Speech and Language Processing. Daniel Jurafsky & James H. Martin. Copyright © 2024.

All
rights reserved. Draft of January 12, 2025.

CHAPTER

24 Discourse Coherence

And even in our wildest and most wandering reveries, nay in our very dreams,
we shall find, if we reflect, that the imagination ran not altogether at adven-
tures, but that there was still a connection upheld among the different ideas,
which succeeded each other. Were the loosest and freest conversation to be
transcribed, there would immediately be transcribed, there would immediately
be observed something which connected it in all its transitions.
David Hume, An enquiry concerning human understanding, 1748

Orson Welles’ movie Citizen Kane was groundbreaking in many ways, perhaps most
notably in its structure. The story of the life of fictional media magnate Charles
Foster Kane, the movie does not proceed in chronological order through Kane’s
life. Instead, the film begins with Kane’s death (famously murmuring “Rosebud”)
and is structured around flashbacks to his life inserted among scenes of a reporter
investigating his death. The novel idea that the structure of a movie does not have
to linearly follow the structure of the real timeline made apparent for 20th century
cinematography the infinite possibilities and impact of different kinds of coherent
narrative structures.
But coherent structure is not just a fact about movies or works of art. Like
movies, language does not normally consist of isolated, unrelated sentences, but
instead of collocated, structured, coherent groups of sentences. We refer to such
discourse a coherent structured group of sentences as a discourse, and we use the word co-
coherence herence to refer to the relationship between sentences that makes real discourses
different than just random assemblages of sentences. The chapter you are now read-
ing is an example of a discourse, as is a news article, a conversation, a thread on
social media, a Wikipedia page, and your favorite novel.
What makes a discourse coherent? If you created a text by taking random sen-
tences each from many different sources and pasted them together, would that be a
local coherent discourse? Almost certainly not. Real discourses exhibit both local coher-
global ence and global coherence. Let’s consider three ways in which real discourses are
locally coherent;
First, sentences or clauses in real discourses are related to nearby sentences in
systematic ways. Consider this example from Hobbs (1979):
(24.1) John took a train from Paris to Istanbul. He likes spinach.
This sequence is incoherent because it is unclear to a reader why the second
sentence follows the first; what does liking spinach have to do with train trips? In
fact, a reader might go to some effort to try to figure out how the discourse could be
coherent; perhaps there is a French spinach shortage? The very fact that hearers try
to identify such connections suggests that human discourse comprehension involves
the need to establish this kind of coherence.
By contrast, in the following coherent example:
(24.2) Jane took a train from Paris to Istanbul. She had to attend a conference.
2 C HAPTER 24 • D ISCOURSE C OHERENCE

the second sentence gives a REASON for Jane’s action in the first sentence. Struc-
tured relationships like REASON that hold between text units are called coherence
coherence relations, and coherent discourses are structured by many such coherence relations.
relations
Coherence relations are introduced in Section 24.1.
A second way a discourse can be locally coherent is by virtue of being “about”
someone or something. In a coherent discourse some entities are salient, and the
discourse focuses on them and doesn’t go back and forth between multiple entities.
This is called entity-based coherence. Consider the following incoherent passage,
in which the salient entity seems to wildly swing from John to Jenny to the piano
store to the living room, back to Jenny, then the piano again:
(24.3) John wanted to buy a piano for his living room.
Jenny also wanted to buy a piano.
He went to the piano store.
It was nearby.
The living room was on the second floor.
She didn’t find anything she liked.
The piano he bought was hard to get up to that floor.
Entity-based coherence models measure this kind of coherence by tracking salient
Centering
Theory entities across a discourse. For example Centering Theory (Grosz et al., 1995), the
most influential theory of entity-based coherence, keeps track of which entities in
the discourse model are salient at any point (salient entities are more likely to be
pronominalized or to appear in prominent syntactic positions like subject or object).
In Centering Theory, transitions between sentences that maintain the same salient
entity are considered more coherent than ones that repeatedly shift between entities.
entity grid The entity grid model of coherence (Barzilay and Lapata, 2008) is a commonly
used model that realizes some of the intuitions of the Centering Theory framework.
Entity-based coherence is introduced in Section 24.3.
topically Finally, discourses can be locally coherent by being topically coherent: nearby
coherent
sentences are generally about the same topic and use the same or similar vocab-
ulary to discuss these topics. Because topically coherent discourses draw from a
single semantic field or topic, they tend to exhibit the surface property known as
lexical cohesion lexical cohesion (Halliday and Hasan, 1976): the sharing of identical or semanti-
cally related words in nearby sentences. For example, the fact that the words house,
chimney, garret, closet, and window— all of which belong to the same semantic
field— appear in the two sentences in (24.4), or that they share the identical word
shingled, is a cue that the two are tied together as a discourse:
(24.4) Before winter I built a chimney, and shingled the sides of my house...
I have thus a tight shingled and plastered house... with a garret and a
closet, a large window on each side....
In addition to the local coherence between adjacent or nearby sentences, dis-
courses also exhibit global coherence. Many genres of text are associated with
particular conventional discourse structures. Academic articles might have sections
describing the Methodology or Results. Stories might follow conventional plotlines
or motifs. Persuasive essays have a particular claim they are trying to argue for,
and an essay might express this claim together with a structured set of premises that
support the argument and demolish potential counterarguments. We’ll introduce
versions of each of these kinds of global coherence.
Why do we care about the local or global coherence of a discourse? Since co-
herence is a property of a well-written text, coherence detection plays a part in any
 24.1 • C OHERENCE R ELATIONS 3

task that requires measuring the quality of a text. For example coherence can help
in pedagogical tasks like essay grading or essay quality measurement that are trying
to grade how well-written a human essay is (Somasundaran et al. 2014, Feng et al.
2014, Lai and Tetreault 2018). Coherence can also help for summarization; knowing
the coherence relationship between sentences can help know how to select informa-
tion from them. Finally, detecting incoherent text may even play a role in mental
health tasks like measuring symptoms of schizophrenia or other kinds of disordered
language (Ditman and Kuperberg 2010, Elvevåg et al. 2007, Bedi et al. 2015, Iter
et al. 2018).

24.1 Coherence Relations

Recall from the introduction the difference between passages (24.5) and (24.6).
(24.5) Jane took a train from Paris to Istanbul. She likes spinach.
(24.6) Jane took a train from Paris to Istanbul. She had to attend a conference.
The reason (24.6) is more coherent is that the reader can form a connection be-
tween the two sentences, in which the second sentence provides a potential REASON
for the first sentences. This link is harder to form for (24.5). These connections
coherence between text spans in a discourse can be specified as a set of coherence relations.
relation
The next two sections describe two commonly used models of coherence relations
and associated corpora: Rhetorical Structure Theory (RST), and the Penn Discourse
TreeBank (PDTB).

24.1.1 Rhetorical Structure Theory

The most commonly used model of discourse organization is Rhetorical Structure
RST Theory (RST) (Mann and Thompson, 1987). In RST relations are defined between
nucleus two spans of text, generally a nucleus and a satellite. The nucleus is the unit that
satellite is more central to the writer’s purpose and that is interpretable independently; the
satellite is less central and generally is only interpretable with respect to the nucleus.
Some symmetric relations, however, hold between two nuclei.
Below are a few examples of RST coherence relations, with definitions adapted
from the RST Treebank Manual (Carlson and Marcu, 2001).
Reason: The nucleus is an action carried out by an animate agent and the satellite
is the reason for the nucleus.
(24.7) [NUC Jane took a train from Paris to Istanbul.] [SAT She had to attend a
conference.]

Elaboration: The satellite gives additional information or detail about the situation
presented in the nucleus.
(24.8) [NUC Dorothy was from Kansas.] [SAT She lived in the midst of the great
Kansas prairies.]

Evidence: The satellite gives additional information or detail about the situation
presented in the nucleus. The information is presented with the goal of convince the
reader to accept the information presented in the nucleus.
(24.9) [NUC Kevin must be here.] [SAT His car is parked outside.]
4 C HAPTER 24 • D ISCOURSE C OHERENCE

Attribution: The satellite gives the source of attribution for an instance of reported
speech in the nucleus.
(24.10) [SAT Analysts estimated] [NUC that sales at U.S. stores declined in the
quarter, too]

List: In this multinuclear relation, a series of nuclei is given, without contrast or

explicit comparison:
(24.11) [NUC Billy Bones was the mate; ] [NUC Long John, he was quartermaster]

RST relations are traditionally represented graphically; the asymmetric Nucleus-

Satellite relation is represented with an arrow from the satellite to the nucleus:

evidence

Kevin must be here. His car is parked outside

We can also talk about the coherence of a larger text by considering the hierar-
chical structure between coherence relations. Figure 24.1 shows the rhetorical struc-
ture of a paragraph from Marcu (2000a) for the text in (24.12) from the Scientific
American magazine.
(24.12) With its distant orbit–50 percent farther from the sun than Earth–and slim
atmospheric blanket, Mars experiences frigid weather conditions. Surface
temperatures typically average about -60 degrees Celsius (-76 degrees
Fahrenheit) at the equator and can dip to -123 degrees C near the poles. Only
the midday sun at tropical latitudes is warm enough to thaw ice on occasion,
but any liquid water formed in this way would evaporate almost instantly
because of the low atmospheric pressure.

Title 2-9
(1)
evidence
Mars

2-3 4-9
background elaboration-additional

(2) (3) 4-5 6-9

WIth its Mars
distant orbit experiences
<p> -- 50 frigid weather List Contrast
percent conditions.
farther from (4) (5) 6-7 8-9
the sun than Surface and can dip
Earth -- </p> temperatures to -123 purpose explanation-argumentative
and slim typically average degrees C
atmospheric about -60 near the (6) (7) (8) (9)
blanket, degrees Celsius poles. Only the to thaw ice but any liquid water because of
<p> (-76 degrees midday sun at on occasion, formed in this way the low
Fahrenheit)</p> tropical latitudes would evaporate atmospheric
at the equator is warm enough almost instantly pressure.

Figure 24.1 A discourse tree for the Scientific American text in (24.12), from Marcu (2000a). Note that
asymmetric relations are represented with a curved arrow from the satellite to the nucleus.

The leaves in the Fig. 24.1 tree correspond to text spans of a sentence, clause or
EDU phrase that are called elementary discourse units or EDUs in RST; these units can
also be referred to as discourse segments. Because these units may correspond to
arbitrary spans of text, determining the boundaries of an EDU is an important task
for extracting coherence relations. Roughly speaking, one can think of discourse
 24.1 • C OHERENCE R ELATIONS 5

segments as being analogous to constituents in sentence syntax, and indeed as we’ll

see in Section 24.2 we generally draw on parsing algorithms to infer discourse struc-
ture.
There are corpora for many discourse coherence models; the RST Discourse
TreeBank (Carlson et al., 2001) is the largest available discourse corpus. It con-
sists of 385 English language documents selected from the Penn Treebank, with full
RST parses for each one, using a large set of 78 distinct relations, grouped into 16
classes. RST treebanks exist also for Spanish, German, Basque, Dutch and Brazilian
Portuguese (Braud et al., 2017).
Now that we’ve seen examples of coherence, we can see more clearly how a
coherence relation can play a role in summarization or information extraction. For
example, the nuclei of a text presumably express more important information than
the satellites, which might be dropped in a summary.

24.1.2 Penn Discourse TreeBank (PDTB)

PDTB The Penn Discourse TreeBank (PDTB) is a second commonly used dataset that
embodies another model of coherence relations (Miltsakaki et al. 2004, Prasad et al.
2008, Prasad et al. 2014). PDTB labeling is lexically grounded. Instead of asking
annotators to directly tag the coherence relation between text spans, they were given
discourse a list of discourse connectives, words that signal discourse relations, like because,
connectives
although, when, since, or as a result. In a part of a text where these words marked a
coherence relation between two text spans, the connective and the spans were then
annotated, as in Fig. 24.13, where the phrase as a result signals a causal relationship
between what PDTB calls Arg1 (the first two sentences, here in italics) and Arg2
(the third sentence, here in bold).
(24.13) Jewelry displays in department stores were often cluttered and uninspired.
And the merchandise was, well, fake. As a result, marketers of faux gems
steadily lost space in department stores to more fashionable
rivals—cosmetics makers.
(24.14) In July, the Environmental Protection Agency imposed a gradual ban on
virtually all uses of asbestos. (implicit=as a result) By 1997, almost all
remaining uses of cancer-causing asbestos will be outlawed.
Not all coherence relations are marked by an explicit discourse connective, and
so the PDTB also annotates pairs of neighboring sentences with no explicit signal,
like (24.14). The annotator first chooses the word or phrase that could have been its
signal (in this case as a result), and then labels its sense. For example for the am-
biguous discourse connective since annotators marked whether it is using a C AUSAL
or a T EMPORAL sense.
The final dataset contains roughly 18,000 explicit relations and 16,000 implicit
relations. Fig. 24.2 shows examples from each of the 4 major semantic classes, while
Fig. 24.3 shows the full tagset.
Unlike the RST Discourse Treebank, which integrates these pairwise coherence
relations into a global tree structure spanning an entire discourse, the PDTB does not
annotate anything above the span-pair level, making no commitment with respect to
higher-level discourse structure.
There are also treebanks using similar methods for other languages; (24.15)
shows an example from the Chinese Discourse TreeBank (Zhou and Xue, 2015).
Because Chinese has a smaller percentage of explicit discourse connectives than
English (only 22% of all discourse relations are marked with explicit connectives,
6 C HAPTER 24 • D ISCOURSE C OHERENCE

Class Type
Example
TEMPORAL The parishioners of St. Michael and All Angels stop to chat at
SYNCHRONOUS
the church door, as members here always have. (Implicit while)
In the tower, five men and women pull rhythmically on ropes
attached to the same five bells that first sounded here in 1614.
CONTINGENCY REASON Also unlike Mr. Ruder, Mr. Breeden appears to be in a position
to get somewhere with his agenda. (implicit=because) As a for-
mer White House aide who worked closely with Congress,
he is savvy in the ways of Washington.
COMPARISON CONTRAST The U.S. wants the removal of what it perceives as barriers to
investment; Japan denies there are real barriers.
EXPANSION CONJUNCTION Not only do the actors stand outside their characters and make
it clear they are at odds with them, but they often literally stand
on their heads.
Figure 24.2 The four high-level semantic distinctions in the PDTB sense hierarchy

Temporal Comparison
• Asynchronous • Contrast (Juxtaposition, Opposition)
• Synchronous (Precedence, Succession) •Pragmatic Contrast (Juxtaposition, Opposition)
• Concession (Expectation, Contra-expectation)
• Pragmatic Concession

Contingency Expansion
• Cause (Reason, Result) • Exception
• Pragmatic Cause (Justification) • Instantiation
• Condition (Hypothetical, General, Unreal • Restatement (Specification, Equivalence, Generalization)
Present/Past, Factual Present/Past)
• Pragmatic Condition (Relevance, Implicit As- • Alternative (Conjunction, Disjunction, Chosen Alterna-
sertion) tive)
• List
Figure 24.3 The PDTB sense hierarchy. There are four top-level classes, 16 types, and 23 subtypes (not all
types have subtypes). 11 of the 16 types are commonly used for implicit ¯ argument classification; the 5 types in
italics are too rare in implicit labeling to be used.

compared to 47% in English), annotators labeled this corpus by directly mapping

pairs of sentences to 11 sense tags, without starting with a lexical discourse connec-
tor.
(24.15) [Conn 为] [Arg2 推动图们江地区开发] ，[Arg1 韩国捐款一百万美元
设立了图们江发展基金]
“[In order to] [Arg2 promote the development of the Tumen River region],
[Arg1 South Korea donated one million dollars to establish the Tumen
River Development Fund].”
These discourse treebanks have been used for shared tasks on multilingual dis-
course parsing (Xue et al., 2016).

24.2 Discourse Structure Parsing

Given a sequence of sentences, how can we automatically determine the coherence
discourse
parsing relations between them? This task is often called discourse parsing (even though
for PDTB we are only assigning labels to leaf spans and not building a full parse
 24.2 • D ISCOURSE S TRUCTURE PARSING 7

tree as we do for RST).

24.2.1 EDU segmentation for RST parsing

RST parsing is generally done in two stages. The first stage, EDU segmentation,
extracts the start and end of each EDU. The output of this stage would be a labeling
like the following:
(24.16) [Mr. Rambo says]e1 [that a 3.2-acre property]e2 [overlooking the San
Fernando Valley]e3 [is priced at $4 million]e4 [because the late actor Erroll
Flynn once lived there.]e5
Since EDUs roughly correspond to clauses, early models of EDU segmentation
first ran a syntactic parser, and then post-processed the output. Modern systems
generally use neural sequence models supervised by the gold EDU segmentation in
datasets like the RST Discourse Treebank. Fig. 24.4 shows an example architecture
simplified from the algorithm of Lukasik et al. (2020) that predicts for each token
whether or not it is a break. Here the input sentence is passed through an encoder
and then passed through a linear layer and a softmax to produce a sequence of 0s
and 1, where 1 indicates the start of an EDU.

EDU break 0 0 0 1

softmax

linear layer

ENCODER
Mr. Rambo says that …
Figure 24.4 Predicting EDU segment beginnings from encoded text.

24.2.2 RST parsing

Tools for building RST coherence structure for a discourse have long been based on
syntactic parsing algorithms like shift-reduce parsing (Marcu, 1999). Many modern
RST parsers since Ji and Eisenstein (2014) draw on the neural syntactic parsers we
saw in Chapter 20, using representation learning to build representations for each
span, and training a parser to choose the correct shift and reduce actions based on
the gold parses in the training set.
We’ll describe the shift-reduce parser of Yu et al. (2018). The parser state con-
sists of a stack and a queue, and produces this structure by taking a series of actions
on the states. Actions include:
• shift: pushes the first EDU in the queue onto the stack creating a single-node
subtree.
• reduce(l,d): merges the top two subtrees on the stack, where l is the coherence
relation label, and d is the nuclearity direction, d ∈ {NN, NS, SN}.
As well as the pop root operation, to remove the final tree from the stack.
Fig. 24.6 shows the actions the parser takes to build the structure in Fig. 24.5.
 8 C HAPTER 24 • D ISCOURSE C OHERENCE

elab e1 : American Telephone & Telegraph Co. said it

e2 : will lay off 75 to 85 technicians here , effective Nov. 1.
e3 : The workers install , maintain and repair its private branch exchanges,
attr elab e4 : which are large intracompany telephone networks.
e1 e2 e3 e4

Figure
Figure 24.5example
1: An ExampleofRST
RSTdiscourse tree,tree,
discourse showing four{e
where EDUs. Figure from Yu et al. (2018).
1 , e2 , e3 , e4 } are EDUs, attr and elab are
discourse relation labels, and arrows indicate the nuclearities of discourse relations.
Step Stack Queue Action Relation

RST discourse parsing. Other 1 studies? still adopt e1 , ediscrete

2 , e3 , e4
syntax features SH proposed by statistical ? models,
2 e1 e2 , e3 , e4 SH ?
feeding them into neural 3network emodels (Braud et al., 2016; Braud et al., 2017).
1 , e2 e3 , e4 RD(attr,SN) ?
The above approaches 4model syntax e1:2 trees in ane3explicit , e4 way, requiring SH discrete syntaxed parsing outputs
1 e2
as inputs for RST parsing.5 These eapproaches 1:2 , e 3 may suffer
e 4 from the error SH propagation problem.ed 1 e2 Syntax trees
produced by a supervised6 syntax , e3 , e4model could
e1:2parsing ? have errors, RD(elab,NS)which may propagate ed 1 einto
2 discourse
parsing models. The problem 7 e1:2 be
could , e3:4
extremely serious ? when RD(elab,SN)
inputs of discourse parsing ed1 e2 , ed 3 e4 different
have
distributions with the training 8 data eof1:4the supervised ? syntax parser. Recently, PR Zhang ed1 e2et
, ed , e\
3 e4(2017)
al. 1:2 e3:4
suggest
an alternative method, which extracts syntax features from a Bi-Affine dependency parser (Dozat and
Figure 24.6 Parsing the example of Fig. 24.5 using a shift-reduce parser. Figure from Yu
Manning, 2016), Table
and 1: An
the(2018). example
method givesofcompetitive
the transition-based
performancessystem on relation for RST discourse
extraction. parsing.
It actually
et al.
represents syntax trees implicitly, thus it can reduce the error propagation problem.
In this work, we investigate
The Yu the implicit
et al. (2018)syntaxuses anfeature extraction approach
encoder-decoder architecture, for RST
whereparsing.
the encoder In ad-
The initial
dition, state ais
we propose an empty
represents
transition-based state,
the input andmodel
span
neural ofthe final
words forand state
this EDUs represents
task, a full
using aishierarchical
which able to result. biLSTM.
incorporate There Theare three kinds of
various
actionsflexibly.
features in our transition
Wefirst
exploitbiLSTM system:layer represents
hierarchical bi-directionalthe words LSTMs inside an EDU, and
(Bi-LSTMs) the second
to encode texts,represents
and further
the EDU sequence. Given an input
enhance the transition-based model with dynamic oracle. Based 1on 2the proposed sentence w , w , ..., w m , the words
model, canwebestudy
repre-the
• Shift (SH),
effectiveness sented
of our which
proposed as usual
removes (by
implicit the static embeddings,
firstfeatures.
syntax EDU inWe combinations
theconduct
queueexperiments with character
onto the stack, embeddings
formingRST
on a standard or
a single-node
dis- subtree.
course TreeBank (Carlsontags, or etcontextual
al., 2003). embeddings)
First, we resulting
evaluate thein an input
performance word of representation
our proposed sequence
transition-
• Reduce (RD) xw1 ,(l,d),
w , ...,which
xthat Themerges
xwm .model result of the
the top two subtrees
word-level biLSTM onthen
is thea stack,
sequence where is a discourse relation
of hw lvalues:
based baseline, finding 2 the is able to achieve strong performances after applying dynamic
label, and d 2 {NN, NS,
oracle. Then we evaluate the effectiveness SN} indicates the relation nuclearity (nuclear (N) or satellite
depen- (S)).
hw1 of
, hw2implicit
, ..., hwm syntax
= biLSTM(x featuresw1 extracted
, xw2 , ..., xwmfrom
) a Bi-Affine (24.17)
dency• Pop
parser.Root (PR),
Results showwhich that thepops
implicit outsyntax
the top tree on
features the stack,
are effective, marking
giving better the decodingthan
performances being completed,
An EDU of span ws , ws+1 , ..., wt then has biLSTM output representation hws , hws+1 , ..., htw ,
explicitwhen
Tree-LSTM
the stack(Li et
holdsal., 2015b).
only one Our codes will
subtreepooling: be released
and the queue is empty. for public under the Apache License
and is represented by average
2.0 at https://github.com/yunan4nlp/NNDisParser.
InGiven
summary,
the weRSTmainly
tree makeas shown the following
in Figure two1, contributions
e it can 1 be X int this
generated
w
work:by (1)thewe following
propose a transition-action sequence: {SH,
based neural RST discourse parsing model with x
dynamic= oracle, (2) hwek compare three different syntactic (24.18)
SH, RD(attr,SN), SH, SH, RD(elab,NS), RD(elab,SN), t − s + 1 PR}. Table 1 shows the decoding
k=s
integration approaches proposed by us. The rest of the paper is organized as follows. Section 2 describes
process in detail.
our proposed modelsThe
By this way,
second layer
including
we naturally
uses this input toneural
the transition-based
convert
compute RST discourse
a final
model, therepresentation
parsing
dynamic oracle of the into predicting
sequence
strategy andofthe
a sequence of
transition
implicit actions,
syntax EDU
feature where eachapproach.
representations
extraction line includes
h : Sectiona 3state
e
presentsand thenext step action
experiments referringourtomodels.
to evaluate the tree.
Section 4 shows the related work. Finally,hsection , h , ...,5hdraws
e e e conclusions.
= biLSTM(x e e
, x , ..., xe ) (24.19)
2.2 Encoder-Decoder 1 2 n 1 2 n

2Previous
Transition-based
The Discourse
RSTParsing
decoder is then
transition-based a feedforward network W that outputs an action o based on a
discourse parsing studies exploit statistical models, using manually-
concatenation of the top three subtrees on the stack (so , s1 , s2 ) plus the first EDU in
designed
We follow Jidiscrete features
and Eisenstein
the (q(Sagae,
queue(2014),0 ):
2009;aHeilman
exploiting and Sagae,
transition-based framework2015; forWang et al., 2017).
RST discourse parsing.In this work, we
The framework
propose is conceptually simple
a transition-based neuraland flexible
model fortoRST
support arbitraryparsing,
discourse features, which
which has been widely
follows an encoder-decoder
o =DyerW(h t , ht , ht , he ) (24.20) a
used in a number
framework. of NLP
Given tasks (Zhu
an input sequenceet al., of
2013;
EDUs {eet1s0al., 2015;
, ...,s2enZhang
, e2s1 q0 theetencoder
}, al., 2016). In addition,
computes the input represen-
transition-based model formalizes a certain task into predicting a sequence e of actions, which is essential
tations {h1 , h2 , ...,
e e e
hn },
where theand the decoder
representation of predicts
the EDU next on thestepqueue actions
h comes conditioned
directly on
fromthetheencoder outputs.
similar to sequence-to-sequence models proposed recently (Bahdanau et q0 al., 2014). In the following,
we first describe encoder, and the three hidden vectors representing partial trees
the transition system for RST discourse parsing, and then introduce our neural network are computed by
2.2.1 Encoder average pooling over the encoder output for the EDUs in those trees:
model by its encoder and decoder parts, respectively. Thirdly, we present our proposed dynamic oracle
We follow
strategy Li toet enhance
aiming al. (2016), using hierarchical
the transition-based model.Bi-LSTMs
Then j to encode the the source method
EDU inputs, where the
t 1 weX introduce integration of
first-layer is used to represent sequencial
implicit syntax features. Finally we describe the training words inside
hs = method of our of EDUs, e and the second layer
hkneural network models.(24.21) is used to represent
j−i+1
sequencial EDUs. Given an input sentence {w1 , w2 , ...,k=i wm }, first we represent each word by its form
2.1 The Transition-based System
(e.g., wi ) and POS tag (e.g. ti ), concatenating their neural embeddings. By this way, the input vectors
The transition-based
of the framework converts
first-layer Bi-LSTM are {xw a structural learning problem
w into a sequence ofemb(t
action predic-
1 , x2 , ..., xm }, where xi = emb(wi ) i ), and then we apply
w w
tions, whose key point is a transition system. A transition system consists of two parts: states and actions.
Bi-LSTM directly, obtaining:
The states are used to store partially-parsed results and the actions are used to control state transitions.
w w w w w w
 24.2 • D ISCOURSE S TRUCTURE PARSING 9

Training first maps each RST gold parse tree into a sequence of oracle actions, and
then uses the standard cross-entropy loss (with l2 regularization) to train the system
to take such actions. Give a state S and oracle action a, we first compute the decoder
output using Eq. 24.20, apply a softmax to get probabilities:

exp(oa )
pa = P (24.22)
exp(oa0 )
a0 ∈A

and then computing the cross-entropy loss:

λ
LCE () = − log(pa ) + ||Θ||2 (24.23)
2
RST discourse parsers are evaluated on the test section of the RST Discourse Tree-
bank, either with gold EDUs or end-to-end, using the RST-Pareval metrics (Marcu,
2000b). It is standard to first transform the gold RST trees into right-branching bi-
nary trees, and to report four metrics: trees with no labels (S for Span), labeled
with nuclei (N), with relations (R), or both (F for Full), for each metric computing
micro-averaged F1 over all spans from all documents (Marcu 2000b, Morey et al.
2017).

24.2.3 PDTB discourse parsing

shallow
PDTB discourse parsing, the task of detecting PDTB coherence relations between
discourse spans, is sometimes called shallow discourse parsing because the task just involves
parsing
flat relationships between text spans, rather than the full trees of RST parsing.
The set of four subtasks for PDTB discourse parsing was laid out by Lin et al.
(2014) in the first complete system, with separate tasks for explicit (tasks 1-3) and
implicit (task 4) connectives:
1. Find the discourse connectives (disambiguating them from non-discourse uses)
2. Find the two spans for each connective
3. Label the relationship between these spans
4. Assign a relation between every adjacent pair of sentences
Many systems have been proposed for Task 4: taking a pair of adjacent sentences
as input and assign a coherence relation sense label as output. The setup often fol-
lows Lin et al. (2009) in assuming gold sentence span boundaries and assigning each
adjacent span one of the 11 second-level PDTB tags or none (removing the 5 very
rare tags of the 16 shown in italics in Fig. 24.3).
A simple but very strong algorithm for Task 4 is to represent each of the two
spans by BERT embeddings and take the last layer hidden state corresponding to
the position of the [CLS] token, pass this through a single layer tanh feedforward
network and then a softmax for sense classification (Nie et al., 2019).
Each of the other tasks also have been addressed. Task 1 is to disambiguat-
ing discourse connectives from their non-discourse use. For example as Pitler and
Nenkova (2009) point out, the word and is a discourse connective linking the two
clauses by an elaboration/expansion relation in (24.24) while it’s a non-discourse
NP conjunction in (24.25):
(24.24) Selling picked up as previous buyers bailed out of their positions and
aggressive short sellers—anticipating further declines—moved in.
(24.25) My favorite colors are blue and green.
10 C HAPTER 24 • D ISCOURSE C OHERENCE

Similarly, once is a discourse connective indicating a temporal relation in (24.26),

but simply a non-discourse adverb meaning ‘formerly’ and modifying used in (24.27):
(24.26) The asbestos fiber, crocidolite, is unusually resilient once it enters the
lungs, with even brief exposures to it causing symptoms that show up
decades later, researchers said.
(24.27) A form of asbestos once used to make Kent cigarette filters has caused a
high percentage of cancer deaths among a group of workers exposed to it
more than 30 years ago, researchers reported.
Determining whether a word is a discourse connective is thus a special case
of word sense disambiguation. Early work on disambiguation showed that the 4
PDTB high-level sense classes could be disambiguated with high (94%) accuracy
used syntactic features from gold parse trees (Pitler and Nenkova, 2009). Recent
work performs the task end-to-end from word inputs using a biLSTM-CRF with
BIO outputs (B - CONN, I - CONN, O) (Yu et al., 2019).
For task 2, PDTB spans can be identified with the same sequence models used to
find RST EDUs: a biLSTM sequence model with pretrained contextual embedding
(BERT) inputs (Muller et al., 2019). Simple heuristics also do pretty well as a base-
line at finding spans, since 93% of relations are either completely within a single
sentence or span two adjacent sentences, with one argument in each sentence (Biran
and McKeown, 2015).

24.3 Centering and Entity-Based Coherence

A second way a discourse can be coherent is by virtue of being “about” some entity.
This idea that at each point in the discourse some entity is salient, and a discourse
is coherent by continuing to discuss the same entity, appears early in functional lin-
guistics and the psychology of discourse (Chafe 1976, Kintsch and Van Dijk 1978),
and soon made its way to computational models. In this section we introduce two
entity-based models of this kind of entity-based coherence: Centering Theory (Grosz et al.,
1995), and the entity grid model of Barzilay and Lapata (2008).

24.3.1 Centering
Centering
Theory Centering Theory (Grosz et al., 1995) is a theory of both discourse salience and
discourse coherence. As a model of discourse salience, Centering proposes that at
any given point in the discourse one of the entities in the discourse model is salient:
it is being “centered” on. As a model of discourse coherence, Centering proposes
that discourses in which adjacent sentences CONTINUE to maintain the same salient
entity are more coherent than those which SHIFT back and forth between multiple
entities (we will see that CONTINUE and SHIFT are technical terms in the theory).
The following two texts from Grosz et al. (1995) which have exactly the same
propositional content but different saliences, can help in understanding the main
Centering intuition.
(24.28) a. John went to his favorite music store to buy a piano.
b. He had frequented the store for many years.
c. He was excited that he could finally buy a piano.
d. He arrived just as the store was closing for the day.
 24.3 • C ENTERING AND E NTITY-BASED C OHERENCE 11

(24.29) a. John went to his favorite music store to buy a piano.

b. It was a store John had frequented for many years.
c. He was excited that he could finally buy a piano.
d. It was closing just as John arrived.
While these two texts differ only in how the two entities (John and the store) are
realized in the sentences, the discourse in (24.28) is intuitively more coherent than
the one in (24.29). As Grosz et al. (1995) point out, this is because the discourse
in (24.28) is clearly about one individual, John, describing his actions and feelings.
The discourse in (24.29), by contrast, focuses first on John, then the store, then back
to John, then to the store again. It lacks the “aboutness” of the first discourse.
Centering Theory realizes this intuition by maintaining two representations for
backward-
looking each utterance Un . The backward-looking center of Un , denoted as Cb (Un ), rep-
center
resents the current salient entity, the one being focused on in the discourse after Un
forward-looking
center
is interpreted. The forward-looking centers of Un , denoted as C f (Un ), are a set
of potential future salient entities, the discourse entities evoked by Un any of which
could serve as Cb (the salient entity) of the following utterance, i.e. Cb (Un+1 ).
The set of forward-looking centers C f (Un ) are ranked according to factors like
discourse salience and grammatical role (for example subjects are higher ranked
than objects, which are higher ranked than all other grammatical roles). We call the
highest-ranked forward-looking center C p (for “preferred center”). C p is a kind of
prediction about what entity will be talked about next. Sometimes the next utterance
indeed talks about this entity, but sometimes another entity becomes salient instead.
We’ll use here the algorithm for centering presented in Brennan et al. (1987),
which defines four intersentential relationships between a pair of utterances Un and
Un+1 that depend on the relationship between Cb (Un+1 ), Cb (Un ), and C p (Un+1 );
these are shown in Fig. 24.7.

Cb (Un+1 ) = Cb (Un ) Cb (Un+1 ) 6= Cb (Un )

or undefined Cb (Un )
Cb (Un+1 ) = C p (Un+1 ) Continue Smooth-Shift
Cb (Un+1 ) 6= C p (Un+1 ) Retain Rough-Shift
Figure 24.7 Centering Transitions for Rule 2 from Brennan et al. (1987).

The following rules are used by the algorithm:

Rule 1: If any element of C f (Un ) is realized by a pronoun in utterance

Un+1 , then Cb (Un+1 ) must be realized as a pronoun also.
Rule 2: Transition states are ordered. Continue is preferred to Retain is
preferred to Smooth-Shift is preferred to Rough-Shift.

Rule 1 captures the intuition that pronominalization (including zero-anaphora)

is a common way to mark discourse salience. If there are multiple pronouns in an
utterance realizing entities from the previous utterance, one of these pronouns must
realize the backward center Cb ; if there is only one pronoun, it must be Cb .
Rule 2 captures the intuition that discourses that continue to center the same en-
tity are more coherent than ones that repeatedly shift to other centers. The transition
table is based on two factors: whether the backward-looking center Cb is the same
from Un to Un+1 and whether this discourse entity is the one that is preferred (C p )
in the new utterance Un+1 . If both of these hold, a CONTINUE relation, the speaker
has been talking about the same entity and is going to continue talking about that
12 C HAPTER 24 • D ISCOURSE C OHERENCE

entity. In a RETAIN relation, the speaker intends to SHIFT to a new entity in a future
utterance and meanwhile places the current entity in a lower rank C f . In a SHIFT
relation, the speaker is shifting to a new salient entity.
Let’s walk though the start of (24.28) again, repeated as (24.30), showing the
representations after each utterance is processed.
(24.30) John went to his favorite music store to buy a piano. (U1 )
He was excited that he could finally buy a piano. (U2 )
He arrived just as the store was closing for the day. (U3 )
It was closing just as John arrived (U4 )
Using the grammatical role hierarchy to order the C f , for sentence U1 we get:
C f (U1 ): {John, music store, piano}
C p (U1 ): John
Cb (U1 ): undefined
and then for sentence U2 :
C f (U2 ): {John, piano}
C p (U2 ): John
Cb (U2 ): John
Result: Continue (C p (U2 )=Cb (U2 ); Cb (U1 ) undefined)
The transition from U1 to U2 is thus a CONTINUE. Completing this example is left
as exercise (1) for the reader

24.3.2 Entity Grid model

Centering embodies a particular theory of how entity mentioning leads to coher-
ence: that salient entities appear in subject position or are pronominalized, and that
discourses are salient by means of continuing to mention the same entity in such
ways.
entity grid The entity grid model of Barzilay and Lapata (2008) is an alternative way to
capture entity-based coherence: instead of having a top-down theory, the entity-grid
model using machine learning to induce the patterns of entity mentioning that make
a discourse more coherent.
The model is based around an entity grid, a two-dimensional array that repre-
sents the distribution of entity mentions across sentences. The rows represent sen-
tences, and the columns represent discourse entities (most versions of the entity grid
model focus just on nominal mentions). Each cell represents the possible appearance
of an entity in a sentence, and the values represent whether the entity appears and its
grammatical role. Grammatical roles are subject (S), object (O), neither (X), or ab-
sent (–); in the implementation of Barzilay and Lapata (2008), subjects of passives
are represented with O, leading to a representation with some of the characteristics
of thematic roles.
Fig. 24.8 from Barzilay and Lapata (2008) shows a grid for the text shown in
Fig. 24.9. There is one row for each of the six sentences. The second column, for
the entity ‘trial’, is O – – – X, showing that the trial appears in the first sentence as
direct object, in the last sentence as an oblique, and does not appear in the middle
sentences. The third column, for the entity Microsoft, shows that it appears as sub-
ject in sentence 1 (it also appears as the object of the preposition against, but entities
that appear multiple times are recorded with their highest-ranked grammatical func-
tion). Computing the entity grids requires extracting entities and doing coreference
 present in sentences 1 and 6 (as O and X, respectively) but is absent from the rest of the
sentences. Also note that the grid in Table 1 takes coreference resolution into account.
Even though the same entity appears in different linguistic forms, for example, Microsoft
Corp., Microsoft, and the company , it is mapped to a single entry in the grid (see the
column introduced by Microsoft in Table 1).
Computational Linguistics Volume 34, Number 1

a feature space with

24.3transitions
• C
Table
of length two is illustrated in Table 3. The second13row
1 ENTERING AND E NTITY-BASED C OHERENCE
(introduced by d1 ) is the
A feature
fragment vector representation
of the entity of thearegrid
grid. Noun phrases in Table
represented by1.
their head nouns. Grid cells
correspond to grammatical roles: subjects (S), objects (O), or neither (X).

Government
Competitors
Department
3.3 Grid Construction: Linguistic Dimensions

Microsoft

Netscape
Evidence

Earnings
Products

Software
Markets

Brands

Tactics
Case
Trial

Suit
One of the central research issues in developing entity-based models of coherence is
determining what sources 1 S ofO linguistic
S X O – –knowledge
– – – – – are – –essential
– 1 for accurate prediction,
and how to encode them 2 succinctly
– – O – – in X aS discourse – – – 2
O – – – – representation. Previous approaches
3 – – S O – – – – S O O – – – – 3
tend to agree on the 4features – – S of – –entity
– – – distribution
– – – S – – related
– 4 to local coherence—the
disagreement
Barzilay lies in the5 way
and Lapata – – these
– – –features
– – – –are – modeled.
– – S O – 5 Modeling Local Coherence
Our study of alternative 6 – X S encodings
– – – – – is – –not – O 6duplication of previous ef-
– –a –mere
forts (Poesio
Figure 24.8 Part et al.
of2004) that grid
the entity focusforonthe
linguistic aspects
text in Fig. 24.9. of parameterization.
Entities Because
are listed by their head we
are interested in an automatically
noun; each cell represents whether constructed model, we have to take into account
an entity appears as subject (S), object (O), neither (X), or com-
6
putational and learning issues when
is absent (–). Figure from Barzilay and Lapata (2008).
Table 2 considering alternative representations. Therefore,
our exploration
Summary augmented of thewithparameter space is guided
syntactic annotations bycomputation.
for grid three considerations: the linguistic
importance of a parameter, the accuracy of its automatic computation, and the size of the
1resulting
[The Justice Department]
feature space. From S is conducting an [anti-trust trial] O against [Microsoft Corp.] X
the linguistic side, we focus on properties of entity distri-
with [evidence]X that [the company]S is increasingly attempting to crush [competitors]O .
bution that are tightly linked
2 [Microsoft]O is accused of trying to to
local coherence,
forcefully and[markets]
buy into at the same time allow for multiple
X where [its own
interpretations
products]S are notduring the encoding
competitive enoughprocess.
to unseatComputational considerations
[established brands] O.
prevent us
3from
[Theconsidering
case]S revolves discourse representations
around [evidence] that cannot
O of [Microsoft] be computed
S aggressively reliably by exist-
pressuring
[Netscape]
ing tools. For O into merging
instance, we[browser
could not software] O.
experiment with the granularity of an utterance—
4sentence
[Microsoft] S claims [its tactics] S are commonplace and good economically.
versus clause—because available clause separators introduce substantial noise
5 [The government]S may file [a civil suit]O ruling that [conspiracy]S to curb [competition]O
into a grid[collusion]
through construction. Finally, we exclude representations that will explode the size of
X is [a violation of the Sherman Act] O .
6the feature space,
[Microsoft] S
thereby
continues to increasing
show the earnings]
[increased amount Oofdespite
data required
[the trial]for
X.
training the model.

Figure
Entity24.9 A discourse
Ex traction. with thecomputation
The accurate entities marked and annotated
of entity classes iswithkey grammatical
to computing func-
mean-
tions.
ingfulFigure
When entity from
a noun Barzilay
grids. is and Lapata
Inattested
previous more (2008).
than once with
implementations a different grammatical
of entity-based models, classes roleof in the
coref-
same
erent sentence,
nouns have we default
been extractedto the role with the (Miltsakaki
manually highest grammatical
and Kukich ranking:
2000; subjects
Karamanis are
ranked
resolution
et al. 2004;higher
to clusterthan
Poesio them objects,
et al. into which
2004), in
discourse turn are
but thisentities ranked
is not (Chapter higher
an option23) than
forasour the
well rest. For
as parsing
model. example,
the
An obvious
the entityto
solution
sentences forMicrosoft
identifying
get is mentioned
grammatical entity twiceisintoSentence
classes
roles. employ an 1 with the grammatical
automatic coreferenceroles x (for
resolution
Microsoft
tool that Corp.) and s
determines (for the
which noun company
phrases ), but
refer istorepresented
the same only in
entity bya sdocument.
in the grid (see
In the1 and
Tables
resulting grid, columns that are dense (like the column for Microsoft) in-
Current 2). approaches recast coreference resolution as a classification task. A pair
dicate entities that are mentioned often in the texts; sparse columns (like the column
of NPs is classified as coreferring or not based on constraints that are learned from
foranearnings)
annotated indicate
corpus. entities
A separate that are mentioned rarely.
3.2 Entity Grids as Feature Vectorsclustering mechanism then coordinates the possibly
In the entity pairwise
contradictory grid model, coherence is
classifications andmeasured
constructs by apatterns
partition of local
on theentity
set oftran-NPs. In
sition. For example, Department is a subject in sentence
A fundamental assumption underlying our approach is that the distribution of system.
our experiments, we employ Ng and Cardie’s (2002) 1,
coreference and then not
resolution men-
entities
tioned
The
in in sentence
system
coherent decides
texts 2; this
whether
exhibits iscertain
thetwo transition
NPs are[ Scoreferent
regularities –]. The transitions
reflected by
in exploiting are athus
grid topology. sequences
wealth
Some ofoflexical,
these
, O X , –}n which
grammatical,
{Sregularities aresemantic,
can be
formalized and positional
extracted
in as features.
Centering continuous
Theory It is trained
ascells on the
from
constraints each MUC
on (6–7) data
column.
transitions Eachof sets
the
and yields
transition
local focus hasstate-of-the-art
ina adjacent
probability; performance
sentences. Grids (70.4
the probability of of F-measure theongrid
[ S –] intexts
coherent areMUC-6
fromand
likely Fig.
to 63.4
24.8
have onisMUC-7).
some 0.08dense
(itcolumns
occurs 6(i.e., times columns
out of the with 75just
totala transitions
few gaps, such as Microsoft
of length two). Fig. in 24.10
Table shows
1) and the many
distribution over transitions of length 2 for the text of Fig. 24.9 (shown as the first 1).
sparse columns which will consist mostly of gaps (see markets and earnings in Table
One
row d1would
Table ),3 and 2further expect that entities corresponding to dense columns are more often
other documents.
subjects
Example or of aobjects. These characteristics
feature-vector document representationwill be less pronounced
using all transitionsin low-coherence
of length two given texts.
syntactic
Inspiredcategories S , O , X , and
by Centering –.
Theory, our analysis revolves around patterns of local entity
transitions. A local entity transition is a sequence {S, O, X, –}n that represents entity
SS SO SX S– OS OO OX O– XS XO XX X– –S –O –X ––
occurrences and their syntactic roles in n adjacent sentences. Local transitions can be
d1 .01
easily .01 0 from.08
obtained a grid .01as0continuous
0 .09subsequences
0 0 0 of each
.03 column.
.05 .07 Each .03 transition
.59
d2 have
will .02 .01 .01 .02
a certain 0
probability .07 in0 a given
.02 grid.
.14 For .14 instance,
.06 .04 .03 .07 0.1 .36of the
the probability
d3 .02 0 0 .03 .09 0 .09 .06 0 0 0 .05 .03 .07 .17 .39
transition [ S –] in the grid from Table 1 is 0.08 (computed as a ratio of its frequency
[i.e., six]
Figure 24.10 divided by the
A feature totalfor
vector number of transitions
representing documentsofusing length all two [i.e., 75]).
transitions Each2.text
of length
can thusdbe
Document 1 isviewed
the textas inaFig.
distribution
24.9. Figure defined over transition
from Barzilay and Lapatatypes.(2008).
8
We can now go one step further and represent each text by a fixed set of transition
sequences using a standard feature vector notation. Each grid rendering j of a document
The transitions and their probabilities can then be used as features for a machine
di corresponds to a feature vector Φ(x ij ) = (p1 (x ij ), p2 (x ij ), . . . , pm (x ij )), where m is the
learning model. This model can be a text classifier trained to produce human-labeled
number of all predefined entity transitions, and pt (x ij ) the probability of transition t
coherence
in grid x ijscores (for example
. This feature vector from humans labeling
representation is usefully each text as coherent
amenable to machine or learning
inco-
herent).
algorithms But such (see data is expensive in
our experiments to gather.
SectionsBarzilay and Lapata (2005)
4–6). Furthermore, it allows introduced
the consid-
a eration
simplifying of large innovation:
numbers of coherence
transitions models
whichcan couldbe potentially
trained by uncoverself-supervision:
novel entity
distribution
trained patterns relevant
to distinguish the natural for coherence
original order assessment or other in
of sentences coherence-related
a discourse from tasks.
Note that considerable latitude is available when specifying the transition types to
be included in a feature vector. These can be all transitions of a given length (e.g., two
or three) or the most frequent transitions within a document collection. An example of

7
14 C HAPTER 24 • D ISCOURSE C OHERENCE

a modified order (such as a randomized order). We turn to these evaluations in the

next section.

24.3.3 Evaluating Neural and Entity-based coherence

Entity-based coherence models, as well as the neural models we introduce in the
next section, are generally evaluated in one of two ways.
First, we can have humans rate the coherence of a document and train a classifier
to predict these human ratings, which can be categorial (high/low, or high/mid/low)
or continuous. This is the best evaluation to use if we have some end task in mind,
like essay grading, where human raters are the correct definition of the final label.
Alternatively, since it’s very expensive to get human labels, and we might not
yet have an end-task in mind, we can use natural texts to do self-supervision. In
self-supervision we pair up a natural discourse with a pseudo-document created by
changing the ordering. Since naturally-ordered discourses are more coherent than
random permutation (Lin et al., 2011), a successful coherence algorithm should pre-
fer the original ordering.
Self-supervision has been implemented in 3 ways. In the sentence order dis-
crimination task (Barzilay and Lapata, 2005), we compare a document to a random
permutation of its sentences. A model is considered correct for an (original, per-
muted) test pair if it ranks the original document higher. Given k documents, we can
compute n permutations, resulting in kn pairs each with one original document and
one permutation, to use in training and testing.
In the sentence insertion task (Chen et al., 2007) we take a document, remove
one of the n sentences s, and create n − 1 copies of the document with s inserted into
each position. The task is to decide which of the n documents is the one with the
original ordering, distinguishing the original position for s from all other positions.
Insertion is harder than discrimination since we are comparing documents that differ
by only one sentence.
Finally, in the sentence order reconstruction task (Lapata, 2003), we take a
document, randomize the sentences, and train the model to put them back in the
correct order. Again given k documents, we can compute n permutations, resulting
in kn pairs each with one original document and one permutation, to use in training
and testing. Reordering is of course a much harder task than simple classification.

24.4 Representation learning models for local coherence

The third kind of local coherence is topical or semantic field coherence. Discourses
cohere by talking about the same topics and subtopics, and drawing on the same
semantic fields in doing so.
The field was pioneered by a series of unsupervised models in the 1990s of this
lexical cohesion kind of coherence that made use of lexical cohesion (Halliday and Hasan, 1976):
the sharing of identical or semantically related words in nearby sentences. Morris
and Hirst (1991) computed lexical chains of words (like pine, bush trees, trunk) that
occurred through a discourse and that were related in Roget’s Thesaurus (by being in
the same category, or linked categories). They showed that the number and density
TextTiling of chain correlated with the topic structure. The TextTiling algorithm of Hearst
(1997) computed the cosine between neighboring text spans (the normalized dot
product of vectors of raw word counts), again showing that sentences or paragraph in
24.4 • R EPRESENTATION LEARNING MODELS FOR LOCAL COHERENCE 15

a subtopic have high cosine with each other, but not with sentences in a neighboring
subtopic.
A third early model, the LSA Coherence method of Foltz et al. (1998) was the
first to use embeddings, modeling the coherence between two sentences as the co-
sine between their LSA sentence embedding vectors1 , computing embeddings for a
sentence s by summing the embeddings of its words w:
sim(s,t) = cos(s, t)
X X
= cos( w, w) (24.31)
w∈s w∈t

and defining the overall coherence of a text as the average similarity over all pairs of
adjacent sentences si and si+1 :
n−1
1 X
coherence(T ) = cos(si , si+1 ) (24.32)
n−1
i=1

Modern neural representation-learning coherence models, beginning with Li et al.

(2014), draw on the intuitions of these early unsupervised models for learning sen-
tence representations and measuring how they change between neighboring sen-
tences. But the new models also draw on the idea pioneered by Barzilay and Lapata
(2005) of self-supervision. That is, unlike say coherence relation models, which
train on hand-labeled representations for RST or PDTB, these models are trained to
distinguish natural discourses from unnatural discourses formed by scrambling the
order of sentences, thus using representation learning to discover the features that
matter for at least the ordering aspect of coherence.
Here we present one such model, the local coherence discriminator (LCD) (Xu
et al., 2019). Like early models, LCD computes the coherence of a text as the av-
erage of coherence scores between consecutive pairs of sentences. But unlike the
early unsupervised models, LCD is a self-supervised model trained to discriminate
consecutive sentence pairs (si , si+1 ) in the training documents (assumed to be coher-
ent) from (constructed) incoherent pairs (si , s0 ). All consecutive pairs are positive
examples, and the negative (incoherent) partner for a sentence si is another sentence
uniformly sampled from the same document as si .
Fig. 24.11 describes the architecture of the model fθ , which takes a sentence
pair and returns a score, higher scores for more coherent pairs. Given an input
sentence pair s and t, the model computes sentence embeddings s and t (using any
sentence embeddings algorithm), and then concatenates four features of the pair: (1)
the concatenation of the two vectors (2) their difference s − t; (3) the absolute value
of their difference |s − t|; (4) their element-wise product s t. These are passed
through a one-layer feedforward network to output the coherence score.
The model is trained to make this coherence score higher for real pairs than for
negative pairs. More formally, the training objective for a corpus C of documents d,
each of which consists of a list of sentences si , is:
XX
0
Lθ = E [L( fθ (si , si+1 ), fθ (si , s ))] (24.33)
p(s0 |si )
d∈C si ∈d

E p(s0 |si ) is the expectation with respect to the negative sampling distribution con-
ditioned on si : given a sentence si the algorithms samples a negative sentence s0
1 See Chapter 6 for more on LSA embeddings; they are computed by applying SVD to the term-
document matrix (each cell weighted by log frequency and normalized by entropy), and then the first
300 dimensions are used as the embedding.
 16 C HAPTER 24 • D ISCOURSE C OHERENCE

tion: The role of the loss function is

ge f + = f✓ (si , si+1 ) to be high while
0
i , s ) to be low. Common losses such as
og loss can all be used. Through exper-
lidation, we found that margin loss to
r for this problem. Specifically, L takes
m: L(f + , f ) = max(0, ⌘ f + + f )
the margin hyperparameter.

samples: Technically, we are free to

y sentence s0 to form a negative pair
However, because of potential differ-
enre, topic and writing style, such neg-
Figure 24.11 The architecture of the LCD model of document coherence, showing the
ht cause the discriminative model to ofFigure
computation 1: for
the score Generic
a pair ofarchitecture
sentences s andfor t.our proposed
Figure from Xu model.
et al. (2019).
unrelated to coherence. Therefore, we
sentences from the same documentuniformlyto over the other sentences in the same document. L is a loss function that
negative pairs. Specifically, suppose si scores,
takes two 4.2 one Pre-trained
for a positiveGenerative
pair and one for Model as the
a negative pair, with the goal of
encouraging f + = fSentence Encoder
θ (si , si+1 ) to be high and f − = fθ (si , s0 )) to be low. Fig. 24.11
m document dk with length nuse k , then
the margin loss l( f + , f − ) = max(0, η − f + + f − ) where η is the margin hyper-
a uniform distribution over theparameter.
nk 1 Our model can work with any pre-trained sen-
{sj }j 6= i from dk . For a document withXu et al. tence
(2019) encoder,
also give aranging from the
useful baseline most that
algorithm simplistic
itself has quite high
es, there are n 1 positive pairs, and
performance average
in measuring GloVe (Pennington
perplexity: train anetRNN al., 2014)
language embed-
model on the data,
and
n 2)/2 negative pairs. It turns out that compute dings to more sophisticated supervised or unsu-the preceding
the log likelihood of sentence s i in two ways, once given
context (conditional log likelihood) and once with no context (marginal log likeli-
tic number of negatives provides a rich pervised pre-trained sentence encoders (Conneau
hood). The difference between these values tells us how much the preceding context
rning signal, while at the same improved
time, is theet al., 2017).of As
predictability si , a mentioned
predictabilityinmeasure
the introduction,
of coherence.
ohibitively large to be effectively Training
cov- since generative models can
models to predict longer contexts than just often be turned into pairs of sen-
consecutive
sampling procedure. In practice, tenceswe sentence
can result in evenencoder,
stronger generative coherence model
discourse representations. can
For example a Trans-
former
new set of negatives each time we see language model trained with a contrastive
be leveraged by our model to benefit from the sentence objective to predict text
up
t, hence after many epochs, we can ef-to a distance of ±2 sentences improves performance
advantages of both generative and discriminative on various discourse coher-
ence tasks (Iter et al., 2020).
over the space for even very long doc- training, similar to (Kiros et al., 2015; Peters et al.,
Language-model style models are generally evaluated by the methods of Sec-
ection 5.7 discusses further details on 2018). they
tion 24.3.3, although Aftercaninitialization,
also be evaluated we on freeze the and
the RST genera-
PDTB coherence
relation tasks.tive model parameters to avoid overfitting.
In Section 5, we will experimentally show that
el Architecture while we do benefit from strong pre-trained en-
24.5 Global
c neural architecture that we use for f✓
Coherence
coders, the fact that our local discriminative model
ed in Figure 1. We assume the use of improves over previous methods is independent of
rained sentence encoder, whichAisdiscourse
dis- must the choice of sentence
also cohere globally encoder.
rather than just at the level of pairs of sen-
he next section. tences. Consider stories, for example. The narrative structure of stories is one of
the oldest kinds 5 ofExperiments
global coherence to be studied. In his influential Morphology of
n input sentence pair, the sentence en-
the Folktale, Propp (1968) models the discourse structure of Russian folktales via
s the sentences to real-valued vectors
a kind S of plot5.1
grammar. His model includes a set of character categories he called
Evaluation Tasks
then compute the concatenationdramatis
of the personae, like Hero, Villain, Donor, or Helper, and a set of events he
features: (1) concatenation of thecalled two Following
functions Nguyen
(like “Villain commits and kidnapping”,
Joty (2017) “Donor and other testspre-
Hero”, or “Hero
T ); (2) element-wise differenceisSpursued”)
T; that have
vious to occur
work, in particular
we evaluate ourorder,
models alongonwiththeother
dis- components.
Propp
t-wise product S ⇤ T ; (4) absolute valueshows crimination
that the plots of each
and of the fairy
insertion tales Additionally,
tasks. he studies can be werepresented as
-wise difference |S T |. The concate- evaluate on the paragraph reconstruction task in
ure representation is then fed to a one- open-domain settings, in a similar manner to Li
to output the coherence score. and Jurafsky (2017).
ice, we make our overall coherence In the discrimination task, a document is com-
 24.5 • G LOBAL C OHERENCE 17

a sequence of these functions, different tales choosing different subsets of functions,

but always in the same order. Indeed Lakoff (1972) showed that Propp’s model
amounted to a discourse grammar of stories, and in recent computational work Fin-
layson (2016) demonstrates that some of these Proppian functions could be induced
from corpora of folktale texts by detecting events that have similar actions across
stories. Bamman et al. (2013) showed that generalizations over dramatis personae
could be induced from movie plot summaries on Wikipedia. Their model induced
latent personae from features like the actions the character takes (e.g., Villains stran-
gle), the actions done to them (e.g., Villains are foiled and arrested) or the descriptive
words used of them (Villains are evil).
In this section we introduce two kinds of such global discourse structure that
have been widely studied computationally. The first is the structure of arguments:
the way people attempt to convince each other in persuasive essays by offering
claims and supporting premises. The second is somewhat related: the structure of
scientific papers, and the way authors present their goals, results, and relationship to
prior work in their papers.

24.5.1 Argumentation Structure

The first type of global discourse structure is the structure of arguments. Analyzing
argumentation
mining people’s argumentation computationally is often called argumentation mining.
The study of arguments dates back to Aristotle, who in his Rhetorics described
pathos three components of a good argument: pathos (appealing to the emotions of the
ethos listener), ethos (appealing to the speaker’s personal character), and logos (the logical
logos structure of the argument).
Most of the discourse structure studies of argumentation have focused on logos,
particularly via building and training on annotated datasets of persuasive essays or
other arguments (Reed et al. 2008, Stab and Gurevych 2014a, Peldszus and Stede
2016, Habernal and Gurevych 2017, Musi et al. 2018). Such corpora, for exam-
claims ple, often include annotations of argumentative components like claims (the central
premises component of the argument that is controversial and needs support) and premises
(the reasons given by the author to persuade the reader by supporting or attacking
argumentative the claim or other premises), as well as the argumentative relations between them
relations
like SUPPORT and ATTACK.
Consider the following example of a persuasive essay from Stab and Gurevych
(2014b). The first sentence (1) presents a claim (in bold). (2) and (3) present two
premises supporting the claim. (4) gives a premise supporting premise (3).
“(1) Museums and art galleries provide a better understanding
about arts than Internet. (2) In most museums and art galleries, de-
tailed descriptions in terms of the background, history and author are
provided. (3) Seeing an artwork online is not the same as watching it
with our own eyes, as (4) the picture online does not show the texture
or three-dimensional structure of the art, which is important to study.”
Thus this example has three argumentative relations: SUPPORT(2,1), SUPPORT(3,1)
and SUPPORT(4,3). Fig. 24.12 shows the structure of a much more complex argu-
ment.
While argumentation mining is clearly related to rhetorical structure and other
kinds of coherence relations, arguments tend to be much less local; often a persua-
sive essay will have only a single main claim, with premises spread throughout the
text, without the local coherence we see in coherence relations.
 military purposes]Claim6 , I strongly believe that [this technology is beneficial to
humanity]MajorClaim2 . It is likely that [this technology bears some important cures which
will significantly improve life conditions]Claim7 .
The conclusion of the essay starts with an attacking claim followed by the restatement of
the major claim. The last sentence includes another claim that summarizes the most im-
portant points of the author’s argumentation. Figure 2 shows the entire argumentation
structure of the example essay.
18 C HAPTER 24 • D ISCOURSE C OHERENCE

Figure
Figure 2
24.12 Argumentation structure of a persuasive essay. Arrows indicate argumentation relations, ei-
Argumentation structure of the example essay. Arrows indicate argumentative relations.
ther of SUPPORT (with arrowheads) or ATTACK (with circleheads); P denotes premises. Figure from Stab and
Arrowheads denote argumentative support relations and circleheads attack relations. Dashed
Gurevych (2017).relations that are encoded in the stance attributes of claims. “P” denotes premises.
lines indicate

Algorithms for detecting argumentation structure often include classifiers for

distinguishing claims, premises, or non-argumentation, together with relation629 clas-
sifiers for deciding if two spans have the SUPPORT, ATTACK, or neither relation
(Peldszus and Stede, 2013). While these are the main focus of much computational
work, there is also preliminary efforts on annotating and detecting richer semantic
relationships (Park and Cardie 2014, Hidey et al. 2017) such as detecting argumen-
argumentation tation schemes, larger-scale structures for argument like argument from example,
schemes
or argument from cause to effect, or argument from consequences (Feng and
Hirst, 2011).
Another important line of research is studying how these argument structure (or
other features) are associated with the success or persuasiveness of an argument
(Habernal and Gurevych 2016, Tan et al. 2016, Hidey et al. 2017. Indeed, while it
is Aristotle’s logos that is most related to discourse structure, Aristotle’s ethos and
pathos techniques are particularly relevant in the detection of mechanisms of this
persuasion sort of persuasion. For example scholars have investigated the linguistic realization
of features studied by social scientists like reciprocity (people return favors), social
proof (people follow others’ choices), authority (people are influenced by those
with power), and scarcity (people value things that are scarce), all of which can
be brought up in a persuasive argument (Cialdini, 1984). Rosenthal and McKeown
(2017) showed that these features could be combined with argumentation structure
to predict who influences whom on social media, Althoff et al. (2014) found that
linguistic models of reciprocity and authority predicted success in online requests,
while the semisupervised model of Yang et al. (2019) detected mentions of scarcity,
commitment, and social identity to predict the success of peer-to-peer lending plat-
forms.
See Stede and Schneider (2018) for a comprehensive survey of argument mining.

24.5.2 The structure of scientific discourse

Scientific papers have a very specific global structure: somewhere in the course of
the paper the authors must indicate a scientific goal, develop a method for a solu-
tion, provide evidence for the solution, and compare to prior work. One popular
 24.6 • S UMMARY 19

annotation scheme for modeling these rhetorical goals is the argumentative zon-
argumentative
zoning ing model of Teufel et al. (1999) and Teufel et al. (2009), which is informed by the
idea that each scientific paper tries to make a knowledge claim about a new piece
of knowledge being added to the repository of the field (Myers, 1992). Sentences
in a scientific paper can be assigned one of 15 tags; Fig. 24.13 shows 7 (shortened)
examples of labeled sentences.

Category Description Example

A IM Statement of specific research goal, or“The aim of this process is to examine the role that
hypothesis of current paper training plays in the tagging process”
OWN M ETHOD New Knowledge claim, own work: “In order for it to be useful for our purposes, the
methods following extensions must be made:”
OWN R ESULTS Measurable/objective outcome of own “All the curves have a generally upward trend but
work always lie far below backoff (51% error rate)”
U SE Other work is used in own work “We use the framework for the allocation and
transfer of control of Whittaker....”
G AP W EAK Lack of solution in field, problem with “Here, we will produce experimental evidence
other solutions suggesting that this simple model leads to serious
overestimates”
S UPPORT Other work supports current work or is “Work similar to that described here has been car-
supported by current work ried out by Merialdo (1994), with broadly similar
conclusions.”
A NTISUPPORT Clash with other’s results or theory; su- “This result challenges the claims of...”
periority of own work
Figure 24.13 Examples for 7 of the 15 labels from the Argumentative Zoning labelset (Teufel et al., 2009).

Teufel et al. (1999) and Teufel et al. (2009) develop labeled corpora of scientific
articles from computational linguistics and chemistry, which can be used as supervi-
sion for training standard sentence-classification architecture to assign the 15 labels.

24.6 Summary
In this chapter we introduced local and global models for discourse coherence.
• Discourses are not arbitrary collections of sentences; they must be coherent.
Among the factors that make a discourse coherent are coherence relations
between the sentences, entity-based coherence, and topical coherence.
• Various sets of coherence relations and rhetorical relations have been pro-
posed. The relations in Rhetorical Structure Theory (RST) hold between
spans of text and are structured into a tree. Because of this, shift-reduce
and other parsing algorithms are generally used to assign these structures.
The Penn Discourse Treebank (PDTB) labels only relations between pairs of
spans, and the labels are generally assigned by sequence models.
• Entity-based coherence captures the intuition that discourses are about an
entity, and continue mentioning the entity from sentence to sentence. Cen-
tering Theory is a family of models describing how salience is modeled for
discourse entities, and hence how coherence is achieved by virtue of keeping
the same discourse entities salient over the discourse. The entity grid model
gives a more bottom-up way to compute which entity realization transitions
lead to coherence.
20 C HAPTER 24 • D ISCOURSE C OHERENCE

• Many different genres have different types of global coherence. Persuasive

essays have claims and premises that are extracted in the field of argument
mining, scientific articles have structure related to aims, methods, results, and
comparisons.

Bibliographical and Historical Notes

Coherence relations arose from the independent development of a number of schol-
ars, including Hobbs (1979) idea that coherence relations play an inferential role for
the hearer, and the investigations by Mann and Thompson (1987) of the discourse
structure of large texts. Other approaches to coherence relations and their extrac-
SDRT tion include Segmented Discourse Representation Theory (SDRT) (Asher and Las-
carides 2003, Baldridge et al. 2007) and the Linguistic Discourse Model (Polanyi
1988, Scha and Polanyi 1988, Polanyi et al. 2004). Wolf and Gibson (2005) argue
that coherence structure includes crossed bracketings, which make it impossible to
represent as a tree, and propose a graph representation instead. A compendium of
over 350 relations that have been proposed in the literature can be found in Hovy
(1990).
RST parsing was first proposed by Marcu (1997), and early work was rule-based,
focused on discourse markers (Marcu, 2000a). The creation of the RST Discourse
TreeBank (Carlson et al. 2001, Carlson and Marcu 2001) enabled a wide variety
of machine learning algorithms, beginning with the shift-reduce parser of Marcu
(1999) that used decision trees to choose actions, and continuing with a wide variety
of machine learned parsing methods (Soricut and Marcu 2003, Sagae 2009, Hernault
et al. 2010, Feng and Hirst 2014, Surdeanu et al. 2015, Joty et al. 2015) and chunkers
(Sporleder and Lapata, 2005). Subba and Di Eugenio (2009) integrated sophisticated
semantic information into RST parsing. Ji and Eisenstein (2014) first applied neural
models to RST parsing neural models, leading to the modern set of neural RST
models (Li et al. 2014, Li et al. 2016, Braud et al. 2017, Yu et al. 2018, inter alia) as
well as neural segmenters (Wang et al. 2018). and neural PDTB parsing models (Ji
and Eisenstein 2015, Qin et al. 2016, Qin et al. 2017).
Barzilay and Lapata (2005) pioneered the idea of self-supervision for coher-
ence: training a coherence model to distinguish true orderings of sentences from
random permutations. Li et al. (2014) first applied this paradigm to neural sentence-
representation, and many neural self-supervised models followed (Li and Jurafsky
2017, Logeswaran et al. 2018, Lai and Tetreault 2018, Xu et al. 2019, Iter et al.
2020)
Another aspect of global coherence is the global topic structure of a text, the way
the topics shift over the course of the document. Barzilay and Lee (2004) introduced
an HMM model for capturing topics for coherence, and later work expanded this
intuition (Soricut and Marcu 2006, Elsner et al. 2007, Louis and Nenkova 2012, Li
and Jurafsky 2017).
The relationship between explicit and implicit discourse connectives has been
a fruitful one for research. Marcu and Echihabi (2002) first proposed to use sen-
tences with explicit relations to help provide training data for implicit relations, by
removing the explicit relations and trying to re-predict them as a way of improv-
ing performance on implicit connectives; this idea was refined by Sporleder and
Lascarides (2005), (Pitler et al., 2009), and Rutherford and Xue (2015). This rela-
 B IBLIOGRAPHICAL AND H ISTORICAL N OTES 21

tionship can also be used as a way to create discourse-aware representations. The

DisSent algorithm (Nie et al., 2019) creates the task of predicting explicit discourse
markers between two sentences. They show that representations learned to be good
at this task also function as powerful sentence representations for other discourse
tasks.
The idea of entity-based coherence seems to have arisen in multiple fields in the
mid-1970s, in functional linguistics (Chafe, 1976), in the psychology of discourse
processing (Kintsch and Van Dijk, 1978), and in the roughly contemporaneous work
of Grosz, Sidner, Joshi, and their colleagues. Grosz (1977) addressed the focus of
attention that conversational participants maintain as the discourse unfolds. She de-
fined two levels of focus; entities relevant to the entire discourse were said to be in
global focus, whereas entities that are locally in focus (i.e., most central to a partic-
ular utterance) were said to be in immediate focus. Sidner 1979; 1983 described a
method for tracking (immediate) discourse foci and their use in resolving pronouns
and demonstrative noun phrases. She made a distinction between the current dis-
course focus and potential foci, which are the predecessors to the backward- and
forward-looking centers of Centering theory, respectively. The name and further
roots of the centering approach lie in papers by Joshi and Kuhn (1979) and Joshi
and Weinstein (1981), who addressed the relationship between immediate focus and
the inferences required to integrate the current utterance into the discourse model.
Grosz et al. (1983) integrated this work with the prior work of Sidner and Grosz.
This led to a manuscript on centering which, while widely circulated since 1986,
remained unpublished until Grosz et al. (1995). A collection of centering papers ap-
pears in Walker et al. (1998). See Karamanis et al. (2004) and Poesio et al. (2004) for
a deeper exploration of centering and its parameterizations, and the History section
of Chapter 23 for more on the use of centering on coreference.
The grid model of entity-based coherence was first proposed by Barzilay and
Lapata (2005) drawing on earlier work by Lapata (2003) and Barzilay, and then
extended by them Barzilay and Lapata (2008) and others with additional features
(Elsner and Charniak 2008, 2011, Feng et al. 2014, Lin et al. 2011) a model that
projects entities into a global graph for the discourse (Guinaudeau and Strube 2013,
Mesgar and Strube 2016), and a convolutional model to capture longer-range entity
dependencies (Nguyen and Joty, 2017).
Theories of discourse coherence have also been used in algorithms for interpret-
ing discourse-level linguistic phenomena, including verb phrase ellipsis and gap-
ping (Asher 1993, Kehler 1993), and tense interpretation (Lascarides and Asher
1993, Kehler 1994, Kehler 2000). An extensive investigation into the relationship
between coherence relations and discourse connectives can be found in Knott and
Dale (1994).
Useful surveys of discourse processing and structure include Stede (2011) and
Webber et al. (2012).
Andy Kehler wrote the Discourse chapter for the 2000 first edition of this text-
book, which we used as the starting point for the second-edition chapter, and there
are some remnants of Andy’s lovely prose still in this third-edition coherence chap-
ter.
22 C HAPTER 24 • D ISCOURSE C OHERENCE

Exercises
24.1 Finish the Centering Theory processing of the last two utterances of (24.30),
and show how (24.29) would be processed. Does the algorithm indeed mark
(24.29) as less coherent?
24.2 Select an editorial column from your favorite newspaper, and determine the
discourse structure for a 10–20 sentence portion. What problems did you
encounter? Were you helped by superficial cues the speaker included (e.g.,
discourse connectives) in any places?
 Exercises 23

Althoff, T., C. Danescu-Niculescu-Mizil, and D. Jurafsky. Elvevåg, B., P. W. Foltz, D. R. Weinberger, and T. E. Gold-
2014. How to ask for a favor: A case study on the suc- berg. 2007. Quantifying incoherence in speech: an auto-
cess of altruistic requests. ICWSM 2014. mated methodology and novel application to schizophre-
Asher, N. 1993. Reference to Abstract Objects in Dis- nia. Schizophrenia research, 93(1-3):304–316.
course. Studies in Linguistics and Philosophy (SLAP) Feng, V. W. and G. Hirst. 2011. Classifying arguments by
50, Kluwer. scheme. ACL.
Asher, N. and A. Lascarides. 2003. Logics of Conversation. Feng, V. W. and G. Hirst. 2014. A linear-time bottom-up
Cambridge University Press. discourse parser with constraints and post-editing. ACL.
Baldridge, J., N. Asher, and J. Hunter. 2007. Annotation for Feng, V. W., Z. Lin, and G. Hirst. 2014. The impact of deep
and robust parsing of discourse structure on unrestricted hierarchical discourse structures in the evaluation of text
texts. Zeitschrift für Sprachwissenschaft, 26:213–239. coherence. COLING.
Bamman, D., B. O’Connor, and N. A. Smith. 2013. Learning Finlayson, M. A. 2016. Inferring Propp’s functions from se-
latent personas of film characters. ACL. mantically annotated text. The Journal of American Folk-
lore, 129(511):55–77.
Barzilay, R. and M. Lapata. 2005. Modeling local coherence:
An entity-based approach. ACL. Foltz, P. W., W. Kintsch, and T. K. Landauer. 1998. The
measurement of textual coherence with latent semantic
Barzilay, R. and M. Lapata. 2008. Modeling local coher-
analysis. Discourse processes, 25(2-3):285–307.
ence: An entity-based approach. Computational Linguis-
tics, 34(1):1–34. Grosz, B. J. 1977. The representation and use of focus in
a system for understanding dialogs. IJCAI-77. Morgan
Barzilay, R. and L. Lee. 2004. Catching the drift: Prob- Kaufmann.
abilistic content models, with applications to generation
and summarization. HLT-NAACL. Grosz, B. J., A. K. Joshi, and S. Weinstein. 1983. Provid-
ing a unified account of definite noun phrases in English.
Bedi, G., F. Carrillo, G. A. Cecchi, D. F. Slezak, M. Sig- ACL.
man, N. B. Mota, S. Ribeiro, D. C. Javitt, M. Copelli,
and C. M. Corcoran. 2015. Automated analysis of free Grosz, B. J., A. K. Joshi, and S. Weinstein. 1995. Center-
speech predicts psychosis onset in high-risk youths. npj ing: A framework for modeling the local coherence of
Schizophrenia, 1. discourse. Computational Linguistics, 21(2):203–225.
Biran, O. and K. McKeown. 2015. PDTB discourse parsing Guinaudeau, C. and M. Strube. 2013. Graph-based local co-
as a tagging task: The two taggers approach. SIGDIAL. herence modeling. ACL.
Habernal, I. and I. Gurevych. 2016. Which argument is more
Braud, C., M. Coavoux, and A. Søgaard. 2017. Cross-lingual
convincing? Analyzing and predicting convincingness of
RST discourse parsing. EACL.
Web arguments using bidirectional LSTM. ACL.
Brennan, S. E., M. W. Friedman, and C. Pollard. 1987. A
Habernal, I. and I. Gurevych. 2017. Argumentation mining
centering approach to pronouns. ACL.
in user-generated web discourse. Computational Linguis-
Carlson, L. and D. Marcu. 2001. Discourse tagging manual. tics, 43(1):125–179.
Technical Report ISI-TR-545, ISI.
Halliday, M. A. K. and R. Hasan. 1976. Cohesion in English.
Carlson, L., D. Marcu, and M. E. Okurowski. 2001. Building Longman. English Language Series, Title No. 9.
a discourse-tagged corpus in the framework of rhetorical Hearst, M. A. 1997. Texttiling: Segmenting text into multi-
structure theory. SIGDIAL. paragraph subtopic passages. Computational Linguistics,
Chafe, W. L. 1976. Givenness, contrastiveness, definiteness, 23:33–64.
subjects, topics, and point of view. In C. N. Li, ed., Sub- Hernault, H., H. Prendinger, D. A. duVerle, and M. Ishizuka.
ject and Topic, 25–55. Academic Press. 2010. Hilda: A discourse parser using support vector ma-
Chen, E., B. Snyder, and R. Barzilay. 2007. Incre- chine classification. Dialogue & Discourse, 1(3).
mental text structuring with online hierarchical ranking. Hidey, C., E. Musi, A. Hwang, S. Muresan, and K. McKe-
EMNLP/CoNLL. own. 2017. Analyzing the semantic types of claims and
Cialdini, R. B. 1984. Influence: The psychology of persua- premises in an online persuasive forum. 4th Workshop on
sion. Morrow. Argument Mining.
Ditman, T. and G. R. Kuperberg. 2010. Building coherence: Hobbs, J. R. 1979. Coherence and coreference. Cognitive
A framework for exploring the breakdown of links across Science, 3:67–90.
clause boundaries in schizophrenia. Journal of neurolin- Hovy, E. H. 1990. Parsimonious and profligate approaches to
guistics, 23(3):254–269. the question of discourse structure relations. Proceedings
Elsner, M., J. Austerweil, and E. Charniak. 2007. A unified of the 5th International Workshop on Natural Language
local and global model for discourse coherence. NAACL- Generation.
HLT. Iter, D., K. Guu, L. Lansing, and D. Jurafsky. 2020. Pretrain-
Elsner, M. and E. Charniak. 2008. Coreference-inspired co- ing with contrastive sentence objectives improves dis-
herence modeling. ACL. course performance of language models. ACL.
Elsner, M. and E. Charniak. 2011. Extending the entity grid Iter, D., J. Yoon, and D. Jurafsky. 2018. Automatic detec-
with entity-specific features. ACL. tion of incoherent speech for diagnosing schizophrenia.
Fifth Workshop on Computational Linguistics and Clini-
cal Psychology.
24 Chapter 24 • Discourse Coherence

Ji, Y. and J. Eisenstein. 2014. Representation learning for Louis, A. and A. Nenkova. 2012. A coherence model based
text-level discourse parsing. ACL. on syntactic patterns. EMNLP.
Ji, Y. and J. Eisenstein. 2015. One vector is not enough: Lukasik, M., B. Dadachev, K. Papineni, and G. Simões.
Entity-augmented distributed semantics for discourse re- 2020. Text segmentation by cross segment attention.
lations. TACL, 3:329–344. EMNLP.
Joshi, A. K. and S. Kuhn. 1979. Centered logic: The role Mann, W. C. and S. A. Thompson. 1987. Rhetorical structure
of entity centered sentence representation in natural lan- theory: A theory of text organization. Technical Report
guage inferencing. IJCAI-79. RS-87-190, Information Sciences Institute.
Joshi, A. K. and S. Weinstein. 1981. Control of inference: Marcu, D. 1997. The rhetorical parsing of natural language
Role of some aspects of discourse structure – centering. texts. ACL.
IJCAI-81. Marcu, D. 1999. A decision-based approach to rhetorical
Joty, S., G. Carenini, and R. T. Ng. 2015. CODRA: A novel parsing. ACL.
discriminative framework for rhetorical analysis. Compu- Marcu, D. 2000a. The rhetorical parsing of unrestricted
tational Linguistics, 41(3):385–435. texts: A surface-based approach. Computational Linguis-
tics, 26(3):395–448.
Karamanis, N., M. Poesio, C. Mellish, and J. Oberlander.
2004. Evaluating centering-based metrics of coherence Marcu, D., ed. 2000b. The Theory and Practice of Discourse
for text structuring using a reliably annotated corpus. Parsing and Summarization. MIT Press.
ACL. Marcu, D. and A. Echihabi. 2002. An unsupervised approach
Kehler, A. 1993. The effect of establishing coherence in el- to recognizing discourse relations. ACL.
lipsis and anaphora resolution. ACL. Mesgar, M. and M. Strube. 2016. Lexical coherence graph
modeling using word embeddings. ACL.
Kehler, A. 1994. Temporal relations: Reference or discourse
coherence? ACL. Miltsakaki, E., R. Prasad, A. K. Joshi, and B. L. Webber.
2004. The Penn Discourse Treebank. LREC.
Kehler, A. 2000. Coherence, Reference, and the Theory of
Grammar. CSLI Publications. Morey, M., P. Muller, and N. Asher. 2017. How much
progress have we made on RST discourse parsing? a
Kintsch, W. and T. A. Van Dijk. 1978. Toward a model of replication study of recent results on the rst-dt. EMNLP.
text comprehension and production. Psychological re-
Morris, J. and G. Hirst. 1991. Lexical cohesion computed by
view, 85(5):363–394.
thesaural relations as an indicator of the structure of text.
Knott, A. and R. Dale. 1994. Using linguistic phenomena Computational Linguistics, 17(1):21–48.
to motivate a set of coherence relations. Discourse Pro- Muller, P., C. Braud, and M. Morey. 2019. ToNy: Contextual
cesses, 18(1):35–62. embeddings for accurate multilingual discourse segmen-
Lai, A. and J. Tetreault. 2018. Discourse coherence in the tation of full documents. Workshop on Discourse Relation
wild: A dataset, evaluation and methods. SIGDIAL. Parsing and Treebanking.
Lakoff, G. 1972. Structural complexity in fairy tales. In The Musi, E., M. Stede, L. Kriese, S. Muresan, and A. Rocci.
Study of Man, 128–50. School of Social Sciences, Uni- 2018. A multi-layer annotated corpus of argumenta-
versity of California, Irvine, CA. tive text: From argument schemes to discourse relations.
Lapata, M. 2003. Probabilistic text structuring: Experiments LREC.
with sentence ordering. ACL. Myers, G. 1992. “In this paper we report...”: Speech acts and
Lascarides, A. and N. Asher. 1993. Temporal interpretation, scientific facts. Journal of Pragmatics, 17(4):295–313.
discourse relations, and common sense entailment. Lin- Nguyen, D. T. and S. Joty. 2017. A neural local coherence
guistics and Philosophy, 16(5):437–493. model. ACL.
Li, J. and D. Jurafsky. 2017. Neural net models of open- Nie, A., E. Bennett, and N. Goodman. 2019. DisSent: Learn-
domain discourse coherence. EMNLP. ing sentence representations from explicit discourse rela-
tions. ACL.
Li, J., R. Li, and E. H. Hovy. 2014. Recursive deep models
for discourse parsing. EMNLP. Park, J. and C. Cardie. 2014. Identifying appropriate support
for propositions in online user comments. First workshop
Li, Q., T. Li, and B. Chang. 2016. Discourse parsing with on argumentation mining.
attention-based hierarchical neural networks. EMNLP.
Peldszus, A. and M. Stede. 2013. From argument diagrams
Lin, Z., M.-Y. Kan, and H. T. Ng. 2009. Recognizing im- to argumentation mining in texts: A survey. International
plicit discourse relations in the Penn Discourse Treebank. Journal of Cognitive Informatics and Natural Intelligence
EMNLP. (IJCINI), 7(1):1–31.
Lin, Z., H. T. Ng, and M.-Y. Kan. 2011. Automatically eval- Peldszus, A. and M. Stede. 2016. An annotated corpus of
uating text coherence using discourse relations. ACL. argumentative microtexts. 1st European Conference on
Lin, Z., H. T. Ng, and M.-Y. Kan. 2014. A pdtb-styled end- Argumentation.
to-end discourse parser. Natural Language Engineering, Pitler, E., A. Louis, and A. Nenkova. 2009. Automatic sense
20(2):151–184. prediction for implicit discourse relations in text. ACL
Logeswaran, L., H. Lee, and D. Radev. 2018. Sentence IJCNLP.
ordering and coherence modeling using recurrent neural Pitler, E. and A. Nenkova. 2009. Using syntax to disam-
networks. AAAI. biguate explicit discourse connectives in text. ACL IJC-
NLP.
 Exercises 25

Poesio, M., R. Stevenson, B. Di Eugenio, and J. Hitzeman. Stab, C. and I. Gurevych. 2014b. Identifying argumentative
2004. Centering: A parametric theory and its instantia- discourse structures in persuasive essays. EMNLP.
tions. Computational Linguistics, 30(3):309–363. Stab, C. and I. Gurevych. 2017. Parsing argumentation struc-
Polanyi, L. 1988. A formal model of the structure of dis- tures in persuasive essays. Computational Linguistics,
course. Journal of Pragmatics, 12. 43(3):619–659.
Polanyi, L., C. Culy, M. van den Berg, G. L. Thione, and Stede, M. 2011. Discourse processing. Morgan & Claypool.
D. Ahn. 2004. A rule based approach to discourse pars- Stede, M. and J. Schneider. 2018. Argumentation Mining.
ing. Proceedings of SIGDIAL. Morgan & Claypool.
Prasad, R., N. Dinesh, A. Lee, E. Miltsakaki, L. Robaldo, Subba, R. and B. Di Eugenio. 2009. An effective discourse
A. K. Joshi, and B. L. Webber. 2008. The Penn Discourse parser that uses rich linguistic information. NAACL HLT.
TreeBank 2.0. LREC.
Surdeanu, M., T. Hicks, and M. A. Valenzuela-Escarcega.
Prasad, R., B. L. Webber, and A. Joshi. 2014. Reflections on 2015. Two practical rhetorical structure theory parsers.
the Penn Discourse Treebank, comparable corpora, and NAACL HLT.
complementary annotation. Computational Linguistics,
40(4):921–950. Tan, C., V. Niculae, C. Danescu-Niculescu-Mizil, and
L. Lee. 2016. Winning arguments: Interaction dynam-
Propp, V. 1968. Morphology of the Folktale, 2nd edition. ics and persuasion strategies in good-faith online discus-
University of Texas Press. Original Russian 1928. Trans- sions. WWW-16.
lated by Laurence Scott.
Teufel, S., J. Carletta, and M. Moens. 1999. An annotation
Qin, L., Z. Zhang, and H. Zhao. 2016. A stacking gated scheme for discourse-level argumentation in research ar-
neural architecture for implicit discourse relation classifi- ticles. EACL.
cation. EMNLP.
Teufel, S., A. Siddharthan, and C. Batchelor. 2009. To-
Qin, L., Z. Zhang, H. Zhao, Z. Hu, and E. Xing. 2017. Ad-
wards domain-independent argumentative zoning: Ev-
versarial connective-exploiting networks for implicit dis-
idence from chemistry and computational linguistics.
course relation classification. ACL.
EMNLP.
Reed, C., R. Mochales Palau, G. Rowe, and M.-F. Moens.
Walker, M. A., A. K. Joshi, and E. Prince, eds. 1998. Cen-
2008. Language resources for studying argument. LREC.
tering in Discourse. Oxford University Press.
Rosenthal, S. and K. McKeown. 2017. Detecting influencers
Wang, Y., S. Li, and J. Yang. 2018. Toward fast and accurate
in multiple online genres. ACM Transactions on Internet
neural discourse segmentation. EMNLP.
Technology (TOIT), 17(2).
Webber, B. L., M. Egg, and V. Kordoni. 2012. Discourse
Rutherford, A. and N. Xue. 2015. Improving the inference
structure and language technology. Natural Language
of implicit discourse relations via classifying explicit dis-
Engineering, 18(4):437–490.
course connectives. NAACL HLT.
Wolf, F. and E. Gibson. 2005. Representing discourse coher-
Sagae, K. 2009. Analysis of discourse structure with syn-
ence: A corpus-based analysis. Computational Linguis-
tactic dependencies and data-driven shift-reduce parsing.
tics, 31(2):249–287.
IWPT-09.
Xu, P., H. Saghir, J. S. Kang, T. Long, A. J. Bose, Y. Cao,
Scha, R. and L. Polanyi. 1988. An augmented context free
and J. C. K. Cheung. 2019. A cross-domain transferable
grammar for discourse. COLING.
neural coherence model. ACL.
Sidner, C. L. 1979. Towards a computational theory of defi-
Xue, N., H. T. Ng, S. Pradhan, A. Rutherford, B. L. Web-
nite anaphora comprehension in English discourse. Tech-
ber, C. Wang, and H. Wang. 2016. CoNLL 2016 shared
nical Report 537, MIT Artificial Intelligence Laboratory,
task on multilingual shallow discourse parsing. CoNLL-
Cambridge, MA.
16 shared task.
Sidner, C. L. 1983. Focusing in the comprehension of defi-
Yang, D., J. Chen, Z. Yang, D. Jurafsky, and E. H. Hovy.
nite anaphora. In M. Brady and R. C. Berwick, eds, Com-
2019. Let’s make your request more persuasive: Model-
putational Models of Discourse, 267–330. MIT Press.
ing persuasive strategies via semi-supervised neural nets
Somasundaran, S., J. Burstein, and M. Chodorow. 2014. on crowdfunding platforms. NAACL HLT.
Lexical chaining for measuring discourse coherence qual-
Yu, N., M. Zhang, and G. Fu. 2018. Transition-based neural
ity in test-taker essays. COLING.
RST parsing with implicit syntax features. COLING.
Soricut, R. and D. Marcu. 2003. Sentence level discourse
Yu, Y., Y. Zhu, Y. Liu, Y. Liu, S. Peng, M. Gong, and
parsing using syntactic and lexical information. HLT-
A. Zeldes. 2019. GumDrop at the DISRPT2019 shared
NAACL.
task: A model stacking approach to discourse unit seg-
Soricut, R. and D. Marcu. 2006. Discourse generation using mentation and connective detection. Workshop on Dis-
utility-trained coherence models. COLING/ACL. course Relation Parsing and Treebanking 2019.
Sporleder, C. and A. Lascarides. 2005. Exploiting linguistic Zhou, Y. and N. Xue. 2015. The Chinese Discourse Tree-
cues to classify rhetorical relations. RANLP-05. Bank: a Chinese corpus annotated with discourse rela-
Sporleder, C. and M. Lapata. 2005. Discourse chunking and tions. Language Resources and Evaluation, 49(2):397–
its application to sentence compression. EMNLP. 431.
Stab, C. and I. Gurevych. 2014a. Annotating argument com-
ponents and relations in persuasive essays. COLING.