful, yet it remains
unclear in what conditions they arise and how
to mitigate their impact. In this work, we
first identify internal model symptoms of hal-
lucinations by analyzing the relative token
contributions to the generation in contrastive
hallucinated vs. non-hallucinated outputs gen-
erated via source perturbations. We then show
that these symptoms are reliable indicators
of natural hallucinations, by using them to de-
sign a lightweight hallucination detector which
outperforms both model-free baselines and
strong classifiers based on quality estimation
or large pre-trained models on manually an-
notated English-Chinese and German-English
translation test beds.
1
Introduction
While neural language generation models can gen-
erate high quality text in many settings, they
also fail in counter-intuitive ways, for instance
by ‘‘hallucinating’’ (Wiseman et al., 2017; Lee
et al., 2018; Falke et al., 2019). In the most se-
vere case, known as ‘‘detached hallucinations’’
(Raunak et al., 2021), the output is completely
detached from the source, which not only reveals
fundamental limitations of current models, but
also risks misleading users and undermining trust
(Bender et al., 2021; Martindale and Carpuat,
2018). Yet, we lack a systematic understanding
of the conditions where hallucinations arise, as
hallucinations occur infrequently among transla-
tions of naturally occurring text. As a workaround,
prior work has largely focused on black-box de-
tection methods which train neural classifiers on
synthetic data constructed by heuristics (Falke
546
et al., 2019; Zhou et al., 2021), and on studying
hallucinations given artificially perturbed inputs
(Lee et al., 2018; Shi et al., 2022).
In this paper, we address the problem by first
identifying the internal model symptoms that
characterize hallucinations given artificial inputs
and then testing the discovered symptoms on
texts. Specifically, we
translations of natural
study hallucinations in Neural Machine Trans-
lation (NMT) using two types of interpretability
techniques: saliency analysis and perturbations.
We use saliency analysis (Bach et al., 2015; Voita
et al., 2021) to compare the relative contributions
of various tokens to the hallucinated vs. non-
hallucinated outputs generated by diverse adver-
sarial perturbations in the inputs (Table 1) inspired
by Lee et al. (2018) and Raunak et al. (2021). Re-
sults surprisingly show that source contribution
patterns are stronger indicators of hallucinations
than the relative contributions of the source
and target, as had been previously hypothesized
(Voita et al., 2021). We discover two distinctive
source contribution patterns, including 1) con-
centrated contribution from a small subset of
source tokens, and 2) the staticity of the source
contribution distribution along the generation
steps (§ 3).
We further show that the symptoms identified
generalize to hallucinations on natural inputs by
using them to design a lightweight hallucination
classifier (§ 4) that we evaluate on manually an-
notated hallucinations from English-Chinese and
German-English NMT (Table 1). Our study shows
that our
introspection-based detection model
largely outperforms model-free baselines and the
classifier based on quality estimation scores. Fur-
thermore, it is more accurate and robust to domain
shift than black-box detectors based on large pre-
trained models (§ 5).
Transactions of the Association for Computational Linguistics, vol. 11, pp. 546–564, 2023. https://doi.org/10.1162/tacl a 00563
Action Editor: Ivan Titov. Submission batch: 7/2022; Revision batch: 12/2022; Published 6/2023.
c(cid:2) 2023 Association for Computational Linguistics. Distributed under a CC-BY 4.0 license.
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
Counterfactual hallucination from perturbation
Source
Republicans Abroad are not running a similar election, nor will they have delegates
at the convention. Recent elections have emphasized the value of each vote.
Good NMT
Perturbed
Source
Hallucination
Repulicans Abroad ar not runing a simila election, nor will they have delegates at
the convention. Recent elections have emphasized the value o each vote.
Gloss: The big ear comments that administrators have the right to retain or delete
any content in the comments under their jurisdiction.
Natural hallucination
Source
Hallucination
DAS GRUNDRECHT JEDES EINZELNEN AUF FREIE WAHL DES
BERUFS, DER AUSBILDUNGSST ¨ATTE SOWIE DES AUSBILDUNGS –
UND BESCH ¨AFTIGUNGSORTS MUSS GEWAHRT BLEIBEN.
Gloss: The fundamental right of every individual to freely choose their profession,
their training institution and their employment place must remain guaranteed.
THE PRIVACY OF ANY OTHER CLAIM, EXTRAINING STANDARDS,
EXTRAINING OR EMPLOYMENT OR EMPLOYMENT WILL BE LIABLE.
Table 1: Contrasting counterfactual English-Chinese hallucinations derived from source perturba-
tions (top) with a natural hallucination produced by a German-English NMT model (bottom).
Before presenting these two studies, we review
current findings about the conditions in which hal-
lucinations arise and formulate three hypotheses
capturing potential hallucination symptoms.
2 Hallucinations: Definition
and Hypotheses
The term ‘‘hallucinations’’ has varying definitions
in MT and natural language generation. We adopt
the most widely used one, which refers to output
text that is unfaithful to the input (Maynez et al.,
2020; Zhou et al., 2021; Xiao and Wang, 2021;
Ji et al., 2022), while others include fluency cri-
teria as part of the definition (Wang and Sennrich,
2020; Martindale et al., 2019). Different from pre-
vious work that aims to detect partial hallucina-
tions at the token level (Zhou et al., 2021), we
focus on detached hallucinations where a major
part of the output is unfaithful to the input, as these
represent severe errors, as illustrated in Table 1.
Prior work on understanding the conditions
that lead to hallucinations has focused on training
conditions and data noise (Ji et al., 2022). For
MT, Raunak et al. (2021) show that hallucinations
under perturbed inputs are caused by training
samples in the long tail that tend to be memorized
by Transformer models, while natural hallucina-
tions given unperturbed inputs can be linked to
corpus-level noise. Briakou and Carpuat (2021)
show that models trained on samples where
the source and target side diverge semanti-
cally output degenerated text more frequently.
Wang and Sennrich (2020) establish a link
between MT hallucinations under domain shift and
exposure bias by showing that Minimum Risk
Training, a training objective which addresses
exposure bias, can reduce the frequency of halluci-
nations. However, these insights do not yet provide
practical strategies for handling MT hallucinations.
A complementary approach to diagnosing hal-
lucinations is to identify their symptoms via model
introspection at inference time. However, there
lacks a systematic study of hallucinations from
the model’s internal perspective. Previous work is
either limited to an interpretation method that is
tied to an outdated model architecture (Lee et al.,
2018) or to pseudo-hallucinations (Voita et al.,
2021). In this paper, we propose to shed light on
the decoding behavior of hallucinations on both
artificially perturbed and natural inputs through
model introspection based on Layerwise Rele-
vance Propagation (LRP) (Bach et al., 2015), which
is applicable to a wide range of neural model ar-
chitectures. We focus on MT tasks with the widely
used Transformer model (Vaswani et al., 2017),
and examine existing and new hypotheses for how
547
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
hallucinations are produced. These hypotheses
share the intuition that anomalous patterns of
contributions from source tokens are indicative of
hallucinations, but operationalize it differently.
The Low Source Contribution Hypothesis
introduced by Voita et al. (2021) states that hal-
lucinations occur when NMT overly relies on the
target context over the source. They test the hy-
pothesis by inspecting the relative source and
target contributions to NMT predictions on Trans-
former models using LRP. However, their study
is limited to pseudo-hallucinations produced by
force decoding with random target prefixes. This
work will test this hypothesis on actual hallucina-
tions generated by NMT models.
The Local Source Contribution Hypothesis
introduced by Lee et al. (2018) states that hallu-
cinations occur when NMT model overly relies on
a small subset of source tokens across all gener-
ation steps. They test it by visualizing the dot-
product attention in RNN models, but it is unclear
whether these findings generalize to other model
architectures. In addition, they only study hallu-
cinations caused by random token insertion. This
work will test this hypothesis on hallucinations
under various types of source perturbations as
well as on natural inputs, and will rely on LRP to
quantify token contributions more precisely than
with attention.
Inspired by the previous observation on atten-
tion matrices that an NMT model attends repeat-
edly to the same source tokens throughout
inference when it hallucinates (Lee et al., 2018;
Berard et al., 2019b) or generates a low-quality
translation (Rikters and Fishel, 2017), we for-
malize this observation as the Static Source
Contribution Hypothesis—the distribution of
source contributions remains static along infer-
ence steps when an NMT model hallucinates.
While prior work (Lee et al., 2018; Berard et al.,
2019b; Rikters and Fishel, 2017) focuses on the
static attention to the EOS or full-stop tokens, this
hypothesis is agnostic about which source tokens
contribute. Unlike the Low Source Contribution
Hypothesis, this hypothesis exclusively relies on
the source and does not make any assumption
about relative source versus target contributions.
Unlike the Local Source Contribution Hypothe-
sis, this hypothesis is agnostic to the proportion of
source tokens contributing to a translation.
In this work, we evaluate in a controlled fash-
ion how well each hypothesis explains detached
hallucinations, first on artificially perturbed sam-
ples that let us contrast hallucinated vs. non-
hallucinated outputs in controlled settings (§ 3),
and second on natural source inputs that let us
test the generalizability of these hypotheses when
they are used to automatically detect hallucina-
tions in more realistic settings (§ 5).1
3 Study of Hallucinations under
Perturbations via Model Introspection
Hallucinations are typically rare and difficult to
identify in natural datasets. To test the aforemen-
tioned hypotheses at scale, we first exploit the fact
that source perturbations exacerbate NMT halluci-
nations (Lee et al., 2018; Raunak et al., 2021).
We construct a perturbation-based counterfac-
tual hallucination dataset on English→Chinese by
automatically identifying hallucinated NMT trans-
lations given perturbed source inputs and contrast
them with the NMT translations of the original
source (§ 3.1). This dataset lets us directly test
the three hypotheses by computing the relative to-
ken contributions to the model’s predictions using
LRP (§ 3.2), and conduct a controlled compari-
son of patterns on the original and hallucinated
samples (§ 3.4).
3.1 Perturbation-based Hallucination Data
To construct the dataset, we randomly select 50k
seed sentence pairs to perturb from the NMT train-
ing corpora, and then we apply the following
perturbations on the source sentences:2
• We randomly misspell words by deleting
characters with a probability of 0.1, as
Karpukhin et al. (2019) show that a few
misspellings can lead to egregious errors in
the output.
• We randomly title-case words with a proba-
bility of 0.1, as Berard et al. (2019a) find that
this often leads to severe output errors.
• We insert a random token at the beginning
of the source sentence, as Lee et al. (2018)
and Raunak et al. (2021) find it a reliable
trigger of hallucinations. The inserted token
is chosen from 100 most frequent, 100 least
frequent, mid-frequency tokens (randomly
1Code and data are released at https://github.com
/weijia-xu/hallucinations-in-nmt.
2For better contrastive analysis, we select samples with
source length of n = 30 and clip the output length by T = 15.
548
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
sampled 100 tokens from the remaining
tokens), and punctuations.
Inspired by Lee et al. (2018), we then iden-
tify hallucinations using heuristics that compare
the translations from the original and perturbed
sources. We select samples whose original NMT
translations y(cid:4) are of reasonable quality compared
to the reference y (i.e., bleu(y, y(cid:4)) > 0.3). The
translation of a perturbed source sentence ˜y is
identified as a hallucination if it is very different
from the translation of the original source (i.e.,
bleu(y(cid:4), ˜y) < 0.03) and is not a copy of the per-
turbed source ˜x (i.e., bleu(˜x, ˜y) < 0.5).3 This
results in 623, 270, and 1307 contrastive pairs
of the original (non-hallucinated) and hallucinated
translations under misspelling, title-casing, and
insertion perturbations, respectively.
We further divide the contrastive pairs into
degenerated and non-degenerated hallucinations.
Degenerated hallucinations are ‘‘bland, incoher-
ent, or get stuck in repetitive loops’’ (Holtzman
et al., 2020), i.e., hallucinated translations that
contain 3 more repetitive n-grams than the source
are identified as degenerated hallucinations, while
the non-degenerated group contains relatively
fluent but hallucinated translations.
3.2 Measuring Relative Token Contributions
We test the three source contribution hypothe-
ses described in § 2 on the resulting dataset by
contrasting the contributions of relevant tokens
to the generation of a hallucinated versus a non-
hallucinated translation using LRP (Bach et al.,
2015). LRP decomposes the prediction of a neu-
ral model computed over an input instance into
relevance scores for input dimensions. Specifi-
cally, LRP decomposes a neural model into several
layers of computation and measures the relative
influence score R(l)
for input neuron i at layer
i
l. Different from other interpretation methods
that measure the absolute influence of each input
dimension (Alvarez-Melis and Jaakkola, 2017;
Ma et al., 2018; He et al., 2019), LRP adopts the
principal that the relative influence R(l)
from all
i
neurons at each layer should sum up to a constant:
(cid:2)
i
R(1)
i =
(cid:2)
i
R(2)
i = . . . =
(cid:2)
i
R(L)
i = C (1)
To back-propagate the influence scores from
the last layer to the first layer (i.e., the input
layer), we need to decompose the relevance score
R(l+1)
of a neuron j at layer l + 1 into messages
j
R(l,l+1)
sent from the neuron j at layer l + 1 to
i←j
each input neuron i at layer l under the follow-
ing rules:
i←j = vijR(l+1)
R(l,l+1)
j
,
(cid:2)
vij = 1
(2)
i
There exist several versions of LRP, including
LRP-ε, LRP-αβ, and LRP-γ, which compute vij dif-
ferently (Bach et al., 2015; Binder et al., 2016;
Montavon et al., 2019). Following Voita et al.
(2021), we use LRP-αβ (Bach et al., 2015; Binder
et al., 2016), which defines vij such that the
relevance scores are positive at each step. Con-
sider first the simplest case of linear layers with
non-linear activation functions:
u(l+1)
j
= g(zj), zj =
zij + bj, zij = wiju(l)
i
i
(3)
where u(l)
is the i-th neuron at layer l, wij is the
i
weight connecting the neurons u(l)
, bj
is a bias term, and g is a non-linear activation
function. The αβ rule considers the positive and
negative contributions separately:
i and u(l+1)
j
(cid:2)
ij = max(zij, 0), b+
z+
ij = min(zij, 0), b−
z−
j = min(bj, 0)
and defines vij by the following equation:
j = max(bj, 0)
vij = α ·
z+
ij(cid:3)
ij + b+
i z+
j
+ β ·
z−
ij(cid:3)
ij + b−
i z−
j
(4)
Following Voita et al. (2021), we use α = 1,
β = 0. This rule is directly applicable to linear,
convolutional, maxpooling, and feed-forward lay-
ers. To back-propagate relevance scores through
attention layers in the Transformer encoder-
decoder model (Vaswani et al., 2017), we follow
the propagation rules in Voita et al. (2021), where
the weighting vij is obtained by performing a
first order Taylor expansion of each neuron u(l+1)
.
In the context of NMT, LRP ensures that, at
each generation step t, the sum of contributions
Rt(xi) and Rt(yj) from source tokens xi and
target prefix tokens yj remains equal:
j
(cid:2)
(cid:2)
Rt(xi) +
Rt(yj) = 1
(5)
i
j
(7)
The ratio will be lower on hallucinated samples
than on original samples if the hypothesis holds.
We compute the standardized mean difference
in High-Contribution Ratio between the halluci-
nated and original samples (Table 2).8 The nega-
tive score differences in LRP-based scores support
the hypothesis, which is consistent with the find-
ings of Lee et al. (2018) based on attention
weights. However, the attention-based score pat-
terns are not consistent on degenerated and non-
degenerated samples.
8λ0 is set to yield the largest score difference for each
measurement type.
Furthermore, we investigate whether there is
any positional bias for the local source con-
tribution. We visualize the normalized source
contribution ¯R(xi) averaged over all samples
with a source length of 30 in Figure 2. The
source contribution of the hallucinated samples
is disproportionately high at the beginning of a
source sequence. By contrast, on the original sam-
ples, the normalized contribution is higher at the
end of the source sequence, which could be a
way for the model to decide when to finish gen-
eration. The positional bias exists not only on
hallucinations under insertions at the beginning
of the source, but also on hallucinations under
misspelling and title-casing perturbations that are
applied at random positions.
Third, we examine the Static Source Contri-
bution Hypothesis hypothesis by first visualiz-
ing the source contributions Rt(xi) at varying
source and generation positions on individual
pairs of original and hallucinated samples. The
heatmaps of source contributions for the example
from Table 1 are shown in Figure 3. On the orig-
inal outputs, the source contribution distribution
in each column changes dynamically when mov-
ing horizontally along target generation steps. By
contrast, when the model hallucinates, the source
contribution distribution remains roughly static.
To quantify this pattern, we introduce Source
Contribution Staticity, which measures how the
source contribution distribution shifts over gen-
eration steps. Specifically, given a window size
k, we first divide the target sequence into several
non-overlapping segments, each containing k to-
kens. Then, we compute the average vector over
the contribution vectors Rt = [Rt(x0) . . . Rt(xn)]
at steps t within each segment. Finally, we measure
the cosine similarity between the average contri-
bution vectors of adjacent segments and average
over the cosine similarity scores at all positions
as the final score sk of window size k. Figure 4
illustrates this process for a window size of 2.
Table 2 shows the standardized mean differ-
ence in Source Contribution Staticity between
the hallucinated and original samples in the de-
generated and non-degenerated groups,
taking
the maximum staticity score among window
sizes k ∈ [1, 3]
for each sample. The pos-
itive differences in LRP-based scores supports
the Static Source Contribution Hypothesis—the
source contribution distribution is more static on
551
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
Figure 2: Normalized source contribution ¯R(xi) (Eq. 6) at each source token position averaged over the orig-
inal or hallucinated samples under (a) misspelling, (b) title-casing, and (c) insertion perturbations.
Figure 3: Heatmaps of relative contributions of source
tokens (y-axis) at each generation step (x-axis) com-
puted on the example of the original translation and the
counterfactual hallucination from the perturbed source
in Table 1. The source contribution distribution re-
mains static across almost all generation steps on the
hallucinated sample, unlike on the original sample.
the hallucinated samples than that on the original
samples. Furthermore, LRP distinguishes hallucina-
tions from non-hallucinations better than attention,
especially on non-degenerated samples where the
translation outputs contain no repetitive loops.
In summary, we find that, when generating a
hallucination under source perturbations, the NMT
model tends to rely on a small proportion of
the source tokens, especially the tokens at the
beginning of the source sentence. In addition, the
distribution of the source contributions is more
static on hallucinated translations than that on
non-hallucinated translations. We turn to applying
these insights on natural hallucinations next.
4 A Classifier to Detect Natural
Hallucinations
Figure 4: Computing the Source Contribution Staticity
of window size k = 2 given the source contribution
vectors Rt = [Rt(x0) . . . Rt(xn)] at generation step t.
Classifier We build a small multi-layer percep-
tron (MLP) with a single hidden layer and the
following input features:
• Normalized Source Contribution of the
first K1 source tokens and the last K1 source
tokens: ¯R(xi)|i = 1, . . . , K1, n − K1 +
1, . . . , n (where n is the length of the source
sequence and K1 is a hyper-parameter), as
we showed in the Local Source Contribution
Hypothesis that the contributions of the be-
ginning and end tokens distribute differently
between hallucinated and non-hallucinated
samples.
• Source Contribution Staticity sk given the
source contributions Rt(xi) and a window
size k as defined in § 3.4. We include
the similarity scores of window sizes k =
{1, 2, . . . , K2} as input features, where K2 is
a hyper-parameter.
Based on these findings, we design features for a
lightweight hallucination detector trained on sam-
ples automatically constructed by perturbations.
This yields small classifiers with input dimension
of 9. For each language pair, we train 20 classifiers
552
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
with different random seeds and select the model
with the highest validation F1 score.
Data Generation We construct
the training
and validation data using the same approach to
constructing the perturbation-based hallucination
dataset (§ 3.1), but with longer seed pairs—we
randomly select seed sentence pairs with source
length between 20 and 60 from the training cor-
pora. We split the synthetic data randomly into
the training (around 1k samples) and validation
(around 200 samples) sets with roughly equal
number of positive and negative samples.
5 Detecting Natural Hallucinations
While the hallucination classifier is trained on
hallucinations from perturbations, we collect more
realistic data to evaluate it against a wide range of
relevant models.
5.1 Natural Hallucination Evaluation Set
We build a test bed for detached hallucination
detection for different language pairs and trans-
lation directions (En-Zh and De-En), and release
the data together with the underlying NMT models
(described in § 3.3).
Since hallucinations are rare, we collect sam-
ples from large pools of out-of-domain data for
our models to obtain enough positive examples
of hallucinations for a meaningful test set. We
use TED talk transcripts from the IWSLT15 train-
ing set (Cettolo et al., 2015) for En-Zh, and the
JRC-Acquis corpus (Steinberger et al., 2006) of
legislation from the European Union for De-En.
To increase the chance of finding hallucinations,
we select around 200, 50, and 50 translation out-
puts with low BLEU, low COMET (Rei et al., 2020a),
or low LASER similarity (Artetxe and Schwenk,
2019) scores, respectively. We further combine
them with 50 randomly selected samples.
Three bilingual annotators assess the faithful-
ness of the NMT output given each input. While
we ultimately need a binary annotation of outputs
as hallucinated or not, annotators were asked to
choose one of five labels to improve consistency:
• Detached hallucination: a translation with
large segments that are unrelated to the
source.
• Faithful
translation: a translation that
is
faithful to the source.
Detached hallucination
Non hallucination, including:
Faithful translation
Incomplete translation
Locally unfaithful
Incomprehensible but aligned
Total
En-Zh De-En
111
189
154
80
58
5
408
153
17
31
33
423
Table 3: Human annotation label distribution on
the En-Zh and De-En natural hallucination test sets
(with random tie breaking on fine-grained labels;
there are no ties on binary labels post-aggregation).
• Incomplete translation: a translation that is
partially correct but misses part(s) of the
source.
• Locally unfaithful: a translation that contains
a few unfaithful phrases but is otherwise
faithful.
• Incomprehensible but aligned: a translation
that is incomprehensible even though most
phrases can be aligned to the source.
All labels except for the ‘‘detached hallucina-
tion’’ are aggregated into the ‘‘non-hallucination’’
category. The inter-annotator agreement on aggre-
gated labels is substantial, with a Fleiss’s Kappa
(Fleiss, 1971) score of F K = 0.77 for De-En
and F K = 0.64 for En-Zh. Disagreements are
resolved by majority voting for De-En, and by
adjudication by a bilingual speaker for En-Zh.
This yields 27% of detached hallucinations on
En-Zh and 45% on De-En. The non-hallucinated
NMT outputs span all the fine-grained categories
above, as can be seen in Table 3. Hallucinations
are over-represented compared to what one might
expect in the wild, but this is necessary to provide
enough positive examples of hallucinations for
evaluation.
5.2 Experimental Conditions
5.2.1 Introspection-based Classifiers
We implement the LRP-based classifier described
in § 4. To lower the cost of computing source
contributions, we clip the source length at 40,
and only consider the influence back-propagated
through the most recent 10 target tokens—prior
work shows that nearby context is more influen-
tial than distant context (Khandelwal et al., 2018).
553
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
We tune the hyper-parameters K1 and K2
within the space K1 ∈ {1, 3, 5, 7, 9}, K2 ∈
{4, 8, 12, 16} based on the average F1 accuracy
on the validation set over three runs. We compare
it with an attention-based classifier, which uses
the same features, but computes token contribu-
tions using attention weights averaged over all
attention heads.
5.2.2 Model-free Baselines
We use three simple baselines to characterize the
task. The random classifier that predicts halluci-
nation with a probability of 0.5. The degenera-
tion detector marks as hallucinations degenerated
outputs that contain K more repetitive n-grams
than the source, where K is a hyper-parameter
tuned on the perturbation-based hallucination data.
The NMT probability scores are used as a coarse
model signal to detect hallucinations based on the
heuristic that the model is less confident when
producing a hallucination. The output is classi-
fied as a hallucination if the probability score is
lower than a threshold tuned on the perturbation-
based hallucination data.
5.2.3 Quality Estimation Classifier
We also compare the introspection-based classi-
fiers with a baseline classifier based on the state-
of-the-art quality estimation model—COMET-QE
(Rei et al., 2020b). Given a source sentence and
its NMT translation, we compute the COMET-QE
score and classify the translation as a halluci-
nation if the score is below a threshold tuned on
the perturbation-based validation set.
5.2.4 Large Pre-trained Classifiers
We further compare the introspection-based clas-
sifiers with classifiers that rely on large pre-trained
multilingual models, to compare the discrimi-
native power of the source contribution patterns
from the NMT model itself to extrinsic semanti-
cally driven discrimination criteria.
We use the cosine distance between the LASER
representations (Artetxe and Schwenk, 2019;
Heffernan et al., 2022) of the source and the NMT
translation. It classifies a translation as a halluci-
nation if the distance score is higher than a thresh-
old tuned on the perturbation-based validation set.
Inspired by local hallucination (Zhou et al.,
2021) and cross-lingual semantic divergence
(Briakou and Carpuat, 2020) detection methods,
we build an XLM-R classifier by fine-tuning the
XLM-R model (Conneau et al., 2020) on synthetic
hallucination samples. We randomly select 50K
seed pairs of source and reference sentences with
source lengths between 20 and 60 from the paral-
lel corpus and use the following perturbations to
construct examples of detached hallucinations:
• Map a source sentence to a random target
from the parallel corpus to simulate natural,
detached hallucinations.
• Repeat a random dependency subtree in the
reference many times to simulate degener-
ated hallucinations.
• Drop a random clause from the source
sentence to simulate natural, detached hal-
lucinations.
We then collect diverse non-hallucinated samples:
• Original seed pairs provide faithful transla-
tions.
• Randomly drop a dependency subtree from a
reference to simulate incomplete translations.
• Randomly substitute a phrase in the reference
keeping the same part-of-speech to simulate
translations with locally unfaithful phrases.
The final
training and validation sets contain
around 300k and 700 samples, respectively. We
fine-tune the pre-trained model with a batch size of
32. We use the Adam optimizer (Kingma and Ba,
2015) with decoupled weight decay (Loshchilov
and Hutter, 2019) and an initial learning rate of
2 × 10−5. We fine-tune all models for 5 epochs
and select the checkpoint with the highest F1
score on the validation set.
5.3 Findings
As shown in Table 4, we compare all classifiers
against the baselines by the Precision, Recall, and
F1 scores. Since false positives and false negatives
might have a different impact in practice (e.g.,
does the detector flag examples for review by
humans, or entirely automatically? what is MT
used for?), we also report the Area Under the
Receiver Operating Characteristic Curve (AUC),
which characterizes the discriminative power of
each method at varying threshold settings.
554
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
–
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
Params
De-En
En-Zh
P
R
F1
AUC
P
R
F1
AUC
Model-free Baselines
Random
Degeneration
NMT Score
0
1
1
Quality Estimation Classifier
COMET-QE
363M
Large Pre-trained Classifiers
LASER
XLM-R
45M
125M
Introspection-based Classifiers
Attention-based
LRP-based
< 400
< 400
Ensemble Classifier
LRP + LASER
LRP + XLM-R
45M
125M
44.0
49.1
33.3
49.9
59.3
3.4
46.8
53.7
6.2
50.2
–
37.7
27.6
63.2
35.4
49.8
71.2
91.9
35.5
66.9
51.1
48.0
–
49.8
72.2
71.4
71.8
82.4
32.4
99.1
48.9
89.4
81.6
91.3
54.0
21.0
65.0
33.8
89.5
45.6
54.6
94.9
64.0
83.2
58.9
88.6
75.3
93.3
54.3
87.3
89.0
76.2
67.4
81.2
70.1
91.4
36.0
87.5
71.0
85.6
47.7
86.4
68.6
96.5
100.0
95.3
45.7
21.5
62.7
35.1
–
–
94.5
97.6
59.5
72.4
72.9
83.1
–
–
Table 4: Precision (P), Recall (R), F1, and Area Under the Receiver Operating Characteristic Curve
(AUC) scores of each classifier on English-Chinese (En-Zh) and German-English (De-En) NMT out-
puts (means of three runs). We boldface the highest scores based on independent Student t-test with
Bonferroni correction (p < 0.05). The Params column indicates the total number of parameters used
for each method (in addition to the NMT parameters).
Main Results The LRP-based, XLM-R, and the
LASER classifiers are the best hallucination detec-
tors, reaching AUC scores around 90 for either
or both language pairs, which is considered out-
standing discrimination ability (Hosmer Jr et al.,
2013).
The LRP-based classifier is the best and most
robust hallucination detector overall. It achieves
higher F1 and AUC scores than LASER on both lan-
guage pairs. Additionally, it outperforms XLM-R by
+47 F1 and +46 AUC on De-En, while achieving
competitive performance on En-Zh. This shows
that the source contribution patterns identified on
hallucinations under perturbations (§ 3) general-
ize as symptoms of natural hallucinations even
under domain shift, as the domain gap between
training and evaluation data is bigger on De-En
than En-Zh. It also confirms that LRP provides a
better signal to characterize token contributions
than attention, improving F1 by 14–39 points and
AUC by 21–28 points. These high scores represent
large improvements of 41–54 points on AUC and
20–75 points on F1 over the model-free baselines.
Model-free Baselines These baselines shed
light on the nature of the hallucinations in the
dataset. The degeneration baseline is the best
among them, with 53.7 F1 on De-En and 66.9
F1 on En-Zh, indicating that the Chinese halluci-
nations are more frequently degenerated than the
English hallucinations from German. However,
ignoring the remaining hallucinations is problem-
atic, since they might be more fluent and thus more
likely to mislead readers. The NMT score is a poor
predictor, scoring worse than the random baseline
on De-En, in line with previous reports that NMT
scores do not capture faithfulness well during in-
ference (Wang et al., 2020). Manual inspection
shows that the NMT score can be low when the
output is faithful but contains rare words, and it
can be high for a hallucinated output that contains
mostly frequent words.
Quality Estimation Classifier The COMET-QE
classifier achieves higher AUC and F1 scores
than the model-free classifiers, except for En-Zh,
where the degeneration baseline obtains higher F1
555
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
than the COMET-QE classifier. However, compared
with the LRP-based classifier, COMET-QE lags be-
hind by 9-38 points on F1 and 7-9 points on
AUC. This is consistent with previous findings
that quality estimation models trained on data
with insufficient negative samples (e.g., COMET-
QE) are inadequate for detecting critical MT errors
such as hallucinations (Takahashi et al., 2021;
Sudoh et al., 2021; Guerreiro et al., 2022).
Pre-trained Classifiers The performance of
pre-trained classifiers varies greatly across lan-
guage pairs. LASER achieves a competitive AUC
score to the LRP-based classifier on De-En but
lags behind on En-Zh, perhaps because the LASER
model is susceptible to the many rare tokens in
the En-Zh evaluation data (from TED subtitles).
XLM-R obtains better performance on En-Zh, ap-
proaching that of the LRP-based classifier, but lags
behind greatly on De-En. This suggests that the
XLM-R classifier suffers from domain shift, which
is bigger on De-En (News→Law) than En-Zh
(News→TED). Fine-tuning the model on the syn-
thetic training data generalizes more poorly across
domains. By contrast,
the introspection-based
classifiers are more robust.
Ensemble Classifiers The LASER and XLM-R clas-
sifiers emerge as the top classifiers apart from the
LRP-based one, but they make different errors than
LRP—the confusion matrix comparing their pre-
dictions shows that the LASER and LRP classifiers
agree on 68–78% of samples, while the XLM-R
and LRP classifiers agree on 64–88% of samples.
Thus an ensemble of LRP + LASER or LRP + XLM-R
(which detects hallucinations when the two clas-
sifiers both do so) yields a very high precision (at
the expense of recall).
De-En
En-Zh
F1
AUC
F1
AUC
All features
- Src Contrib
- Staticity
81.2
74.4
50.7
91.4
92.7
76.6
86.4
85.3
58.3
96.5
96.1
80.0
Table 5: Ablating the Normalized Source Con-
tribution (Src Contrib) and Source Contribution
Staticity (Staticity) features used in the LRP-based
classifier. We boldface the highest scores based
on independent student’s t-test with Bonferroni
Correction(p < 0.05).
Source: C) DASS DIE WAREN IN DEM
ZUSTAND IN DIE GEMEINSCHAFT VER-
SANDT WORDEN SIND, IN DEM SIE ZUR
AUSSTELLUNG GESANDT WURDEN;
Correct Translation: C) THAT THE GOODS
WERE SHIPPED TO THE COMMUNITY IN
THE CONDITION IN WHICH THEY ARE
SENT FOR EXHIBITION;
Output: C) THAT THE WOULD BE CON-
SIDERED IN THE COMMUNITY, IN which
YOU WILL BE EXCLUSIVE;
Table 6: Example of a detached hallucination
produced by the De-En NMT being classified as
non-hallucination by the LRP-based classifier.
translations. Additionally, the classifier can fail to
detect hallucinations caused by the mistranslation
of a large span of the source with rare or previously
unseen tokens, rather than by pathological behav-
ior at inference time as shown by the example in
Table 6.
LRP Ablations The LRP-based classifier benefits
the most from Source Contribution Staticity fea-
tures (Table 5). Removing them hurts AUC by
15–17 points and F1 by 28–31 points, confirm-
ing that the Static Source Contribution Hypothesis
holds on natural hallucinations. Ablating the Nor-
malized Source Contribution features also causes
a significant drop in F1 on De-En, while its impact
on En-Zh is not significant.
Error Analysis
Incomprehensible but aligned
translations suffer from the highest false positive
rate for the LRP classifier, followed by incomplete
Toward Practical Detectors Detecting halluci-
nations in the wild is challenging since they tend
to be rare and their frequency may vary greatly
across test cases. We provide a first step in this
direction by stress testing the top classifiers in an
in-domain scenario where hallucinations are ex-
pected to be rare. Specifically, we randomly select
10k English sentences from the News Crawl: ar-
ticles from 2021 from WMT21 (Akhbardeh et al.,
2021) and use the En-Zh NMT model to translate
them into Chinese. We measure the Precision@20
for hallucination detection by manually exam-
ining the top-20 highest scoring hallucination
556
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
predictions for each method. The LASER, XLM-R,
and LRP-based classifiers evaluated above (with-
out fine-tuning in this setting) achieve 35%, 45%,
and 45% Precision@20, respectively (compared
to 0% for the random baseline). More interest-
ingly, after tuning the threshold on the predicted
probabilities (which is originally set to 0.5) so
that each classifier predicts hallucination 1% of
the time, the LRP + LASER ensemble detects 9 hal-
lucinations with a much higher precision of 89%,
and the LRP + XLM-R ensemble detects 12 halluci-
nations with a precision of 83%. These ensemble
detectors thus have the potential to provide use-
ful signals for detecting hallucinations even when
they are needles in a haystack.
5.4 Limitations
Our findings should be interpreted with several
limitations in mind. First, we exclusively study
detached hallucinations in MT. Thus, we do not
elucidate the internal model symptoms that lead to
partial hallucinations (Zhou et al., 2021), although
the methodology in this work could be used to
shed light on this question. Second, we work with
NMT models trained using the parallel data from
WMT without exploiting monolingual data or com-
parable corpora retrieved from collections of
monolingual texts (e.g., WikiMatrix [Schwenk
et al., 2021]). It remains to be seen whether halluci-
nation symptoms generalize to NMT models trained
with more heterogeneous supervision. Finally,
we primarily test the hallucination classifiers in
roughly balanced test sets, while hallucinations
are expected to be rare in practice. We conducted
a small stress test which shows the promise of our
LRP +LASER classifier in more realistic conditions.
However, further work is needed to systematically
evaluate how these classifiers can be used for
hallucination detection in the wild.
6 Related Work
Hallucinations occur in all applications of neural
models to language generation, including abstrac-
tive summarization (Falke et al., 2019; Maynez
et al., 2020), dialogue generation (Duˇsek et al.,
2018), data-to-text generation (Wiseman et al.,
2017), and machine translation (Lee et al., 2018).
Most existing detection approaches view the gen-
eration model as a black-box, by 1) training hal-
lucination classifiers on synthetic data constructed
by heuristics (Zhou et al., 2021; Santhanam et al.,
2021), or 2) using external models to measure
the faithfulness of the outputs, such as question
answering or natural language inference models
(Falke et al., 2019; Durmus et al., 2020). These
approaches ignore the signals from the genera-
tion model itself and could be highly biased by
the heuristics used for synthetic data construc-
tion, or the biases in the external semantic mod-
els trained for other purposes. Concurrent to this
work, Guerreiro et al. (2022) explore glass-box
detection methods based on model confidence
scores or attention patterns (e.g., the proportion
of attention paid to the EOS token and the pro-
portion of source tokens with attention weights
higher than a threshold). They evaluate these
methods based on hallucination recall, and find
that model confidence is a better indicator of hal-
lucinations than attention patterns. In this paper,
we investigated varying types of glass-box pat-
terns based on the relative token contributions
instead of attention, and find that these patterns
yield more accurate hallucination detectors than
model confidence.
Detecting hallucinations in MT has not yet been
directly addressed by the MT quality estimation lit-
erature. Most quality estimation work has focused
on predicting a direct assessment of translation
quality, which does not distinguish adequacy and
fluency errors (Guzm´an et al., 2019; Specia et al.,
2020). More recent task formulations target crit-
ical adequacy errors (Specia et al., 2021), but do
not separate hallucinations from other error types,
despite arguments that hallucinations should be
considered separately from other MT errors (Shi
et al., 2022). The critical error detection task at
WMT 2022 introduces an Additions error category,
which refers to hallucinations where the trans-
lation content is only partially supported by the
source (Zerva et al., 2022). Additions includes
both detached hallucinations (as in this work) and
partial hallucinations. Methods for addressing all
these tasks fall in two categories: 1) black-box
methods based on the source and output alone
(Specia et al., 2009; Kim et al., 2017; Ranasinghe
et al., 2020), and 2) glass-box methods based
on features extracted from the NMT model itself
(Rikters and Fishel, 2017; Yankovskaya et al.,
2018; Fomicheva et al., 2020). Black-box methods
typically use resource-heavy deep neural networks
trained on large amounts of annotated data. Our
work is inspired by the glass-box methods that
557
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
rely on model probabilities, uncertainty quantifi-
cation, and the entropy of the attention distribu-
tion, but shows that relative token contributions
computed through LRP provide sharper features
to characterize hallucinations.
This paper combines interpretability techniques
to identify the symptoms of hallucinations. We
adopt a saliency method to measure the importance
of each input unit through a back-propagation
pass (Simonyan et al., 2014; Bach et al., 2015;
Li et al., 2016a; Ding et al., 2019). While other
saliency-based methods measure an abstract quan-
tity reflecting the importance of each input feature
by the partial derivative of the prediction with
regard to each input unit (Simonyan et al., 2014),
LRP (Bach et al., 2015) measures the proportional
contribution of each input unit. This makes it well-
suited to compare model behavior across sam-
ples. Furthermore, LRP does not require neural
activations to be differentiable and smooth, and
can be applied to a wide range of architectures, in-
cluding RNN (Ding et al., 2017) and Transformer
(Voita et al., 2021). We apply this technique to
analyze counterfactual hallucination samples in-
spired by perturbation methods (Li et al., 2016b;
Feng et al., 2018; Ebrahimi et al., 2018), but cru-
cially show that the insights generalize to natural
hallucinations.
7 Conclusion
We contribute a thorough empirical study of
the notorious but poorly understood hallucination
phenomenon in NMT, which shows that internal
model symptoms exhibited during inference are
strong indicators of hallucinations. Using counter-
factual hallucinations triggered by perturbations,
we show that distinctive source contribution pat-
terns alone indicate hallucinations better than the
relative contributions of the source and target.
We further show that our findings can be used
for detecting natural hallucinations much more
accurately than model-free baselines and quality
estimation models. Our detector also outperforms
black-box classifiers based on pre-trained models.
We release human-annotated test beds of natural
English-Chinese and German-English hallucina-
tions to enable further research. This work opens
a path toward detecting hallucinations in the
wild and improving models to minimize halluci-
nations in MT and other generation tasks.
Acknowledgments
We thank our TACL action editor, the anony-
mous reviewers, and the UMD CLIP lab for their
feedback. Thanks also to Yuxin Xiong for help-
ing examine German outputs. This research is
supported in part by an Amazon Machine Learn-
ing Research Award and by the National Science
Foundation under Award No. 1750695. Any opin-
ions, findings, conclusions, or recommendations
expressed in this material are those of the authors
and do not necessarily reflect the views of the
National Science Foundation.
References
Farhad Akhbardeh, Arkady Arkhangorodsky,
Magdalena Biesialska, Ondˇrej Bojar, Rajen
Chatterjee, Vishrav Chaudhary, Marta R.
Costa-jussa, Cristina Espa˜na-Bonet, Angela
Fan, Christian Federmann, Markus Freitag,
Yvette Graham, Roman Grundkiewicz, Barry
Haddow, Leonie Harter, Kenneth Heafield,
Christopher Homan, Matthias Huck, Kwabena
Amponsah-Kaakyire,
Jungo Kasai, Daniel
Khashabi, Kevin Knight, Tom Kocmi, Philipp
Koehn, Nicholas Lourie, Christof Monz,
Makoto Morishita, Masaaki Nagata, Ajay
Nagesh, Toshiaki Nakazawa, Matteo Negri,
Santanu Pal, Allahsera Auguste Tapo, Marco
Turchi, Valentin Vydrin, and Marcos Zampieri.
2021. Findings of the 2021 conference on ma-
chine translation (WMT21). In Proceedings
of the Sixth Conference on Machine Trans-
lation, pages 1–88, Online. Association for
Computational Linguistics.
David Alvarez-Melis and Tommi Jaakkola. 2017.
A causal framework for explaining the predic-
tions of black-box sequence-to-sequence mod-
els. In Proceedings of the 2017 Conference on
Empirical Methods in Natural Language Pro-
cessing, pages 412–421, Copenhagen, Denmark.
Association for Computational Linguistics.
https://doi.org/10.18653/v1/D17
-1042
Mikel Artetxe and Holger Schwenk. 2019. Mas-
sively multilingual sentence embeddings for
zero-shot cross-lingual
transfer and beyond.
Transactions of the Association for Computa-
tional Linguistics, 7:597–610. https://doi
.org/10.1162/tacl_a_00288
558
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
Sebastian Bach, Alexander Binder, Gr´egoire
Montavon, Frederick Klauschen, Klaus-Robert
M¨uller, and Wojciech Samek. 2015. On pixel-
wise explanations for non-linear classifier de-
cisions by layer-wise relevance propagation.
PloS One, 10(7):e0130140. https://doi.org
/10.1371/journal.pone.0130140, PubMed:
26161953
Emily M. Bender, Timnit Gebru, Angelina
McMillan-Major, and Shmargaret Shmitchell.
2021. On the dangers of stochastic parrots:
Can language models be too big? In Proceed-
ings of the 2021 ACM Conference on Fairness,
Accountability, and Transparency FAccT ’21,
pages 610–623, New York, NY, USA. Asso-
ciation for Computing Machinery. https://
doi.org/10.1145/3442188.3445922
Alexandre Berard,
Ioan Calapodescu, Marc
Dymetman, Claude Roux, Jean-Luc Meunier,
and Vassilina Nikoulina. 2019a. Machine
translation of restaurant reviews: New cor-
pus for domain adaptation and robustness. In
Proceedings of the 3rd Workshop on Neural
Generation and Translation, pages 168–176,
Hong Kong. Association for Computational
Linguistics. https://doi.org/10.18653/v1
/D19-5617
Alexandre Berard, Ioan Calapodescu, and Claude
Roux. 2019b. Naver labs Europe’s systems
for the WMT19 machine translation robustness
task. In Proceedings of the Fourth Conference
on Machine Translation (Volume 2: Shared
Task Papers, Day 1), pages 526–532, Florence,
Italy. Association for Computational Linguis-
tics. https://doi.org/10.18653/v1
/W19-5361
Alexander
Binder,
Gr´egoire Montavon,
Sebastian Lapuschkin, Klaus-Robert M¨uller,
and Wojciech Samek. 2016. Layer-wise rel-
evance propagation for neural networks with
local renormalization layers. In International
Conference on Artificial Neural Networks,
pages 63–71. Springer. https://doi.org
/10.1007/978-3-319-44781-0_8
Ondrej Bojar, Christian Federmann, Mark Fishel,
Yvette Graham, Barry Haddow, Matthias Huck,
Philipp Koehn, and Christof Monz. 2018.
Findings of the 2018 conference on machine
translation (WMT18). In Proceedings of the
559
Third Conference on Machine Translation,
pages 272–307. Association for Computa-
tional Linguistics. https://doi.org/10
.18653/v1/W18-6401
Eleftheria Briakou and Marine Carpuat. 2020.
Detecting fine-grained cross-lingual semantic
divergences without supervision by learning
to rank. In Proceedings of the 2020 Confer-
ence on Empirical Methods in Natural Lan-
guage Processing (EMNLP), pages 1563–1580,
Online. Association for Computational Lin-
guistics. https://doi.org/10.18653/v1
/2020.emnlp-main.121
Eleftheria Briakou
and Marine Carpuat.
2021. Beyond noise: Mitigating the impact
of fine-grained semantic divergences on neu-
ral machine translation. In Proceedings of the
59th Annual Meeting of the Association for
Computational Linguistics and the 11th Inter-
national Joint Conference on Natural Lan-
guage Processing (Volume 1: Long Papers),
pages 7236–7249, Online. Association for
Computational Linguistics. https://doi.org
/10.18653/v1/2021.acl-long.562
Mauro Cettolo, Jan Niehues, Sebastian St¨uker,
and
Luisa Bentivogli, Roldano Cattoni,
Marcello Federico. 2015. The IWSLT 2015
evaluation campaign. In Proceedings of the
12th International Workshop on Spoken Lan-
guage Translation: Evaluation Campaign,
pages 2–14, Da Nang, Vietnam.
Pinzhen Chen, Jindˇrich Helcl, Ulrich Germann,
Laurie Burchell, Nikolay Bogoychev, Antonio
Valerio Miceli Barone, Jonas Waldendorf,
Alexandra Birch, and Kenneth Heafield. 2021.
The University of Edinburgh’s English-German
and English-Hausa submissions to the WMT21
news translation task. In Proceedings of the
Sixth Conference on Machine Translation,
pages 104–109, Online. Association for Com-
putational Linguistics.
Alexis Conneau, Kartikay Khandelwal, Naman
Goyal, Vishrav Chaudhary, Guillaume Wenzek,
Francisco Guzm´an, Edouard Grave, Myle Ott,
Luke Zettlemoyer, and Veselin Stoyanov.
2020. Unsupervised cross-lingual representa-
tion learning at scale. In Proceedings of the
58th Annual Meeting of the Association for
Computational Linguistics, pages 8440–8451,
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
Online. Association for Computational Lin-
guistics. https://doi.org/10.18653
/v1/2020.acl-main.747
Shuoyang Ding, Hainan Xu, and Philipp Koehn.
2019. Saliency-driven word alignment inter-
pretation for neural machine translation. In
Proceedings of
the Fourth Conference on
Machine Translation (Volume 1: Research
Papers), pages 1–12, Florence, Italy. Associa-
tion for Computational Linguistics. https://
doi.org/10.18653/v1/W19-5201
Yanzhuo Ding, Yang Liu, Huanbo Luan, and
Maosong Sun. 2017. Visualizing and un-
derstanding neural machine translation.
In
Proceedings of the 55th Annual Meeting of
the Association for Computational Linguistics
(Volume 1: Long Papers), pages 1150–1159,
Vancouver, Canada. Association for Computa-
tional Linguistics. https://doi.org/10
.18653/v1/P17-1106
Esin Durmus, He He, and Mona Diab. 2020.
FEQA: A question answering evaluation frame-
work for faithfulness assessment in abstractive
summarization. In Proceedings of the 58th An-
nual Meeting of the Association for Computa-
tional Linguistics, pages 5055–5070, Online.
Association for Computational Linguistics.
https://doi.org/10.18653/v1/2020
.acl-main.454
Ondˇrej Duˇsek,
Jekaterina Novikova,
and
Verena Rieser. 2018. Findings of the E2E NLG
challenge. In Proceedings of the 11th Interna-
tional Conference on Natural Language Gen-
eration, pages 322–328, Tilburg University,
The Netherlands. Association for Computa-
tional Linguistics. https://doi.org/10
.18653/v1/W18-6539
Javid Ebrahimi, Daniel Lowd, and Dejing Dou.
2018. On adversarial examples for character-
level neural machine translation. In Proceed-
ings of
the 27th International Conference
on Computational Linguistics, pages 653–663,
Santa Fe, New Mexico, USA. Association for
Computational Linguistics.
Tobias Falke, Leonardo F. R. Ribeiro, Prasetya
Ajie Utama, Ido Dagan, and Iryna Gurevych.
2019. Ranking generated summaries by cor-
rectness: An interesting but challenging ap-
560
plication for natural
language inference. In
Proceedings of the 57th Annual Meeting of
the Association for Computational Linguistics,
pages 2214–2220, Florence, Italy. Association
for Computational Linguistics. https://doi
.org/10.18653/v1/P19-1213
Shi Feng, Eric Wallace, Alvin Grissom II,
Mohit Iyyer, Pedro Rodriguez, and Jordan
Boyd-Graber. 2018. Pathologies of neural
models make interpretations difficult. In Pro-
ceedings of the 2018 Conference on Empirical
Methods in Natural Language Processing,
pages 3719–3728, Brussels, Belgium. Associa-
tion for Computational Linguistics. https://
doi.org/10.18653/v1/D18-1407
Joseph L. Fleiss. 1971. Measuring nominal scale
agreement among many raters. Psychological
Bulletin, 76(5):378. https://doi.org/10
.1037/h0031619
Marina Fomicheva, Shuo Sun, Lisa Yankovskaya,
Fr´ed´eric Blain, Francisco Guzm´an, Mark
Fishel, Nikolaos Aletras, Vishrav Chaudhary,
and Lucia Specia. 2020. Unsupervised qual-
ity estimation for neural machine translation.
Transactions of the Association for Computa-
tional Linguistics, 8:539–555. https://doi
.org/10.1162/tacl_a_00330
Ulrich Germann. 2020. The University of Edin-
burgh’s submission to the German-to-English
and English-to-German tracks in the WMT
2020 news translation and zero-shot
trans-
In Proceedings of
lation robustness tasks.
the Fifth Conference on Machine Translation,
pages 197–201, Online. Association for Com-
putational Linguistics.
Nuno M. Guerreiro, Elena Voita, and Andr´e
F. T. Martins. 2022. Looking for a needle in a
haystack: A comprehensive study of hallucina-
tions in neural machine translation. https://
doi.org/10.48550/arXiv.2208.05309
Francisco Guzm´an, Peng-Jen Chen, Myle
Ott, Juan Pino, Guillaume Lample, Philipp
Koehn, Vishrav Chaudhary, and Marc’Aurelio
Ranzato. 2019. The FLORES evaluation da-
tasets for low-resource machine translation:
Nepali–English and Sinhala–English. In Pro-
ceedings of the 2019 Conference on Empirical
Methods in Natural Language Processing and
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
the 9th International Joint Conference on Nat-
ural Language Processing (EMNLP-IJCNLP),
pages 6098–6111, Hong Kong, China. Associ-
ation for Computational Linguistics. https://
doi.org/10.18653/v1/D19-1632
Shilin He, Zhaopeng Tu, Xing Wang, Longyue
Wang, Michael Lyu, and Shuming Shi. 2019.
Towards understanding neural machine trans-
lation with word importance. In Proceedings
of the 2019 Conference on Empirical Meth-
ods in Natural Language Processing and the
9th International Joint Conference on Natu-
ral Language Processing (EMNLP-IJCNLP),
pages 953–962, Hong Kong, China. Associa-
tion for Computational Linguistics. https://
doi.org/10.18653/v1/D19-1088
Kevin Heffernan, Onur C¸ elebi, and Holger
Schwenk. 2022. Bitext mining using dis-
tilled sentence representations for low-resource
languages. https://doi.org/10.48550
/arXiv.2205.12654
Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes,
and Yejin Choi. 2020. The curious case of
neural text degeneration. In International Con-
ference on Learning Representations.
David W. Hosmer Jr., Stanley Lemeshow, and
Rodney X. Sturdivant. 2013. Applied Logistic
Regression, volume 398. John Wiley & Sons.
https://doi.org/10.1002/9781118548387
Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng
Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin
Bang, Andrea Madotto, and Pascale Fung.
2022. Survey of hallucination in natural lan-
guage generation. https://doi.org/10
.48550/arXiv.2202.03629
Vladimir Karpukhin, Omer Levy,
Jacob
Eisenstein, and Marjan Ghazvininejad. 2019.
Training on synthetic noise improves robust-
ness to natural noise in machine transla-
tion. In Proceedings of the 5th Workshop on
Noisy User-generated Text (W-NUT 2019),
pages 42–47, Hong Kong, China. Association
for Computational Linguistics. https://doi
.org/10.18653/v1/D19-5506
tics (Volume 1: Long Papers), pages 284–294,
Melbourne, Australia. Association for Compu-
tational Linguistics. https://doi.org/10
.18653/v1/P18-1027
Hyun Kim, Jong-Hyeok Lee, and Seung-Hoon
Na. 2017. Predictor-estimator using multi-
level task learning with stack propagation for
neural quality estimation. In Proceedings of
the Second Conference on Machine Transla-
tion, pages 562–568, Copenhagen, Denmark.
Association for Computational Linguistics.
https://doi.org/10.18653/v1/W17-4763
Diederik P. Kingma and Jimmy Ba. 2015.
Adam: A method for stochastic optimization.
In Proceedings of the 3th International Con-
ference on Learning Representations, San
Diego, CA, USA.
Philipp Koehn, Hieu Hoang, Alexandra Birch,
Chris Callison-Burch, Marcello Federico,
Nicola Bertoldi, Brooke Cowan, Wade Shen,
Christine Moran, Richard Zens, Chris Dyer,
Ondˇrej Bojar, Alexandra Constantin, and Evan
Herbst. 2007. Moses: Open source toolkit for
statistical machine translation. In Proceedings
of the 45th Annual Meeting of the Associa-
tion for Computational Linguistics Compan-
ion Volume Proceedings of
the Demo and
Poster Sessions, pages 177–180, Prague, Czech
Republic. Association
for Computational
Linguistics.
Katherine Lee, Orhan Firat, Ashish Agarwal,
Clara Fannjiang, and David Sussillo. 2018.
Hallucinations in neural machine translation.
In NeurIPS Interpretability and Robustness
for Audio, Speech and Language Workshop.
Jiwei Li, Xinlei Chen, Eduard Hovy, and Dan
Jurafsky. 2016a. Visualizing and understand-
ing neural models in NLP. In Proceedings of
the 2016 Conference of the North American
Chapter of the Association for Computational
Linguistics: Human Language Technologies,
pages 681–691, San Diego, California. Associ-
ation for Computational Linguistics. https://
doi.org/10.18653/v1/N16-1082
Urvashi Khandelwal, He He, Peng Qi, and Dan
Jurafsky. 2018. Sharp nearby, fuzzy far away:
How neural language models use context. In
Proceedings of the 56th Annual Meeting of
the Association for Computational Linguis-
Jiwei Li, Will Monroe, and Dan Jurafsky.
2016b. Understanding neural networks through
representation erasure. CoRR, abs/1612.08220.
https://doi.org/10.48550/arXiv.1612
.08220
561
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
Ilya Loshchilov and Frank Hutter. 2019.
Decoupled weight decay regularization. In In-
ternational Conference on Learning Represen-
tations.
Xutai Ma, Ke Li, and Philipp Koehn. 2018. An
analysis of source context dependency in neural
machine translation. In 21st Annual Confer-
ence of the European Association for Machine
Translation, page 189.
Marianna Martindale and Marine Carpuat. 2018.
Fluency over adequacy: A pilot study in mea
suring user trust in imperfect MT. In Proceed-
ings of the 13th Conference of the Association
for Machine Translation in the Americas
(Volume 1: Research Papers), pages 13–25,
Boston, MA. Association for Machine Trans-
lation in the Americas.
Marianna Martindale, Marine Carpuat, Kevin
Duh, and Paul McNamee. 2019. Identifying
fluently inadequate output in neural and sta-
tistical machine translation. In Proceedings of
Machine Translation Summit XVII Volume 1:
Research Track, pages 233–243, Dublin,
Ireland. European Association for Machine
Translation.
Joshua Maynez, Shashi Narayan, Bernd Bohnet,
and Ryan McDonald. 2020. On faithfulness
and factuality in abstractive summarization. In
Proceedings of the 58th Annual Meeting of
the Association for Computational Linguistics,
pages 1906–1919, Online. Association for
Computational Linguistics. https://doi.org
/10.18653/v1/2020.acl-main.173
Gr´egoire Montavon, Alexander Binder, Sebastian
Lapuschkin, Wojciech Samek, and Klaus-
Robert M¨uller. 2019. Layer-wise relevance
propagation: An overview. Explainable AI: In-
terpreting, Explaining and Visualizing Deep
Learning, pages 193–209. https://doi.org
/10.1007/978-3-030-28954-6 10
Tharindu Ranasinghe, Constantin Orasan, and
Ruslan Mitkov. 2020. TransQuest: Translation
quality estimation with cross-lingual transform-
ers. In Proceedings of the 28th International
Conference on Computational Linguistics,
pages 5070–5081, Barcelona, Spain (Online).
International Committee on Computational
Linguistics. https://doi.org/10.18653/v1
/2020.coling-main.445
Vikas Raunak, Arul Menezes, and Marcin
Junczys-Dowmunt. 2021. The curious case of
hallucinations in neural machine translation.
In Proceedings of the 2021 Conference of the
North American Chapter of the Association
for Computational Linguistics: Human Lan-
guage Technologies, pages 1172–1183, Online.
Association for Computational Linguistics.
https://doi.org/10.18653/v1/2021
.naacl-main.92
Ricardo Rei, Craig Stewart, Ana C. Farinha, and
Alon Lavie. 2020a. COMET: A neural frame-
work for MT evaluation. In Proceedings of
the 2020 Conference on Empirical Methods
in Natural Language Processing (EMNLP),
pages 2685–2702, Online. Association for
Computational Linguistics. https://doi.org
/10.18653/v1/2020.emnlp-main.213
Ricardo Rei, Craig Stewart, Ana C. Farinha, and
Alon Lavie. 2020b. Unbabel’s participation
in the WMT20 metrics shared task. In Pro-
ceedings of the Fifth Conference on Machine
Translation, pages 911–920, Online. Associa-
tion for Computational Linguistics.
Mat¯ıs Rikters and Mark Fishel. 2017. Confi-
dence through attention. In Proceedings of
the 16th Machine Translation Summit (MT
Summit 2017). Nagoya, Japan.
Sashank
Santhanam, Behnam Hedayatnia,
Spandana Gella, Aishwarya Padmakumar,
Seokhwan Kim, Yang Liu, and Dilek Hakkani-
Tur. 2021. Rome was built in 1776: A case
study on factual correctness in knowledge-
grounded response generation. CoRR, abs/2110
.05456. https://doi.org/10.48550/arXiv
.2110.05456
Holger Schwenk, Vishrav Chaudhary, Shuo
Sun, Hongyu Gong, and Francisco Guzm´an.
2021. WikiMatrix: Mining 135M parallel sen-
tences in 1620 language pairs from Wikipedia.
the 16th Conference of
In Proceedings of
the European Chapter of
the Association
for Computational Linguistics: Main Volume,
pages 1351–1361, Online. Association for Com-
putational Linguistics. https://doi.org
/10.18653/v1/2021.eacl-main.115
Ruikang Shi, Alvin Grissom II, and Duc Minh
Trinh. 2022. Rare but severe neural machine
562
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
translation errors induced by minimal dele-
tion: An empirical study on Chinese and
English. In Proceedings of the 29th Interna-
tional Conference on Computational Linguis-
tics, pages 5175–5180, Gyeongju, Republic of
Korea. International Committee on Computa-
tional Linguistics.
Karen Simonyan, Andrea Vedaldi, and Andrew
Zisserman. 2014. Deep inside convolutional
networks: Visualising image classification
models and saliency maps. In Proceedings
of the International Conference on Learning
Representations (ICLR). ICLR.
Lucia Specia, Fr´ed´eric Blain, Marina Fomicheva,
Erick Fonseca, Vishrav Chaudhary, Francisco
Guzm´an, and Andr´e F. T. Martins. 2020. Find-
ings of the WMT 2020 shared task on quality
estimation. In Proceedings of the Fifth Confer-
ence on Machine Translation, pages 743–764,
Online. Association for Computational Lin-
guistics.
Lucia Specia, Fr´ed´eric Blain, Marina Fomicheva,
Chrysoula Zerva, Zhenhao Li, Vishrav
Chaudhary, and Andr´e F. T. Martins. 2021.
Findings of the WMT 2021 shared task on
the
quality estimation.
Sixth Conference on Machine Translation,
pages 684–725, Online. Association for Com-
putational Linguistics.
In Proceedings of
Lucia Specia, Marco Turchi, Nicola Cancedda,
Nello Cristianini, and Marc Dymetman. 2009.
Estimating the sentence-level quality of ma-
chine translation systems. In Proceedings of the
13th Annual conference of the European As-
sociation for Machine Translation, Barcelona,
Spain. European Association for Machine
Translation.
Ralf Steinberger, Bruno Pouliquen, Anna
Ignat, Tomaˇz Erjavec,
Widiger, Camelia
Dan Tufis¸, and D´aniel Varga. 2006. The
JRC-Acquis: A multilingual aligned parallel
corpus with 20+ languages. In Proceedings
of the Fifth International Conference on Lan-
guage Resources and Evaluation (LREC’06),
Genoa, Italy. European Language Resources
Association (ELRA).
Katsuhito Sudoh, Kosuke Takahashi,
and
Satoshi Nakamura. 2021. Is this translation
error critical?: Classification-based human and
automatic machine translation evaluation fo-
cusing on critical errors. In Proceedings of the
Workshop on Human Evaluation of NLP Sys-
tems (HumEval), pages 46–55, Online. Asso-
ciation for Computational Linguistics.
Kosuke Takahashi, Yoichi Ishibashi, Katsuhito
Sudoh, and Satoshi Nakamura. 2021. Multi-
lingual machine translation evaluation metrics
fine-tuned on pseudo-negative examples for
WMT 2021 metrics task. In Proceedings of
the Sixth Conference on Machine Transla-
tion, pages 1049–1052, Online. Association for
Computational Linguistics.
Ashish Vaswani, Noam Shazeer, Niki Parmar,
Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,
Łukasz Kaiser, and Illia Polosukhin. 2017. At-
tention is all you need. In Advances in Neural
Information Processing Systems, volume 30,
pages 5998–6008, Long Beach, CA, USA.
Curran Associates, Inc.
Elena Voita, Rico Sennrich, and Ivan Titov.
2021. Analyzing the source and target con-
tributions to predictions in neural machine
translation. In Proceedings of the 59th Annual
Meeting of the Association for Computational
Linguistics and the 11th International Joint
Conference on Natural Language Processing
(Volume 1: Long Papers), pages 1126–1140,
Online. Association for Computational Lin-
guistics. https://doi.org/10.18653
/v1/2021.acl-long.91
Chaojun Wang and Rico Sennrich. 2020. On ex-
posure bias, hallucination and domain shift in
neural machine translation. In Proceedings of
the 58th Annual Meeting of the Association for
Computational Linguistics, pages 3544–3552,
Online. Association for Computational Lin-
guistics. https://doi.org/10.18653/v1
/2020.acl-main.326
Shuo Wang, Zhaopeng Tu, Shuming Shi, and
Yang Liu. 2020. On the inference calibration of
neural machine translation. In Proceedings of
the 58th Annual Meeting of the Association for
Computational Linguistics, pages 3070–3079,
Online. Association for Computational Lin-
guistics. https://doi.org/10.18653/v1
/2020.acl-main.278
Sam Wiseman, Stuart Shieber, and Alexander
Rush. 2017. Challenges in data-to-document
563
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
the 2017
In Proceedings of
generation.
Conference on Empirical Methods in Natu-
ral Language Processing, pages 2253–2263,
Copenhagen, Denmark. Association for Com-
putational Linguistics. https://doi.org/10
.18653/v1/D17-1239
decoded attention and cross-lingual embed-
dings. In Proceedings of the Third Conference
on Machine Translation: Shared Task Papers,
pages 816–821, Belgium, Brussels. Associa-
tion for Computational Linguistics. https://
doi.org/10.18653/v1/W18-6466
Yijun Xiao and William Yang Wang. 2021.
On hallucination and predictive uncertainty in
conditional language generation. In Proceed-
ings of the 16th Conference of the European
Chapter of the Association for Computational
Linguistics: Main Volume, pages 2734–2744,
Online. Association for Computational Lin-
guistics. https://doi.org/10.18653/v1
/2021.eacl-main.236
Weijia Xu and Marine Carpuat. 2018. The
University of Maryland’s Chinese-English neu-
ral machine translation systems at WMT18.
In Proceedings of the Third Conference on
Machine Translation: Shared Task Papers,
pages 535–540, Belgium, Brussels. Associa-
tion for Computational Linguistics. https://
doi.org/10.18653/v1/W18-6431
Elizaveta Yankovskaya, Andre T¨attar, and Mark
Fishel. 2018. Quality estimation with force-
Chrysoula Zerva, Fr´ed´eric Blain, Ricardo
Rei, Piyawat Lertvittayakumjorn, Jos´e G. C.
de Souza, Steffen Eger, Diptesh Kanojia,
Duarte Alves, Constantin Or˘asan, Marina
Fomicheva, Andr´e F. T. Martins, and Lucia
Specia. 2022. Findings of the WMT 2022
shared task on quality estimation. In Proceed-
ings of the Seventh Conference on Machine
Translation, Abu Dhabi. Association for
Computational Linguistics.
Chunting Zhou, Graham Neubig, Jiatao Gu, Mona
Diab, Francisco Guzm´an, Luke Zettlemoyer,
and Marjan Ghazvininejad. 2021. Detecting
in conditional neural
hallucinated content
sequence generation. In Findings of the Asso-
ciation for Computational Linguistics: ACL-
IJCNLP 2021, pages 1393–1404, Online.
Association for Computational Linguistics.
https://doi.org/10.18653/v1/2021
.findings-acl.120
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
5
6
3
2
1
3
1
1
9
3
/
/
t
l
a
c
_
a
_
0
0
5
6
3
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
9
S
e
p
e
m
b
e
r
2
0
2
3
564