Evaluating Document Coherence Modeling

Aili Shen♣, Meladel Mistica♣, Bahar Salehi♣,
Hang Li♦, Timothy Baldwin♣, Jianzhong Qi♣
♣ The University of Melbourne, Australia
♦ AI Lab at ByteDance, China
{aili.shen, misticam, tbaldwin, jianzhong.qi}@unimelb.edu.au
baharsalehi@gmail.com, lihang.lh@bytedance.com

Abstract

While pretrained language models (LMs)
have driven impressive gains over morpho-
syntactic and semantic tasks, their ability to
model discourse and pragmatic phenomena is
less clear. As a step towards a better under-
standing of their discourse modeling capabili-
ties, we propose a sentence intrusion detection
task. We examine the performance of a broad
range of pretrained LMs on this detection task
for English. Lacking a dataset for the task, we
introduce INSteD, a novel intruder sentence
detection dataset, containing 170,000+ doc-
uments constructed from English Wikipedia
and CNN news articles. Our experiments show
that pretrained LMs perform impressively
in in-domain evaluation, but experience a
substantial drop in the cross-domain setting,
indicating limited generalization capacity.
Further results over a novel linguistic probe
dataset show that there is substantial room
for improvement, especially in the cross-
domain setting.

1

Introduction

Rhetorical relations refer to the transition of one
sentence to the next in a span of text (Mann and
Thompson, 1988; Asher and Lascarides, 2003).
They are important as a discourse device that
contributes to the overall coherence, understand-
ing, and flow of the text. These relations span a
tremendous breadth of types, including contrast,
elaboration, narration, and justification. These
connections allow us to communicate coopera-
tively in understanding one another (Grice, 2002;
Wilson and Sperber, 2004). The ability to under-
stand such coherence (and conversely detect
incoherence) is potentially beneficial for down-
stream tasks, such as storytelling (Fan et al., 2019;

621

Hu et al., 2020b), recipe generation (Chandu et al.,
2019), document-level text generation (Park and
Kim, 2015; Holtzman et al., 2018), and essay
scoring (Tay et al., 2018; Li et al., 2018).

However,

there is little work on document
coherence understanding, especially examining
the capacity of pretrained language models (LMs)
to model the coherence of longer documents. To
address this gap, we examine the capacity of
pretrained language models to capture document
coherence, focused around two research questions:
(1) do models truly capture the intrinsic properties
of document coherence? and (2) what types of doc-
ument incoherence can/can’t these models detect?
We propose the sentence intrusion detection
task: (1) to determine whether a document con-
tains an intruder sentence (coarse-grained level);
and (2) to identify the span of any intruder sen-
tence (fine-grained level). We restrict the scope of
the intruder text to a single sentence, noting that in
practice, the incoherent text could span multiple
sentences, or alternatively be sub-sentential.

Existing datasets in document coherence mea-
surement (Chen et al., 2019; Clercq et al., 2014;
Lai and Tetreault, 2018; Mim et al., 2019; Pitler
and Nenkova, 2008; Tien Nguyen and Joty, 2017)
are unsuitable for our task: They are either pro-
hibitively small, or do not specify the span of in-
coherent text. For example, in the dataset of Lai
and Tetreault (2018), each document is assigned a
coherence score, but the span of incoherent text is
not specified. There is thus a need for a large-scale
dataset which includes annotation of the position
of intruder text. Identifying the span of incoherent
text can benefit tasks where explainability and
immediate feedback are important, such as essay
scoring (Tay et al., 2018; Li et al., 2018).

In this work, we introduce a dataset consist-
ing of English documents from two domains:
Wikipedia articles (106K) and CNN news articles

Transactions of the Association for Computational Linguistics, vol. 9, pp. 621–640, 2021. https://doi.org/10.1162/tacl a 00388
Action Editor: Noah Smith. Submission batch: 11/2020; Revision batch: 1/2021; Published 7/2021.
c(cid:4) 2021 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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

2.1 Document Coherence Measurement

Coherence measurement has been studied across
various tasks, such as the document discrimina-
tion task (Barzilay and Lapata, 2005; Elsner et al.,
2007; Barzilay and Lapata, 2008; Elsner and
Charniak, 2011; Li and Jurafsky, 2017; Putra
and Tokunaga, 2017), sentence insertion (Elsner
and Charniak, 2011; Putra and Tokunaga, 2017;
Xu et al., 2019), paragraph reconstruction (Lapata,
2003; Elsner et al., 2007; Li and Jurafsky, 2017;
Xu et al., 2019; Prabhumoye et al., 2020), sum-
mary coherence rating (Barzilay and Lapata
2005; Pitler et al., 2010; Guinaudeau and Strube,
2013; Tien Nguyen and Joty, 2017), readabil-
ity assessment (Guinaudeau and Strube, 2013;
Mesgar and Strube, 2016, 2018), and essay scoring
(Mesgar and Strube, 2018; Somasundaran et al.,
2014; Tay et al., 2018). These tasks differ from
our task of intruder sentence detection as follows.
First, the document discrimination task assigns
coherence scores to a document and its sentence-
permuted versions, where the original document
is considered to be well-written and coherent
and permuted versions incoherent. Incoherence
is introduced by shuffling sentences, while our
intruder sentences are selected from a second doc-
ument, and there is only ever a single intruder
sentence per document. Second, sentence inser-
tion aims to find the correct position to insert a re-
moved sentence back into a document. Paragraph
reconstruction aims to recover the original sen-
tence order of a shuffled paragraph given its first
sentence. These two tasks do not consider sen-
tences from outside of the document of interest.
Third, the aforementioned three tasks are arti-
ficial, and have very limited utility in terms of
real-world tasks, while our task can provide direct
benefit in applications such as essay scoring, in
identifying incoherent (intruder) sentences as a
means of providing user feedback and explainabil-
ity of essay scores. Lastly, in summary coherence
rating, readability assessment, and essay scoring,
coherence is just one dimension of the overall
document quality measurement.

Various methods have been proposed to capture
local and global coherence, while our work aims
to examine the performance of existing pretrained
LMs in document coherence understanding. To
assess local coherence, traditional studies have
used entity matrices, for example, to represent
entity transitions across sentences (Barzilay and

Figure 1: An excerpt of an incoherent document, with
the ‘‘intruder’’ sentence indicated in bold.

(72K). This dataset fills a gap in research pertain-
ing to document coherence: Our dataset is large
in scale, includes both coherent and incoherent
documents, and has mark-up of the position of
any intruder sentence. Figure 1 is an example
document with an intruder sentence. Here, the
highlighted sentence reads as though it should
be an elaboration of the previous sentence, but
clearly exhibits an abrupt change of topic and the
pronoun it cannot be readily resolved.

This paper makes the following contributions:
(1) we propose the sentence intrusion detection
task, and examine how pretrained LMs perform
over the task and hence at document coherence
understanding; (2) we construct a large-scale
dataset from two domains—Wikipedia and CNN
news articles—that consists of coherent and inco-
herent documents, and is accompanied with the
positions of intruder sentences, to evaluate in
both in-domain and cross-domain settings; (3) we
examine the behavior of models and humans, to
better understand the ability of models to model
the intrinsic properties of document coherence;
and (4) we further hand-craft adversarial
test
instances across a variety of linguistic phenomena
to better understand the types of incoherence that
a given model can detect.

2 Related Work

We first review tasks relevant to our proposed
task, then describe existing datasets used in coher-
ence measurement, and finally discuss work on
dataset artefacts and linguistic probes.

622

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Lapata, 2005, 2008). Guinaudeau and Strube
(2013) and Mesgar and Strube (2016) use a graph
to model entity transition sequences. Sentences
in a document are represented by nodes in the
graph, and two nodes are connected if they share
the same or similar entities. Neural models have
also been proposed (Ji and Smith, 2017; Li and
Jurafsky, 2017; Li et al., 2018; Mesgar and Strube,
2018; Mim et al., 2019; Tien Nguyen and Joty,
2017). For example, Tay et al. (2018) capture
local coherence by computing the similarity of
the output of
two LSTMs (Hochreiter and
Schmid-huber, 1997), which they concatenate with
essay representations to score essays. Li et al.
(2018) use multi-headed self-attention to capture
long distance relationships between words, which
are passed to an LSTM layer to estimate essay
coherence scores. Xu et al. (2019) use the average
of local coherence scores between consecutive
pairs of sentences as the document coherence
score.

Another relevant task is disfluency detection
in spontaneous speech transcription (Johnson
and Charniak, 2004; Jamshid Lou et al., 2018).
This task detects the reparandum and repair in
spontaneous speech transcriptions to make the
text fluent by replacing the reparandum with the
repair. Also relevant is language identification
in code-switched text (Adouane et al., 2018a,b;
Mave et al., 2018; Yirmibes¸o˘glu and Eryi˘git,
2018), where disfluency is defined at the language
level (e.g., for a monolingual speaker). Lau et al.
(2015) and Warstadt et al. (2019) predict sentence-
level acceptability (how natural a sentence is).
However, none of tasks are designed to measure
document coherence, although sentence-level
phenomena can certainly impact on document
coherence.

2.2 Document Coherence Datasets

There exist a number of datasets targeted at dis-
course understanding. For example, Alikhani et al.
(2019) construct a multi-modal dataset for under-
standing discourse relations between text and
imagery, such as elaboration and exemplifica-
tion. In contrast, we focus on discourse relations
in a document at the inter-sentential level. The
Penn Discourse Treebank (Miltsakaki et al., 2004;
Prasad et al., 2008) is a corpus of coherent doc-
uments with annotations of discourse connectives
and their arguments, noting that inter-sentential

discourse relations are not always lexically marked
(Webber, 2009).

The most relevant work to ours is the dis-
course coherence dataset of Chen et al. (2019),
which was proposed to evaluate the capabilities
of pretrained LMs in capturing discourse context.
This dataset contains documents (18K Wikipedia
articles and 10K documents from the Ubuntu IRC
channel) with fixed sentence length, and labels
documents only in terms of whether they are
incoherent, without considering the position of
the incoherent sentence. In contrast, our dataset:
(1) provides more fine-grained information (i.e.,
the sentence position); (2) is larger in scale (over
170K documents); (3) contains documents of
varying length; (4) incorporates adversarial filter-
ing to reduce dataset artefacts (see Section 3); and
(5) is accompanied with human annotation over
the Wikipedia subset, allowing us to understand
behavior patterns of machines and humans.

2.3 Dataset Artefacts

Also relevant to this research is work on removing
artefacts in datasets (Zellers et al., 2019; McCoy
et al., 2019; Zellers et al., 2018). For example,
based on analysis of the SWAG dataset (Zellers
et al., 2018), Zellers et al. (2019) find artefacts
such as stylistic biases, which correlate with the
document labeling and mean that naive models
are able to achieve abnormally high results. Sim-
ilarly, McCoy et al. (2019) examine artefacts in
an NLI dataset, and find that naive heuristics that
are not directly related to the task can perform
remarkably well. We incorporate the findings of
such work in the construction of our dataset.

2.4 Linguistic Probes

Adversarial training has been used to craft adver-
sarial examples to obtain more robust models,
either by manipulating model parameters (white-
box attacks) or minimally editing text at
the
character/word/phrase level (black-box attacks).
For example, Papernot et al. (2018) provide a ref-
erence library of adversarial example construction
techniques and adversarial training methods.

As we aim to understand the linguistic prop-
erties that each model has captured, we focus on
black-box attacks (Sato et al., 2018; Cheng et al.,
2020; Liang et al., 2018; Yang et al., 2020;
Samanta and Mehta, 2017). For example, Samanta
and Mehta (2017) construct adversarial examples
for sentiment classification and gender detection

623

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

by deleting, replacing, or inserting words in the
text. For a comprehensive review of such studies,
see Belinkov and Glass (2019).

There is also a rich literature on exploring what
kinds of linguistic phenomena a model has learned
(Hu et al., 2020a; Hewitt and Liang, 2019; Hewitt
and Manning, 2019; Chen et al., 2019; McCoy
et al., 2019; Conneau et al., 2018; Gulordava
et al., 2018; Peters et al., 2018; Tang et al., 2018;
Blevins et al., 2018; Wilcox et al., 2018; Kuncoro
et al., 2018; Tran et al., 2018; Belinkov et al.,
2017). The basic idea is to use learned represen-
tations to predict linguistic properties of interest.
Example linguistic properties are subject–verb
agreement or syntactic structure, while represen-
tations can be word or sentence embeddings. For
example, Marvin and Linzen (2018) construct
minimal sentence pairs, consisting of a grammat-
ical and ungrammtical sentence, to explore the
capacity of LMs in capturing phenomena such
as subject–verb agreement, reflexive anaphora,
and negative polarity items. In our work, we
hand-construct intruder sentences which result in
incoherent documents, based on a broad range of
linguistic phenomena.

3 Dataset Construction

3.1 Dataset Desiderata

To construct a large-scale, low-noise dataset that
truly tests the ability of systems to detect intruder
sentences, we posit five desiderata:

1. Multiple sources: The dataset should not
be too homogeneous in terms of genre or
domain, and should ideally test the ability of
models to generalize across domain.

2. Defences against hacking: Human annota-
tors and machines should not be able to hack
the task and reverse-engineer the labels by
sourcing the original documents.

3. Free of artefacts: The dataset should be free
of artefacts, that allow naive heuristics to
perform well.

4. Topic consistency: The intruder sentence,
which is used to replace a sentence from a
coherent document to obtain an incoherent
document, should be relevant to the topic of
the document, to focus the task on coherence
and not simple topic detection.

5. KB-free: Our goal is NOT to construct a
fact-checking dataset; the intruder sentence
should be determinable based on the content
of the document, without reliance on external
knowledge bases or fact-checking.

3.2 Data Sources

We construct a dataset
from two sources—
Wikipedia and CNN—which differ in style and
genre, satisfying the first desideratum. Similar to
WikiQA (Yang et al., 2015) and HotpotQA (Yang
et al., 2018), we represent a Wikipedia document
by its summary section (i.e., the opening para-
graph), constraining the length to be between 3
and 8 sentences. For CNN, we adopt the dataset of
Hermann et al. (2015) and Nallapati et al. (2016),
which consists of over 100,000 news articles. To
obtain documents with sentence length similar to
those from Wikipedia, we randomly select the
first 3–8 sentences from each article.

To defend against dataset hacks1 that could
expose the labels of the test data (desideratum
2), the Wikipedia test set is randomly sampled
from 37 historical dumps of Wikipedia, where the
selected article has a cosine similarity less than the
historical average of 0.72 with its online version.2
For the training set, we remove this require-
ment and randomly select articles from different
Wikipedia dumps, namely, the articles in the train-
ing set might be the same as their current online
version. For CNN, we impose no such limitations.

3.3 Generating Candidate Positive Samples

We consider the original documents to be coher-
ent. We construct incoherent documents from half
of our sampled documents as follows (satisfying
desiderata 3–5):

1. Given a document D, use bigram hashing
and TF-IDF matching (Chen et al., 2017) to
retrieve the top-10 most similar documents
from a collection of documents from the same
domain, where D is the query text. Let the
set of retrieved documents be RD.

2. Randomly choose a non-opening sentence
S from document D, to be replaced by a
sentence candidate generated later. We do

1Deliberate or otherwise, e.g., via pre-training on the same

version of Wikipedia our dataset was constructed over.

2This threshold was determined by calculating the average
TF-IDF-weighted similarity of the summary section for
documents in all 37 dumps with their current online versions.

624

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

not replace the opening sentence as it is
needed to establish document context.

Source

#docs

Wikipedia
CNN

106,352 (46%)
72,670 (49%)

avg.
#sents
5±1
5±1

avg.
#tokens
126±24
134±32

3. For each document D(cid:5) ∈ RD, randomly
select one non-opening sentence S(cid:5) ∈ D(cid:5)
as an intruder sentence candidate.

4. Calculate the TF-IDF-weighted cosine simi-
larity between sentence S and each candidate
S(cid:5). Remove any candidates with similarity
scores ≥ 0.6, to attempt to generate a KB-free
incoherence.

5. Replace sentence S with each low-similarity
candidate S(cid:5), and use a fine-tuned XLNet-
Large model (Yang et al., 2019) to check
whether it is easy for XLNet-Large to detect
(see Section 5). For documents with both
easy and difficult sentence candidates, we
randomly sample from the difficult sentence
candidates; otherwise, we randomly choose
from all the sentence candidates.

The decision to filter out sentence candidates
with similarity ≥ 0.6 was based on the observa-
tion that more similar sentences often led to the
need for world knowledge to identify the intruder
sentence (violating the fifth desideratum). For
example, given It is the second novel in the first
of three trilogies about Bernard Samson, …, a
candidate intruder sentence candidate with high
similarity is It is the first novel in the first of three
trilogies about Bernard Samson ….

We also trialed other ways of generating inco-
herent samples, such as using sentence S from
document D as the query text to retrieve docu-
ments, and adopting a 2-hop process to retrieve
relevant documents. We found that these methods
resulted in documents that can be identified by
the pretrained models easily.

4 Dataset Analysis

4.1 Statistics of the Dataset

The process described in Section 3 resulted in
106,352 Wikipedia documents and 72,670 CNN
documents, at an average sentence length of 5
in both cases (see Table 1). The percentages of
positive samples (46% and 49%, respectively) are
slightly less than 50% due to our data generation
constraints (detailed in Section 3.3), which can
lead to no candidate intruder sentence S(cid:5) being
generated for original sentence S. We set aside 8%

Table 1: Dataset statistics for INSteD. Numbers
in parentheses are percentages of
incoherent
documents.

of Wikipedia (which we manually tag, as detailed
in Section 4.5) and 20% of CNN for testing.

4.2 Types of Incoherence

To better understand the different types of issues
resulting from our automatic method, we sampled
100 (synthesized) incoherent documents from
Wikipedia and manually classified the causes
of incoherence according to three overlapping
categories (ranked in terms of expected ease of
detection): (1) information structure inconsis-
tency (a break in information flow); (2) logical
inconsistency (a logically inconsistent world state
is generated, such as someone attending school
before they were born); and (3) factual incon-
sistency (where the intruder sentence is factually
incorrect). See Table 2 for a breakdown across the
categories, noting that a single document can be
incoherent across multiple categories. Information
structure inconsistency is the most common form
of incoherence, followed by factual inconsistency.
The 35% of documents with factual inconsisten-
cies break down into 8% (overall) that have other
types of incoherence, and 27% that only have
a factual inconsistency. This is an issue for the
fifth desideratum for our dataset (see Section 3.1),
motivating the need for manual checking of the
dataset
to determine how readily the intruder
sentence can be detected.3

4.3 Evaluation Metrics

We base evaluation of intruder sentence detection
at both the document and sentence levels:

• document level: Does the document contain
an intruder sentence? This is measured based
on classification accuracy (Acc), noting that
the dataset
the
document level (see Table 1). A prediction

is relatively balanced at

3We keep these documents in the dataset, as it is beyond

the scope of this work to filter these documents out.

625

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Incoherence

Information
structure
inconsistency

Logical
inconsistency

Factual
inconsistency

Example

He is currently the senior pastor at Sovereign Grace Church of Louisville. The Church
is led by Senior Pastor Ray Johnston, Senior Pastor Curt Harlow and Senior Pastor
Andrew McCourt, and Senior Pastor Lincoln Brewster. Under Mahaney’s leadership,
Sovereign Grace Church of Louisville is a member of Sovereign Grace Churches.
Michael David, born September 22, 1954, is an American-born American painter. From
1947–1949 he attended the Otis Art Institute, from 1947 to 1950 he also attended the
Art Center College of Design in Los Angeles, and in 1950 the Chouinard Art Institute.
The Newport Tower has 37 floors. It is located on the beachfront on the east side of
Collins Avenue between 68th and 69th Streets. The building was developed by Melvin
Simon & Associates in 1990.

%

58

26

35

Table 2: Types of document incoherence in Wikipedia. Text in bold indicates the intruder sentence.

is ‘‘correct’’ if at least one sentence/none of
the sentences is predicted to be an intruder.

• sentence level: Is a given (non-opening) sen-
tence an intruder sentence? This is measured
based on F1, noting that most (roughly 88%)
sentences are non-intruder sentences.

4.4 Testing for Dataset Artefacts

To test for artefacts, we use XLNet-Large (Yang
et al., 2019) to predict whether each non-opening
sentence is an intruder sentence, in complete iso-
lation of its containing document (i.e., as a stand-
alone sentence classification task). We compare
the performance of XLNet-Large with a majority-
class baseline (‘‘Majority-class’’) that predicts all
sentences to be non-intruder sentences (i.e., from
the original document), where XLNet-Large is
fine-tuned over the Wikipedia/CNN training set,
and tested over the corresponding test set.

For Wikipedia, XLNet-Large obtains an Acc
of 55.4% (vs. 55.1% for Majority-class) and F1
of 3.4% (vs. 0.0% for Majority-class). For CNN,
the results are 50.8% and 1.2%, respectively (vs.
51.0% and 0.0% resp. for Majority-class). These
results suggest that the dataset does not contain
obvious artefacts, at least for XLNet-Large. We
also experiment with a TF-IDF weighted bag-
of-words logistic regression model, achieving
slightly worse results than XLNet-Large (Acc =
55.1%, F1 = 0.05% for Wikipedia, and Acc =
50.6%, F1 = 0.3% for CNN).4

4For RoBERTa-Large (Section 6.1), there were also no
obvious artefacts observed in the standalone sentence setting:
Acc = 55.7% and F1 = 5.3% over Wikipedia, and Acc =
51.3% and F1 = 4.3% over CNN.

4.5 Human Verification

We performed crowdsourcing via Amazon
Mechanical Turk over the Wikipedia test data
to examine how humans perform over this task.
Each Human Intelligence Task (HIT) contained
5 documents and was assigned to 5 workers. For
each document, the task was to identify a single
sentence that ‘‘creates an incoherence or break in
the content flow’’, or in the case of no such sen-
tence, ‘‘None of the above’’, indicating a coherent
document. In the task instructions, workers were
informed that there is at most one intruder sen-
tence per document, and were not able to select
the opening sentence. Among the 5 documents
for each HIT, there was one incoherent document
from the training set, which was pre-identified as
being easily detectable by an author of the paper,
and acts as a quality control item. We include doc-
uments where at least 3 humans assign the same
label as our test dataset (90.3% of the Wikipedia
test dataset), where all the results are reported over
these documents, if not specified.5 Payment was
calibrated to be above Australian minimum wage.
Figure 2 shows the distribution of instances
where different numbers of workers produced
the correct answer (the red bar). For example,
for 6.2% of instances, 2 of 5 workers annotated
correctly. The blue bars indicate the proportion of
incoherent documents where the intruder sentence
was correctly detected by the given number of
annotators (e.g., for 9.3% of incoherent docu-
ments, only 2 of 5 workers were able to identify
the intruder sentence correctly). Humans tend to
agree with each other over coherent documents,
as indicated by the increasing percentages for

5Different people may have different

thresholds in
considering a document to be incoherent, but this is beyond
the scope of our work.

626

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Bi-LSTM with average-pooling; word embed-
dings are initialized as with BoW.

InferSent: Generate representations for the sen-
tence and document with InferSent (Conneau
et al., 2017), and concatenate the two; InferSent
is based on a Bi-LSTM with a max-pooling layer,
trained on SNLI (Bowman et al., 2015).

Skip-Thought: Generate representations
for
the sentence and document with Skip-Thought
(Kiros et al., 2015), and concatenate the two; Skip-
Thought is an encoder–decoder model where the
encoder extracts generic sentence embeddings and
the decoder reconstructs surrounding sentences of
the encoded sentence.

BERT: Generate representations for the concate-
nated sentence and document with BERT (Devlin
et al., 2019), which was pretrained on the tasks
of masked language modeling and next sentence
prediction over Wikipedia and BooksCorpus
(Zhu et al., 2015); we experiment with both BERT-
Large and BERT-Base (the cased versions).

and

sentence

RoBERTa: Generate representations for
the
concatenated
document with
RoBERTa (Liu et al., 2019), which was pre-
trained on the task of masked language modeling
(dynamically masking) and each input consisting
of continuous sentences from the same document
or multiple documents (providing broader context)
over Cc-news, OpenWebTextCorpus, and STO-
RIES (Trinh and Le, 2018), in addition to the same
data BERT was pretrained on; we experiment with
both RoBERTa-Large and RoBERTa-Base.

for

and

ALBERT: Generate representations
the
concatenated
document with
sentence
ALBERT (Lan et al., 2020), which was pre-
trained over the same dataset as BERT but
replaces the next sentence prediction objective
with a sentence-order prediction objective,
to
model document coherence; we experiment with
both ALBERT-Large and ALBERT-xxLarge.

XLNet: Generate representations for the con-
catenated sentence and document with XLNet
(Yang et al., 2019), which was pretrained using a
permutation language modeling objective over
datasets
including Wikipedia, BooksCorpus,
Giga5 (Parker et al., 2011), ClueWeb 2012-B
(Callan et al., 2009), and Common Crawl; we
experiment with both XLNet-Largeand XLNet-
Base (the cased versions). Although XLNet-Large

627

Figure 2: Distribution of instances where different
numbers of humans produce correct answers. Note that
the red bars indicate distributions over all documents
and the blue bars indicate distributions over incoherent
documents.

red bars but decreasing percentages for blue bars
across the x-axis. Intruder sentences in incoherent
documents, however, are harder to detect. One
possible explanation is that the identification of
intruder sentences requires fact-checking, which
workers were instructed not to do (and base their
judgment only on the information in the provided
document); another reason is that intruder sen-
tences disrupt local incoherence with neighboring
sentences, creating confusion as to which is the
intruder sentence (with many of the sentence-level
mis-annotations being off-by-one errors).

5 Models

We model intruder sentence detection as a binary
classification task: Each non-opening sentence in
a document is concatenated with the document,
and a model
is asked to predict whether the
sentence is an intruder sentence to the document.
Our focus is on the task, dataset, and how
existing models perform at document coherence
prediction rather than modeling novelty, and we
thus experiment with pre-existing pre-trained
models. The models are as follows, each of which
is fed into an MLP layer with a softmax output.

BoW: Average the word embeddings for the
combined document (sentence + sequence of
sentences in the document), based on pretrained
300D GloVe embeddings trained on a 840B-token
corpus (Pennington et al., 2014).

Bi-LSTM: Feed the sequence of words in the
into a single-layer 512D
combined document

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

→ Wikipedia

→ CNN

Wiki→Wiki

CNN→Wiki

CNN→CNN

Wiki→CNN

Acc (%)

F1 (%) Acc (%)

F1 (%)

Acc (%)

F1 (%) Acc (%)

F1 (%)

Majority-class
BoW
Bi-LSTM
InferSent
Skip-Thought
BERT-Base
BERT-Large
XLNet-Base
XLNet-Large
RoBERTa-Base
RoBERTa-Large
ALBERT-Large
ALBERT-xxLarge

ALBERT-xxLarge-freeze

Human

57.3
57.3
56.2
57.3
57.3
65.3
67.0
67.8
72.9
69.5
76.1
70.7
81.7

57.3

66.6

0.0
0.0
12.7
0.0
0.0
35.7
39.6
45.0
55.4
47.0
59.8
49.6
71.5

0.0

35.9

57.3
57.3
57.3
57.3
57.3
61.2
64.0
62.2
62.8
63.2
63.7
63.8
66.6

N/A

66.6

0.0
0.0
0.0
0.0
0.0
21.1
29.1
22.4
22.2
26.1
24.6
24.9
33.2

N/A

35.9

50.6
50.6
51.7
50.6
50.6
80.8
82.4
91.2
96.9
92.5
96.0
93.4
96.9

50.6

0.0
0.0
25.1
0.0
0.0
71.6
74.8
86.6
95.0
88.8
94.5
90.8
95.9

0.3

50.6
50.6
50.2
50.6
50.6
57.0
61.5
64.0
80.7
77.6
88.3
72.6
89.1

N/A

0.0
0.0
3.0
0.0
0.0
23.5
35.9
43.3
73.8
68.1
83.5
61.5
86.7

N/A

74.0

57.8

Table 3: Experimental results over Wikipedia and CNN, in both in-domain and cross-domain settings.
Acc is at the document level and F1 is at the sentence level.

is used in removing data artefacts when selecting
the intruder sentences, our experiments suggest
that the comparative results across models (with
or without artefact filtering) are robust.

6 Experiments

6.1 Preliminary Results

In our first experiments, we train the various
models across both Wikipedia and CNN, and
evaluate them in-domain and cross-domain. We
are particularly interested in the cross-domain
setting, to test the true ability of the model to
detect document incoherence, as distinct from
overfitting to domain-specific idiosyncrasies. It
is also worth mentioning that BERT, RoBERTa,
ALBERT, and XLNet are pretrained on multi-
sentence Wikipedia data, and have potentially
memorised sentence pairs, making in-domain
experiments problematic for Wikipedia in partic-
ular. Also of concern in applying models to the
automatically generated data is that it is entirely
possible that an intruder sentence is undetectable
to a human, because no incoherence results from
the sentence substitution (bearing in mind that
only 58% of documents in Table 2 contained
information structure inconsistencies).

From Table 3, we can see that the simpler models
(BoW, Bi-LSTM, InferSent, and Skip-Thought)
perform only at the level of Majority-class at the

document level, for both Wikipedia and CNN.
At the sentence level (F1), once again the models
the level of Majority-class
perform largely at
(F1 = 0.0), other than Bi-LSTM in-domain for
Wikipedia and CNN. In the final row of the
table, we also see that humans are much better at
detecting whether documents are incoherent (at
the document level) than identifying the position
of intruder sentences (at the sentence level), and
that in general, human performance is low. This
is likely the result of the fact that there are only
58% of documents in Table 2 containing informa-
tion structure inconsistencies. We only conducted
crowdsourcing over Wikipedia due to budget lim-
itations and the fact that the CNN documents are
available online, making dataset hacks possible.6
Among the pretrained LMs, ALBERT-xxLarge
achieves the best performance over Wikipedia and
CNN, at both the document and sentence levels.
Looking closer at the Wikipedia results, we find
that BERT-Large achieves a higher precision than
XLNet-Large (71.0% vs. 60.3%), while XLNet-
Large achieves a higher recall (51.3% vs. 27.4%).
ALBERT-xxLarge achieves a precision higher
than BERT-Large (79.7%) and a recall higher
than XLNet-Large (64.9%), leading to the overall

6To have a general idea about the difficulty of the CNN
dataset, one of the authors annotated 100 documents (50
coherent and 50 incoherent documents), randomly sampled
from the test set.

628

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

best performance. Over CNN, ALBERT-xxLarge,
RoBERTa-Large, and XLNet-Large achieve high
precision and recall (roughly 93.0% to 97%).7
The competitive results for ALBERT-xxLarge
over Wikipedia and CNN result from the pre-
training strategies, especially the sentence-order
prediction loss capturing document coherence in
isolation, different from next sentence prediction
loss which conflates topic prediction and coher-
ence prediction in a lower-difficulty single task.
The performance gap for ALBERT, RoBERTa,
and XLNet between the base and large models are
bigger than that of BERT, suggesting that they
benefit from greater model capacity.8

We also examine how pretrained LMs perform
with only the classifier parameters being updated
during training. Here, we focus on exclusively
on ALBERT-xxLarge, given its superiority. As
shown in Figure 3, the pretrained LM ALBERT-
xxLarge is unable to different coherent documents
from incoherent ones, resulting into random guess,
although it considers document coherence during
pretraining. This indicates the necessity of fine-
tuning LMs for document coherent understanding.
Looking to the cross-domain results, again,
ALBERT-xxLarge achieves the best performance
over both Wikipedia and CNN. The lower results
for RoBERTa-Large and XLNet-Large over
Wikipedia may be due to both RoBERTa and
XLNet being pretrained over newswire docu-
ments, and fine-tuning over CNN reducing the
capacity of the model to generalize. ALBERT and
BERT do not suffer from this as they are not pre-
trained over newswire documents. The substantial
drop between the in- and cross-domain settings
for ALBERT, RoBERTa, XLNet, and BERT
indicates that the models have limited capacity
to learn a generalized representation of document
coherence, in addition to the style differences
between Wikipedia and CNN.

7The higher performance for all models/humans over the
CNN dataset indicates that it is easier for models/humans to
identify the presence of intruder sentences. This is can be
explained by the fact that a large proportion of documents
include named entities, making it easier to detect the intruder
sentences. In addition, the database used to retrieve candidate
intruder sentences is smaller compared to that of Wikipedia.
8We also performed experiments where the models were
allowed to predict the first sentence as the intruder sentence.
As expected, model performance drops, e.g., F1 of XLNet-
Large drops from 55.4% to 47.9%, reflecting both the
increased complexity of the task and the lack of (at least) one
previous sentence to provide document context.

Wiki→Wiki

Ubuntu→Wiki

Majority-class
ALBERT-xxLarge
Human

50.0
96.8
98.0

50.0
53.1
98.0

Ubuntu→Ubuntu Wiki→Ubuntu

Majority-class
ALBERT-xxLarge
Human

50.0
58.1
74.0

50.0
58.7
74.0

Table 4: Acc for the dataset of Chen et al. (2019).

6.2 Results over the Existing Dataset

We also examine how ALBERT-xxLarge per-
forms over the coarse-grained dataset of Chen
et al. (2019), where 50 documents from each do-
main were annotated by a native English speaker.
Performance is measured at the document level
only, as the dataset does not include indication of
which sentence is the intruder sentence. As shown
in Table 4, ALBERT-xxLarge achieves an Acc of
96.8% over the Wikipedia subset, demonstrating
that our Wikipedia dataset is more challenging
(Acc of 81.7%) and also underlining the utility of
adversarial filtering in dataset construction. Given
the considerably lower results, one could conclude
that Ubuntu is a good source for a dataset. How-
ever, when one of the authors attempted to perform
the task manually, they found the document-level
task to be extremely difficult as it relied heavily
on expert knowledge of Ubuntu packages, much
more so than document coherence understanding.
In the cross-domain setting, there is a substan-
tial drop over the Wikipedia dataset, which can be
explained by ALBERT-Large failing to generate
a representation of document coherence from the
Ubuntu dataset, due to the high dependence on
domain knowledge as described above, result-
ing in near-random results. The cross-domain
results for ALBERT-xxLarge over Ubuntu are
actually marginally higher than the in-domain
results but still close to random, suggesting that
the in-domain model isn’t able to capture either
document coherence or domain knowledge, and
underlining the relatively minor role of coherence
for the Ubuntu dataset.

6.3 Performance on Documents of Different

Difficulty Levels

One concern with our preliminary experiments
was whether the intruder sentences generate gen-
uine incoherence in the information structure
of the documents. We investigate this question

629

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Humans

# −intruder docs # +intruder docs

Coherent
Incoherent

1385
11

177
404

Table 5: Statistics over documents where all
5 humans agree, where −intruder/+intruder in-
dicates the documents without/with an intruder
sentence.

that

which were annotated as incoherent by all
annotators, we find out
there is a break
in information flow due to references or urls,
even though there is no intruder sentence. For
documents with an intruder sentence (+ intruder),
where humans disagree with the gold-standard
(humans perceive the documents as coherent or
the position of the intruder sentence to be other
than the actual intruder sentence), we find that 98%
of the documents are considered to be coherent.
We randomly sampled 100 documents from these
documents and examined whether the intruder
sentence results in a break in information flow.
We find that fact-checking is needed to identify
the intruder sentence for 93% of the documents.10
Table 6 shows the performance over
the
Wikipedia documents that are annotated con-
sistently by all 5 annotators (from Table 5).
Consistent with the results
from Table 3,
ALBERT-xxLarge achieves the best performance
both in- and cross-domain. To understand the
different behaviors of humans and ALBERT-
xxLarge we analyze documents which only
humans got correct, only ALBERT-xxLarge got
correct, or neither humans nor ALBERT-xxLarge
got correct, as follows:

1. Humans only: 7 incoherent (+intruder) and

73 coherent (−intruder) documents

2. ALBERT-xxLarge only: 181 incoherent
(+intruder) (of which we found 97% to
require fact-checking11) and 9 coherent
(−intruder) documents (of which 8 contain
urls/references, which confused humans)

10Here, the high percentage of incoherent documents with
factual inconsistencies does not necessarily point to a high
percentage of factual inconsistency in the overall dataset, as
humans are more likely to agree with the gold-standard for
coherent documents.

11There are 4 documents that humans identify as incoherent
based on the wrong intruder sentence, due to the intruder
sentence leading to a misleading factual inconsistency.

630

Figure 3: ALBERT-xxLarge vs. humans.

by breaking down the results over the best-
performing model
(ALBERT-xxLarge) based
on the level of agreement between the human
annotations and the generated gold-standard, for
Wikipedia. The results are in Figure 3, where
the x-axis denotes the number of annotators who
agree with the gold-standard: for example, ‘‘2’’
indicates that 2 of 5 annotators were able to assign
the gold-standard labels to the documents.

Our assumption is that the incoherent docu-
ments which humans fail to detect are actually not
perceptibly incoherent,9 and that any advantage
for the models over humans for documents with
low-agreement (with respect to the gold-standard)
is actually due to dataset artefacts. At the docu-
ment level (Acc), there is reasonable correlation
between model and human performance (i.e., the
model struggles on the same documents as the
humans). At the sentence level (F1), there is less
discernible difference in model performance over
documents of varying human difficulty.

6.4 Analysis over Documents with High

Human Agreement

To understand the relationship between human-
assigned labels and the gold-standard, we further
examine documents where all 5 annotators agree,
noting that human-assigned labels can poten-
tially be different from the gold-standard here.
Table 5 shows the statistics of humans over
these documents, with regard to whether there
is an intruder sentence in the documents. En-
couragingly, we can see that humans tend to
agree more over coherent documents (documents
without any intruder sentences) than incoherent
documents (documents with an intruder sentence).
Examining the 11 original coherent documents

9Although the intruder sentence may lead to factual errors,

the annotators were instructed not to do fact checking.

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Wiki→Wiki

CNN→Wiki

Acc (%) F1 (%) Acc (%) F1 (%)

Majority-class
BERT-Large
XLNet-Large
RoBERTa-Large
ALBERT-xxLarge
Human

70.6
76.7
79.1
82.0
85.9
79.5

0.0
42.0
57.0
59.6
68.8
45.4

70.6
75.4
76.6
77.3
78.8
79.5

0.0
36.9
35.4
37.4
42.9
45.4

Table 6: Results over documents annotated con-
sistently by all 5 annotators, where annotations
can be the same as or different from gold-
standard.

3. Neither humans nor models: 223 incoherent
(+intruder) (of which 98.2% and 77.1%
were predicted to be coherent by humans
and ALBERT-xxLarge, respectively, and for
the remainder, the wrong intruder sentence
was identified) and 2 coherent (−intruder)
documents (both of which were poorly
organised, confusing allcomers)

Looking over the incoherent documents that
require fact-checking, no obvious differences are
discernible between the documents that ALBERT-
xxLarge predicts correctly and those it misses.
Our assumption here is that ALBERT-xxLarge is
biased by the pretraining dataset, and that many
of the cases where it makes the wrong prediction
are attributable to mismatches between the text
in our dataset and the Wikipedia version used in
pretraining the model.

6.5 Question Revisited

Q1: Do models truly capture the intrinsic
properties of document coherence?

A: It is certainly true that models that incorpo-
rate a more explicit notion of document coher-
ence into pretraining (e.g., ALBERT) tend to
perform better. In addition, larger-context mod-
els (RoBERTa) and robust
training strategies
(XLNet) during pretraining are also beneficial for
document coherent understanding. This suggests
a tentative yes, but there were equally instances
of strong disagreement with human intuitions
and model predictions for the better-performing
models and evidence to suggest that the models
were performing fact-checking at the same time
as coherence modeling.

Q2: What types of document incoherence

can/can’t these models detect?

A: Over incoherent documents resulting from
fact inconsistencies, where humans tend to fail,
the better-performing models can often make cor-
rect predictions; over incoherent documents with
information structure or logical inconsistencies
which humans can easily detect, ALBERT-Large,
RoBERTa-Large, and XLNet-Large achieve an
Acc ≥ 87%, showing that they can certainly
capture information structure and logical incon-
sistencies to a high degree. That said, the fact
that they misclassify clearly coherent documents
as incoherent suggests that are in part lacking in
their ability to capture document coherence. We
thus can conclude that they can reliably identify
intruder sentences which result in a break in infor-
mation structure or logical flow, but are imperfect
models of document coherence.

7 Linguistic Probes

To further examine the models, we constructed a
language probe dataset.

7.1 Linguistic Probe Dataset Construction

We handcrafted adversarial instances based on a
range of linguistic phenomena that generate infor-
mation structure inconsistencies. In constructing
such a dataset, minimal modifications were made
to the original sentences, to isolate the effect
of the linguistic probe. For each phenomenon,
we hand-constructed roughly 100 adversarial
instances by modifying intruder sentences in
incoherent Wikipedia test documents that were
manually pre-filtered for ease of detection/lack
of confounding effects in the original text. That
is, the linguistic probes for the different phe-
nomena were manually added to incoherent test
documents, within intruder sentences; our interest
here is whether the addition of the linguistic
probes makes it easier for the models to detect
the incoherence. Importantly, we do not provide
any additional training data, meaning there is
no supervision signal specific to the phenomena.
There are roughly 8×100 instances in total,12 with
the eight phenomena being:

1. gender pronoun flip (Gender), converting a
pronoun to its opposite gender (e.g., she →
he);

12There are 100 instances for each phenomenon except for
Demonstrative, where there were only 95 instances in the
Wikipedia test data with singular demonstratives.

631

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

2. animacy downgrade (Animacy↓), downgrad-
ing pronouns and possessive determiners to
their inanimate versions (e.g., she/he/her/him
→ it, and her/his → its);

3. animacy upgrade (Animacy↑), upgrading
pronouns and possessive determiners to
it →
their
third person version (e.g.,
she/he/her/him, and its → her/his);

4. singular demonstrative flip (Demonstrative),
converting singular demonstratives to plural
ones (e.g., this → these and that → those);
5. conjunction flip (Conjunction), converting
conjunctions to their opposites (e.g., but
→ and therefore, and → but, although →
therefore, and vice versa);

6. past tense flip (Past to Future), converting
past to future tense (e.g., was → will be and
led → will lead);

7. sentence negation (Negation), negating the
sentence (e.g., He has [a] . . . warrant … →
He doesn’t have [a] . . . warrant …);

8. number manipulation (Number), changing
numbers to implausible values (e.g., He
served as Chief Operating Officer . . . from
2002 to 2005 → He served as Chief
Operating Officer . . . from 200 BCE to 201
BCE and Line 11 has a length of 51.7 km and
a total of 18 stations. → Line 11 has a length
of 51.7 m and a total of 1.8 stations.).

All the probes generate syntactically correct
sentences, and the first four generally lead to
sentences that are also semantically felicitous,
with the incoherence being at the document level.
For example, in He was never convicted and was
out on parole within a few years, if we replace he
with she, the sentence is felicitous, but if the focus
entity in the preceding and subsequent sentences
is a male, the information flow will be disrupted.
The last four language probes are crafted to
explore the capacity of a model to capture com-
monsense reasoning, in terms of discourse re-
lationships,
tense and polarity awareness, and
understanding of numbers. For Conjunction,
we only focus on explicit connectives within
a sentence. For Past
there can be
intra-sentence inconsistency if there are time-
specific signals, failing which broader document
context is needed to pick up on the tense flip.
Similarly for Negation and Number, the change

to Future,

can lead to inconsistency either intra- or inter-
sententially. For example, He did not appear in
more than 400 films between 1914 and 1941 … is
intra-sententially incoherent.

7.2 Experimental Results

Table 7 lists the performance of pretrained LMs at
recognising intruder sentences within incoherent
documents, with and without the addition of the
respective linguistic probes.13 For a given model,
we break down the results across probes into
two columns: The first column (‘‘F1’’) shows
the sentence-level performance over the original
intruder sentence (without the addition of the lin-
guistic probe), and the second column (‘‘ΔF1’’)
shows the absolute difference in performance with
the addition of the linguistic probe. Our expec-
tation is that results should improve on average
with the inclusion of the linguistic probe (i.e.,
ΔF1 values should be positive), given that we
have reinforced the incoherence generated by the
intruder sentence.

All models achieve near-perfect results with
Gender linguistic probes (i.e., the sum of F1 and
ΔF1 is close to 100), and are also highly success-
ful at detecting Animacy mismatches and Past to
Future (the top half of Table 7). For the probes
in the bottom half of the table, none of the three
models except ALBERT-xxLarge performs par-
ticularly well, especially for Demonstrative. For
each linguistic probe, we observe that the pre-
trained LMs can more easily detect incoherent
text with the addition of these lexical/gram-
matical inconsistencies (except for XLNet-Large
and ALBERT-xxLarge over Demonstrative and
ALBERT-xxLarge over Conjunction).

In the cross-domain setting, the overall per-
formance of XLNet-LargeCNN and ALBERT-
xxLargeCNN drops across all linguistic probes, but
the absolute gain through the inclusion of the
linguistic probe is almost universally larger,
that while domain differences hurt
suggest
the models, they are attuned to the impact of
linguistic probes on document coherence and
thus learning some more general properties of
(in)coherence. On the other hand,
document
BERT-LargeCNN (over Gender, Animacy↓, and
Animacy↑) and RoBERTa-LargeCNN (Gender and
Animacy↑) actually perform better than in-domain.
RoBERTa-LargeCNN achieves the best overall

13Results for coherent documents are omitted due to space.

632

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Gender

Animacy↓

Animacy↑

Past to Future

F1

ΔF1

F1

ΔF1

F1

ΔF1

F1

ΔF1

BERT-Large
XLNet-Large
RoBERTa-Large
ALBERT-xxLarge

BERT-LargeCNN
XLNet-LargeCNN
RoBERTa-LargeCNN
ALBERT-xxLargeCNN

26.5 +65.3
55.8 +41.6
64.9 +32.5
74.0 +25.4

23.9 +70.0
13.6 +83.1
15.4 +82.4
21.6 +72.8

26.3 +53.2
50.0 +45.2
50.7 +38.3
+8.5
71.8

22.2 +60.2
10.0 +71.3
7.9 +64.4
20.2 +51.8

33.6 +45.1
64.0 +23.5
59.7 +21.7
+2.9
81.0

27.6 +51.4
8.0 +71.8
9.8 +73.3
27.6 +33.4

35.6 +42.1
64.9 +16.9
69.2 +19.9
+4.3
79.8

30.6 +14.7
23.2 +27.6
23.4 +40.0
38.0 +30.4

Human

35.8 +53.4

36.6 +45.3

29.8 +53.9

40.9 +34.4

Conjunction

Demonstrative

Negation

Number

F1

ΔF1

F1

ΔF1

F1

ΔF1

F1

ΔF1

BERT-Large
XLNet-Large
RoBERTa-Large
ALBERT-xxLarge

BERT-LargeCNN
XLNet-LargeCNN
RoBERTa-LargeCNN
ALBERT-xxLargeCNN

51.9 +17.3
+3.6
68.6
+0.7
73.0
−1.6
83.5
−1.4
0.0
+1.4
+1.3

38.2
31.0
33.9
41.6

34.8 +15.6
0.0
55.4
0.0
57.9
+1.3
75.2
−5.7
0.0
0.0
0.0

35.6
14.1
17.8
30.9

34.5 +32.2
+8.9
57.7
68.4 +10.9
+2.9
79.5

+4.2
28.8
15.7 +11.8
21.0 +12.4
28.1 +19.2

32.5 +31.2
50.7 +11.3
54.2 +20.0
63.9 +10.4

19.6 +11.7
15.2 +13.1
18.3 +23.6
23.0 +16.0

Human

40.5

+8.7

38.0

+1.0

40.4 +36.8

37.3 +24.2

Table 7: Results over language probes in incoherent Wikipedia test documents. BERT-LargeCNN, XLNet-
LargeCNN, RoBERTa-LargeCNN, and ALBERT-xxLargeCNN are trained over CNN, while BERT-Large,
XLNet-Large, RoBERTa-Large, and ALBERT-xxLarge are trained over Wikipedia. Here, F1 is over the
original incoherent documents (excluding linguistic probes), and ΔF1 indicates the absolute performance
difference resulting from incorporating linguistic probes.

performance over Gender, Animacy↑, and Num-
ber while ALBERT-xxLargeCNN achieves the best
overall performance over Past to Future, Conjunc-
tion, Demonstrative, and Negation. The reason
that the models tend to struggle with Demonstra-
tive and Conjunction is not immediately clear,
and will be explored in future work.

We also conducted human evaluations on this
dataset via Amazon Mechanical Turk, based on the
same methodology as described in Section 4.5
(without explicit instruction to look out for lin-
guistic artefacts, and with a mixture of coherent
and incoherent documents, as per the original
annotation task). As detailed in Table 7, humans
generally benefit from the inclusion of the linguis-
tic probes. Largely consistent with the results for
the models, humans are highly sensitised to the
effects of Gender, Animacy, Past to Future, and
Negation, but largely oblivious to the effects of

Demonstrative and Conjunction. Remarkably, the
best models (ALBERT-xxLarge and RoBERTa-
Large) perform on par with humans in the in-
domain setting, but are generally well below
humans in the cross-domain setting.

8 Conclusion

We propose the new task of detecting whether
there is an intruder sentence in a document, gen-
erated by replacing an original sentence with a
similar sentence from a second document. To
benchmark model performance over this task, we
construct a large-scale dataset consisting of doc-
uments from English Wikipedia and CNN news
articles. Experimental results show that pretrained
LMs that incorporate larger document contexts in
pretraining perform remarkably well in-domain,
but experience a substantial drop cross-domain. In

633

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

follow-up analysis based on human annotations,
substantial divergences from human intuitions
were observed, pointing to limitations in their
ability to model document coherence. Further
results over a linguistic probe dataset show that
pretrained models fail to identify some linguistic
characteristics that affect document coherence,
suggesting room to improve for them to truly
capture document coherence, and motivating the
construction of a dataset with intruder text at the
intra-sentential level.

References

Wafia Adouane, Jean-Philippe Bernardy, and
Simon Dobnik. 2018a. Improving neural net-
work performance by injecting background
knowledge: Detecting code-switching and
borrowing in Algerian texts. In Proceedings
of
the Third Workshop on Computational
to Linguistic Code-Switching,
Approaches
pages 20–28. DOI: https://doi.org/10
.18653/v1/W18-3203

Wafia Adouane, Simon Dobnik, Jean-Philippe
Bernardy, and Nasredine Semmar. 2018b. A
comparison of character neural language model
and bootstrapping for language identification
in multilingual noisy texts. In Proceedings of
the Second Workshop on Subword/Character
LEvel Models, pages 22–31. DOI: https://
doi.org/10.18653/v1/W18-1203

Malihe Alikhani, Sreyasi Nag Chowdhury, Gerard
de Melo, and Matthew Stone. 2019. CITE:
A corpus of image-text discourse relations. In
Proceedings of the 2019 Conference of the
North American Chapter of the Association for
Computational Linguistics: Human Language
Technologies, Volume 1 (Long and Short
Papers), pages 570–575.

Nicholas M. Asher and Alex Lascarides. 2003.
Logics of Conversation, Studies in Natural
Language Processing. Cambridge University
Press.

Regina Barzilay and Mirella Lapata. 2005.
Modeling local coherence: An entity-based
approach. In Proceedings of the 43rd Annual
Meeting of
the Association for Computa-
tional Linguistics, pages 141–148. DOI:
https://doi.org/10.3115/1219840
.1219858

Regina Barzilay and Mirella Lapata. 2008. Model-
ing local coherence: An entity-based approach.
Computational Linguistics, 34(1):1–34. DOI:
https://doi.org/10.1162/coli.2008
.34.1.1

Yonatan Belinkov, Nadir Durrani, Fahim Dalvi,
Hassan Sajjad, and James Glass. 2017. What
do neural machine translation models learn
the
about morphology? In Proceedings of
55th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long
Papers), pages 861–872. DOI: https://
doi.org/10.18653/v1/P17-1080

Yonatan Belinkov and James Glass. 2019. Anal-
ysis methods in neural language processing:
A survey. Transactions of the Association for
Computational Linguistics, 7:49–72. DOI:
https://doi.org/10.1162/tacl a
00254

and

Terra Blevins, Omer

Levy,

Luke
Zettlemoyer. 2018. Deep RNNs encode soft
the
hierarchical syntax.
56th Annual Meeting of the Association for
Computational Linguistics (Volume 2: Short
Papers), pages 14–19. DOI: https://
doi.org/10.18653/v1/P18-2003

In Proceedings of

Samuel R. Bowman, Gabor Angeli, Christopher
Potts, and Christopher D. Manning. 2015. A
large annotated corpus for learning natural
language inference. In Proceedings of the 2015
Conference on Empirical Methods in Natural
Language Processing, pages 632–642. DOI:
https://doi.org/10.18653/v1/D15
-1075

Jamie Callan, Mark Hoy, Changkuk Yoo,
and Le Zhao. 2009. Clueweb09 data set.
https://lemurproject.org/clueweb09/.
Accessed: 15.12.2019.

Khyathi Chandu, Eric Nyberg, and Alan W Black.
2019. Storyboarding of recipes: Grounded con-
textual generation. In Proceedings of the 57th
Annual Meeting of the Association for Compu-
tational Linguistics, pages 6040–6046. DOI:
https://doi.org/10.18653/v1/P19
-1606

Danqi Chen, Adam Fisch, Jason Weston, and
Antoine Bordes. 2017. Reading Wikipedia to

634

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

answer open-domain questions. In Proceedings
of the 55th Annual Meeting of the Association
for Computational Linguistics (Volume 1:
Long Papers),
1870–1879. DOI:
https://doi.org/10.18653/v1/P17
-1171

pages

Mingda Chen, Zewei Chu, and Kevin Gimpel.
2019. Evaluation benchmarks and learning
criteria for discourse-aware sentence rep-
the 2019
In Proceedings of
resentations.
Conference on Empirical Methods in Nat-
ural Language Processing and the 9th
International Joint Conference on Natural
Language Processing, pages 649–662. DOI:
https://doi.org/10.18653/v1/D19
-1060

Minhao Cheng, Jinfeng Yi, Pin-Yu Chen, Huan
Zhang, and Cho-Jui Hsieh. 2020. Seq2sick:
sequence-to-
Evaluating the robustness of
sequence models with adversarial
exam-
the Thirty-Fourth
ples.
AAAI Conference on Artificial Intelligence,
pages 3601–3608. DOI: https://doi.org
/10.1609/aaai.v34i04.5767

In Proceedings of

Orph´ee De Clercq, V´eronique Hoste, Bart
Desmet, Philip van Oosten, Martine De Cock,
and Lieve Macken. 2014. Using the crowd for
readability prediction. Natural Language En-
gineering, 20(3):293–325. DOI: https://
doi.org/10.1017/S1351324912000344

Alexis Conneau, Douwe Kiela, Holger Schwenk,
Lo¨ıc Barrault, and Antoine Bordes. 2017.
Supervised learning of universal
sentence
language infer-
representations from natural
the 2017
In Proceedings of
ence data.
Conference on Empirical Methods in Natural
Language Processing, pages 670–680. DOI:
https://doi.org/10.18653/v1/D17
-1070

Alexis Conneau, German Kruszewski, Guillaume
Lample, Lo¨ıc Barrault, and Marco Baroni.
2018. What you can cram into a single $&!#*
for
vector: Probing sentence embeddings
linguistic properties. In Proceedings of
the
56th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long
Papers), pages 2126–2136. DOI: https://
doi.org/10.18653/v1/P18-1198

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2019. BERT: Pre-training
of deep bidirectional transformers for language
understanding. In Proceedings of
the 2019
Conference of the North American Chapter of
the Association for Computational Linguistics:
Human Language Technologies, Volume 1
(Long and Short Papers), pages 4171–4186.

Micha Elsner, Joseph Austerweil, and Eugene
Charniak. 2007. A unified local and global
model for discourse coherence. In Human
Language Technologies 2007: The Conference
the
the North American Chapter of
of
Association for Computational Linguistics;
Proceedings
the Main Conference,
pages 436–443.

of

Micha Elsner and Eugene Charniak. 2011.
Extending the entity grid with entity-specific
features. In Proceedings of the 49th Annual
Meeting of the Association for Computational
Linguistics: Human Language Technologies,
pages 125–129.

Angela Fan, Mike Lewis, and Yann Dauphin.
2019. Strategies for structuring story gener-
the 57th Annual
ation.
Meeting of the Association for Computational
Linguistics, pages 2650–2660.

In Proceedings of

Herbert Paul Grice. 2002. Logic and conver-
sation. Foundations of Cognitive Psychology,
719–732.

Camille Guinaudeau and Michael Strube. 2013.
Graph-based local coherence modeling.
In
Proceedings of the 51st Annual Meeting of
the Association for Computational Linguistics
(Volume 1: Long Papers), pages 93–103.

Kristina Gulordava, Piotr Bojanowski, Edouard
Grave, Tal Linzen, and Marco Baroni. 2018.
Colorless green recurrent networks dream
the 2018
hierarchically. In Proceedings of
Conference of the North American Chapter of
the Association for Computational Linguistics:
Human Language Technologies, Volume 1
(Long Papers), pages 1195–1205. DOI:
https://doi.org/10.18653/v1/N18
-1108

Karl Moritz Hermann, Tom´aˇs Koˇcisk´y, Edward
Grefenstette, Lasse Espeholt, Will Kay,
Mustafa Suleyman, and Phil Blunsom. 2015.

635

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Teaching machines to read and comprehend.
the 28th International
In Proceedings of
Conference on Neural Information Processing
Systems – Volume 1, pages 1693–1701.

John Hewitt and Percy Liang. 2019. Designing
tasks.
and interpreting probes with control
the 2019 Conference on
In Proceedings of
Empirical Methods
in Natural Language
Processing and the 9th International Joint
Conference on Natural Language Processing,
pages 2733–2743. DOI: https://doi.org
/10.18653/v1/D19-1275

John Hewitt and Christopher D. Manning. 2019.
A structural probe for finding syntax in word
representations. In Proceedings of the 2019
Conference of the North American Chapter of
the Association for Computational Linguistics:
Human Language Technologies, Volume 1
(Long and Short Papers), pages 4129–4138.

Sepp Hochreiter and J¨urgen Schmidhuber. 1997.
Long short-term memory. Neural Computa-
tion, 9(8):1735–1780. DOI: https://doi
.org/10.1162/neco.1997.9.8.1735,
PMID: 9377276

Ari Holtzman,

Jan Buys, Maxwell Forbes,
Antoine Bosselut, David Golub, and Yejin
Choi. 2018. Learning to write with coopera-
the
tive discriminators.
56th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long
Papers), pages 1638–1649. DOI: https://
doi.org/10.18653/v1/P18-1152

In Proceedings of

Jennifer Hu, Sherry Y. Chen, and Roger P. Levy.
2020a. A closer look at the performance of
neural language models on reflexive anaphor
the Society for
licensing. Proceedings of
Computation in Linguistics, 3(1):382–392.

Junjie Hu, Yu Cheng, Zhe Gan, Jingjing Liu,
Jianfeng Gao, and Graham Neubig. 2020b.
What makes a good story? Designing compos-
ite rewards for visual storytelling. In Proceed-
the Thirty-Fourth AAAI Conference
ings of
on Artificial Intelligence, pages 7969–7976.
DOI: https://doi.org/10.1609/aaai
.v34i05.6305

Paria Jamshid Lou, Peter Anderson, and Mark
Johnson. 2018. Disfluency detection using

636

auto-correlational neural networks. In Proceed-
the 2018 Conference on Empirical
ings of
Methods in Natural Language Processing,
pages 4610–4619. DOI: https://doi.org
/10.18653/v1/D18-1490

Yangfeng Ji and Noah A. Smith. 2017. Neural
discourse structure for text categorization. In
Proceedings of the 55th Annual Meeting of
the Association for Computational Linguistics
(Volume 1: Long Papers), pages 996–1005.

Mark Johnson and Eugene Charniak. 2004. A
TAG-based noisy-channel model of speech
repairs. In Proceedings of the 42nd Annual
Meeting of the Association for Computational
Linguistics, pages 33–39. DOI: https://
doi.org/10.3115/1218955.1218960

Ryan Kiros, Yukun Zhu, Ruslan Salakhutdinov,
Richard S. Zemel, Antonio Torralba, Raquel
and Sanja Fidler. 2015. Skip-
Urtasun,
the
thought vectors.
28th International Conference on Neural
Information Processing Systems – Volume 2,
pages 3294–3302.

In Proceedings of

Adhiguna Kuncoro, Chris Dyer, John Hale, Dani
Yogatama, Stephen Clark, and Phil Blunsom.
2018. LSTMs can learn syntax-sensitive depen-
dencies well, but modeling structure makes
the 56th
them better.
the Association for
Annual Meeting of
Computational Linguistics (Volume 1: Long
Papers), pages 1426–1436. DOI: https://
doi.org/10.18653/v1/P18-1132

In Proceedings of

Alice Lai and Joel Tetreault. 2018. Discourse
coherence in the wild: A dataset, evaluation
the 19th
and methods.
Annual SIGdial Meeting on Discourse and
Dialogue, pages 214–223. DOI: https://
doi.org/10.18653/v1/W18-5023

In Proceedings of

Zhenzhong Lan, Mingda Chen, Sebastian Good-
man, Kevin Gimpel, Piyush Sharma, and Radu
Soricut. 2020. ALBERT: A lite BERT for
self-supervised learning of language represen-
tations. In Proceedings of the 8th International
Conference on Learning Representations.

Mirella Lapata. 2003. Probabilistic text struc-
turing: Experiments with sentence ordering.
In Proceedings of the 41st Annual Meeting

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

of
the Association for Computational Lin-
guistics, pages 545–552. DOI: https://
doi.org/10.3115/1075096.1075165

Jey Han Lau, Alexander Clark, and Shalom
Lappin. 2015. Unsupervised prediction of
acceptability judgements. In Proceedings of
the 53rd Annual Meeting of the Association
for Computational Linguistics and the 7th In-
ternational Joint Conference on Natural Lan-
guage Processing (Volume 1: Long Papers),
pages 1618–1628.

Jiwei Li and Dan Jurafsky. 2017. Neural net
models of open-domain discourse coherence.
the 2017 Conference on
In Proceedings of
in Natural Language
Empirical Methods
Processing, pages 198–209.

Xia Li, Minping Chen, Jianyun Nie, Zhenxing
Liu, Ziheng Feng, and Yingdan Cai. 2018.
Coherence-based automated essay scoring
using self-attention. In Proceedings of
the
17th China National Conference on Com-
putational Linguistics, CCL 2018, and the
6th International Symposium on Natural
Language Processing Based on Naturally
Annotated Big Data, pages 386–397. DOI:
https://doi.org/10.1007/978-3-030
-01716-3 32

Proceedings of the 2018 Conference on Em-
pirical Methods in Natural Language Pro-
cessing, pages 1192–1202. DOI: https://
doi.org/10.18653/v1/D18-1151

Deepthi Mave, Suraj Maharjan, and Thamar
Solorio. 2018. Language identification and
analysis of code-switched social media text.
In Proceedings of the Third Workshop on Com-
putational Approaches to Linguistic Code-
Switching, pages 51–61. DOI: https://
doi.org/10.18653/v1/W18-3206

Tom McCoy, Ellie Pavlick, and Tal Linzen.
2019. Right for the wrong reasons: Diagnos-
ing syntactic heuristics in natural
language
inference. In Proceedings of the 57th Annual
Meeting of
the Association for Computa-
tional Linguistics, pages 3428–3448. DOI:
https://doi.org/10.18653/v1/P19
-1334

In Proceedings of

Mohsen Mesgar and Michael Strube. 2016.
Lexical coherence graph modeling using word
the 2016
embeddings.
Conference of the North American Chapter
of
the Association for Computational Lin-
guistics: Human Language Technologies,
pages 1414–1423. DOI: https://doi.org
/10.18653/v1/N16-1167

Bin Liang, Hongcheng Li, Miaoqiang Su, Pan
Bian, Xirong Li, and Wenchang Shi. 2018.
Deep text classification can be fooled. In Pro-
ceedings of the Twenty-Seventh International
Joint Conference on Artificial Intelligence,
pages 4208–4215. DOI: https://doi.org
/10.24963/ijcai.2018/585

Mohsen Mesgar and Michael Strube. 2018. A
neural local coherence model for text quality
the 2018
In Proceedings of
assessment.
Conference on Empirical Methods in Natural
Language Processing, pages 4328–4339. DOI:
https://doi.org/10.18653/v1/D18
-1464

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei
Joshi, Danqi Chen, Omer
Du, Mandar
Levy, Mike Lewis, Luke Zettlemoyer, and
Veselin Stoyanov. 2019. RoBERTa: A robustly
optimized BERT pretraining approach. CoRR,
cs.CL/1907.11692v1.

William C. Mann and Sandra A. Thompson.
1988. Rhetorical structure theory: Toward a
functional theory of text organization. Text,
8(3):243–281. DOI: https://doi.org
/10.1515/text.1.1988.8.3.243

Rebecca Marvin and Tal Linzen. 2018. Targeted
syntactic evaluation of language models. In

Eleni Miltsakaki, Rashmi Prasad, Aravind
Joshi, and Bonnie Webber. 2004. The Penn
the
discourse treebank.
Fourth International Conference on Language
Resources and Evaluation.

In Proceedings of

Farjana Sultana Mim, Naoya

Inoue, Paul
Reisert, Hiroki Ouchi, and Kentaro Inui. 2019.
Unsupervised learning of discourse-aware text
representation. In Proceedings of
the 2019
Student Research Workshop, pages 93–104.

Ramesh Nallapati, Bowen Zhou, C´ıcero Nogueira
dos Santos, C¸ aglar G¨ulc¸ehre, and Bing Xiang.
2016. Abstractive text summarization using

637

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

sequence-to-sequence RNNs and beyond. In
Proceedings of The 20th SIGNLL Conference
on Computational Natural Language Learning,
pages 280–290. DOI: https://doi.org
/10.18653/v1/K16-1028

Nicolas Papernot, Fartash Faghri, Nicholas
Carlini,
Ian Goodfellow, Reuben Feinman,
Alexey Kurakin, Cihang Xie, Yash Sharma,
Tom Brown, Aurko Roy, Alexander Matyasko,
Vahid Behzadan, Karen Hambardzumyan,
Zhishuai Zhang, Yi-Lin Juang, Zhi Li, Ryan
Sheatsley, Abhibhav Garg, Jonathan Uesato,
Willi Gierke, Yinpeng Dong, David Berthelot,
Paul Hendricks, Jonas Rauber, and Rujun Long.
2018. Technical report on the CleverHans
v2.1.0 adversarial examples library. CoRR,
cs.LG/1610.00768v6.

Cesc C. Park and Gunhee Kim. 2015. Expressing
an image stream with a sequence of natural
In Proceedings of Advances in
sentences.
Neural Information Processing Systems 28,
pages 73–81.

Robert Parker, David Graff,

Junbo Kong,
Ke Chen, and Kazuaki Maeda. 2011. English
linguistic data con-
gigaword fifth edition,
sortium. Google Scholar.

Jeffrey

Socher,

Pennington, Richard

and
Christopher Manning. 2014. GloVe: Global
vectors for word representation. In Proceed-
ings of
the 2014 Conference on Empirical
Methods in Natural Language Processing,
pages 1532–1543. DOI: https://doi.org
/10.3115/v1/D14-1162

Matthew Peters, Mark Neumann,

Luke
Zettlemoyer, and Wen-tau Yih. 2018. Dis-
secting contextual word embeddings: Archi-
In Proceedings
tecture and representation.
of
on Empirical
2018 Conference
Methods in Natural Language Processing,
pages 1499–1509. DOI: https://doi.org
/10.18653/v1/D18-1179

the

evaluation of

Emily Pitler, Annie Louis, and Ani Nenkova.
2010. Automatic
linguistic
quality in multi-document summarization. In
Proceedings of the 48th Annual Meeting of
the Association for Computational Linguistics,
pages 544–554.

In Proceedings of

Emily Pitler and Ani Nenkova. 2008. Revisiting
readability: A unified framework for predicting
the 2008
text quality.
Conference on Empirical Methods in Natural
Language Processing, pages 186–195. DOI:
https://doi.org/10.3115/1613715
.1613742

Shrimai Prabhumoye, Ruslan Salakhutdinov, and
Alan W Black. 2020. Topological sort for
sentence ordering. In Proceedings of the 58th
Annual Meeting of the Association for Compu-
tational Linguistics, pages 2783–2792. DOI:
https://doi.org/10.18653/v1/2020
.acl-main.248

Rashmi Prasad, Nikhil Dinesh, Alan Lee,
Eleni Miltsakaki, Livio Robaldo, Aravind
Joshi, and Bonnie Webber. 2008. The Penn
discourse TreeBank 2.0. In Proceedings of the
Sixth International Conference on Language
Resources and Evaluation.

Jan Wira Gotama Putra and Takenobu Tokunaga.
2017. Evaluating text coherence based on
semantic similarity graph. In Proceedings of
TextGraphs-11: the Workshop on Graph-based
Methods for Natural Language Processing,
pages 76–85.

Suranjana Samanta and Sameep Mehta. 2017.
Towards crafting text adversarial samples.
CoRR, cs.LG/1707.02812v1.

Motoki Sato, Jun Suzuki, Hiroyuki Shindo, and
Yuji Matsumoto. 2018. Interpretable adver-
sarial perturbation in input embedding space
for text. In Proceedings of the Twenty-Seventh
International Joint Conference on Artificial
Intelligence, pages 4323–4330. DOI: https://
doi.org/10.24963/ijcai.2018/601

Swapna Somasundaran,

Jill Burstein,

and
Martin Chodorow. 2014. Lexical chaining
for measuring discourse coherence quality in
test-taker essays. In Proceedings of the 25th
International Conference on Computational
Linguistics: Technical Papers, pages 950–961.

Gongbo Tang, Mathias M¨uller, Annette Rios, and
Rico Sennrich. 2018. Why self-attention? A
targeted evaluation of neural machine trans-
lation architectures. In Proceedings of the 2018
Conference on Empirical Methods in Natural

638

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

Language Processing,
4263–4272.
DOI: https://doi.org/10.18653/v1
/D18-1458

pages

Deirdre Wilson and Dan Sperber. 2004. Rele-
vance theory, The Handbook of Pragmatics,
Blackwell, pages 607–632.

Yi Tay, Minh C. Phan, Luu Anh Tuan, and
Siu Cheung Hui. 2018. SkipFlow: Incorporating
neural
end-to-end
automatic text scoring. In Proceedings of the
Thirty-Second AAAI Conference on Artificial
Intelligence, pages 5948–5955.

coherence

features

for

Dat Tien Nguyen and Shafiq Joty. 2017. A neural
In Proceedings of
local coherence model.
the 55th Annual Meeting of the Association
for Computational Linguistics (Volume 1:
Long Papers),
1320–1330. DOI:
https://doi.org/10.18653/v1/P17
-1121

pages

Ke Tran, Arianna Bisazza, and Christof Monz.
2018. The importance of being recurrent
for modeling hierarchical structure. In Pro-
ceedings of the 2018 Conference on Empirical
Methods in Natural Language Processing,
pages 4731–4736. DOI: https://doi.org
/10.18653/v1/D18-1503

Trieu H. Trinh and Quoc V. Le. 2018. A simple
method for commonsense reasoning. CoRR,
cs.AI/1806.02847v2.

Alex Warstadt, Amanpreet Singh, and Samuel
Bowman. 2019. Neural network acceptability
the Association
judgments. Transactions of
for Computational Linguistics, 7(0). DOI:
https://doi.org/10.1162/tacl a
00290

Bonnie Webber. 2009. Genre distinctions for dis-
course in the Penn TreeBank. In Proceedings
of the Joint Conference of the 47th Annual
Meeting of the ACL and the 4th International
Joint Conference on Natural Language Pro-
cessing of the AFNLP, pages 674–682. DOI:
https://doi.org/10.3115/1690219
.1690240

Ethan Wilcox, Roger Levy, Takashi Morita, and
Richard Futrell. 2018. What do RNN language
models learn about filler–gap dependencies?
In Proceedings of the 2018 EMNLP Workshop
BlackboxNLP: Analyzing and Interpreting Neu-
ral Networks for NLP, pages 211–221. DOI:
https://doi.org/10.18653/v1/W18
-5423

Peng Xu, Hamidreza Saghir, Jin Sung Kang,
Teng Long, Avishek Joey Bose, Yanshuai Cao,
and Jackie Chi Kit Cheung. 2019. A cross-
domain transferable neural coherence model.
In Proceedings of the 57th Annual Meeting of
the Association for Computational Linguistics,
pages 678–687.

Puyudi Yang, Jianbo Chen, Cho-Jui Hsieh,
Jane-Ling Wang, and Michael I. Jordan. 2020.
Greedy attack and gumbel attack: Generating
adversarial examples for discrete data. Journal
of Machine Learning Research, 21:43:1–43:36.

Yi Yang, Wen-tau Yih, and Christopher Meek.
2015. WikiQA: A challenge dataset for open-
domain question answering. In Proceedings
on Empirical
of
Methods in Natural Language Processing,
pages 2013–2018. DOI: https://doi.org
/10.18653/v1/D15-1237

2015 Conference

the

Zhilin Yang, Zihang Dai, Yiming Yang, Jaime
Carbonell, Russ R Salakhutdinov, and Quoc V
Le. 2019. Xlnet: Generalized autoregressive
In
language understanding.
pretraining for
Proceedings of
the Thirty-third Conference
on Neural Information Processing Systems,
pages 5754–5764.

Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua
Bengio, William Cohen, Ruslan Salakhutdinov,
and Christopher D. Manning. 2018. HotpotQA:
A dataset for diverse, explainable multi-hop
question answering. In Proceedings of the 2018
Conference on Empirical Methods in Natural
Language Processing, pages 2369–2380. DOI:
https://doi.org/10.18653/v1/D18
-1259

language

code-switching
pair.

and G¨uls¸en Eryi˘git.
Zeynep Yirmibes¸o˘glu
between
2018. Detecting
In Pro-
Turkish-English
the 2018 EMNLP Workshop
ceedings of
on Noisy
W-NUT: The
User-generated Text, pages 110–115. DOI:
https://doi.org/10.18653/v1/W18
-6115

4th Workshop

Rowan Zellers, Yonatan Bisk, Roy Schwartz,
and Yejin Choi. 2018. SWAG: A large-scale

639

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

adversarial dataset
for grounded common-
sense inference. In Proceedings of the 2018
Conference on Empirical Methods in Natural
Language Processing, pages 93–104. DOI:
https://doi.org/10.18653/v1/D18
-1009

Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali
Farhadi, and Yejin Choi. 2019. HellaSwag: Can
a machine really finish your sentence? In Pro-
ceedings of the 57th Annual Meeting of the
Association for Computational Linguistics,

4791–4800. DOI: https://doi

pages
.org/10.18653/v1/P19-1472

Yukun Zhu, Ryan Kiros, Richard S. Zemel,
Ruslan
Salakhutdinov, Raquel Urtasun,
Antonio Torralba, and Sanja Fidler. 2015.
Aligning books and movies: Towards story-
like visual explanations by watching movies
and reading books. In Proceedings of
the
2015 IEEE Conference on Computer Vision
and Pattern Recognition, pages 19–27. DOI:
https://doi.org/10.1109/ICCV.2015.11

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
3
8
8
1
9
2
9
6
9
9

/

/
t

l

a
c
_
a
_
0
0
3
8
8
p
d

.

f

b
y
g
u
e
s
t

t

o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3

640
Download pdf