Introduction to the Special Issue on Language - 麻省理工学院人工智能研究专业

Introduction to the Special Issue on Language
in Social Media: Exploiting Discourse and
Other Contextual Information

Farah Benamara
Paul Sabatier University
IRIT-Universit´e de Toulouse
benamara@irit.fr

Diana Inkpen
渥太华大学
School of Electrical Engineering and
计算机科学
Diana.Inkpen@uottawa.ca

Maite Taboada
Simon Fraser University
语言学系
mtaboada@sfu.ca

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Social media content is changing the way people interact with each other and share information,
personal messages, and opinions about situations, 物体, and past experiences. Most social
media texts are short online conversational posts or comments that do not contain enough
information for natural language processing (自然语言处理) 工具, as they are often accompanied by
non-linguistic contextual information, including meta-data (例如, the user’s proﬁle, the social
network of the user, and their interactions with other users). Exploiting such different types of
context and their interactions makes the automatic processing of social media texts a challenging
research task. 的确, simply applying traditional text mining tools is clearly sub-optimal, 作为,
typically, these tools take into account neither the interactive dimension nor the particular nature
of this data, which shares properties with both spoken and written language. This special issue
contributes to a deeper understanding of the role of these interactions to process social media data
from a new perspective in discourse interpretation. This introduction ﬁrst provides the necessary
background to understand what context is from both the linguistic and computational linguistic
perspectives, then presents the most recent context-based approaches to NLP for social media.
We conclude with an overview of the papers accepted in this special issue, highlighting what we
believe are the future directions in processing social media texts.

提交材料已收到: 10 九月 2018; 接受出版: 10 九月 2018.

土井:10.1162/coli a 00333

计算语言学

体积 44, 数字 4

1. 介绍

Social media content has, for many people and organizations, changed the way we
interact and share information. This content (ranging from blogs, fora, reviews, 和
various social networking sites) has speciﬁc characteristics that are often referred to as
the ﬁve V’s: 体积, variety, velocity, veracity, and value.

Social media texts are more difﬁcult to process than traditional texts because of
the nature of the social conversations—posted in real-time. The texts are unstructured
and are presented in many formats and written by different people in many languages
and styles. Typographic errors are common, and chat and in-group slang have become
increasingly prevalent on social networking sites like Facebook and Twitter.

此外, most social media texts are short online conversational posts or com-
ments that do not contain enough information for natural language processing (自然语言处理)
工具. They are often accompanied by non-linguistic contextual information, 包括
meta-data such as the social network of each user and their interactions with other
用户. Because the conversation ﬂow is not necessarily sequential, as users can write
(and hence reply) at different times, these conversations are often called asynchronous.
Exploiting this kind of contextual information and meta-data could compensate for
the lack of information from the texts themselves. Such rich contextual information
makes the automatic processing of social media content a challenging research task.
的确, simply applying traditional text mining tools is clearly sub-optimal, as it takes
into account neither the interactive dimension nor the particular nature of these data,
which share properties with both spoken and written language. Most research on
NLP for social media focuses primarily on content-based processing of the linguistic
信息, using lexical semantics (例如, discovering new word senses or multi-word
expressions) or semantic analysis (opinion extraction, irony detection, event and topic
detection, geo-location detection) (Aiello et al. 2013; 戈什等人. 2015; Inkpen et al. 2015;
Londhe, Srihari, and Gopalakrishnan 2016).1 Other research explores the interactions
between content and extra-linguistic or extra-textual features, showing that combining
linguistic data with network and/or user context improves performance over a base-
line that uses only textual information. 例如, user proﬁles like age, 性别,
and location can be used to enhance subjectivity detection (including sentiment and
情感) (Volkova, Coppersmith, and Van Durme 2014; Volkova and Bachrach 2016),
vote predictions (佩辛和吴 2014), or language identiﬁcation (Saloot et al. 2016).
还, information from the conversational thread structure (例如, links between previous
posts) or valuable external sources can serve as contextual constraints to better capture
the sentiment or the ﬁgurative reading of an utterance (Mukherjee and Bhattacharyya
2012; Karoui et al. 2015; 华莱士, Choe, and Charniak 2015)2. 最后, the social network,
like social relationships, can enable grouping users according to speciﬁc communities
regarding the topics or the sentiments they share (Deitrick and Hu 2013; West et al.
2014).

Besides social media processing, the interaction of contextual information derived
from sentences, 话语, and other forms of linguistic and extra-linguistic information
have shown their effectiveness in language technology in general (Taboada and Mann
2006; Webber, Egg, 给科登 2012). This shows that computational linguistics is

1 See Farzindar and Inkpen (2017) for an overview of the main NLP approaches for social media.
2 See Benamara, Taboada, and Mathieu (2017) for a recent overview of context-based approaches to

evaluative language processing.

664

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Benamara, Inkpen, and Taboada

Special Issue on Language in Social Media

currently experiencing a discourse turn, a growing awareness of how multiple sources of
信息, and especially information from context and discourse, can have a positive
impact on a range of computational applications. This turn is particularly notable in the
research community, where several workshops have been recently organized in major
NLP international conferences to account for the role discourse and context can have
in various NLP tasks (例如, the DiscoMT series on discourse in machine translation,
CompPrag on computational pragmatics, SocialNLP on NLP for Social Media, 和
many of the papers at *SEM or SemEval workshops).

This special issue invited contributions that implement such approaches, 但不是

restricted exclusively to applications in evaluative language and sentiment analysis.

Before giving an overview of the papers accepted in this special issue (部分 4), 我们
provide some background on what context is from both the linguistic and computational
linguistic perspectives (部分 2). We then focus on current context-based approaches
to NLP for social media (部分 3). We end this introduction by highlighting what we
believe are the future directions in processing social media texts.

2. Context in Computational Linguistics

Context is a pervasive term in linguistics and no single coherent deﬁnition of context is
可用的 (巴赫 1997; Recanati 2008; Jaszczolt 2012; Korta and Perry 2015). An intuitive
view is to consider the distinctions between the linguistic information formed by mor-
phological, 句法的, or textual material surrounding a word, and any other contextual
information surrounding the utterance. Bunt and Black (2000) discuss the following
non-exhaustive aspects of contextual information:

•

Discourse context: What has been said before in the conversation (IE。,
objects that have been introduced in the preceding discourse).

Attitudinal or epistemic context: This encompasses the speaker’s
知识, the hearer’s knowledge, and the common ground (IE。, what is
known to both the speaker and the hearer about the domain of the
话语).

Spatio-temporal properties of the situation in which the utterance occurs,
like the relative time and place of speaking.

Physical and perceptual context: Objects that are known to be present or
visible in the speaker’s and the hearer’s environment; actions and events
perceivable in that environment. The textual form of an utterance (例如
punctuation and layout) is also important.

Social context: The social relationship of the people involved in
沟通. A sentence like President, leave me alone is only shocking
because we know one does not usually address a president this way.

The question is then: How can these different sources of information interact to
make computers understand natural language texts? There are two possible options to
回答这个问题: Consider each source of information as a separate stage, involving
a linear process starting with words and ending with extra-linguistic context; or incor-
porate contextual information at an earlier stage. The ﬁrst option being computationally
inefﬁcient due in particular to the ambiguity of words and sentences when processed in

665

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

isolation, this special issue adopted the second option, as explained in the subsequent
sections.

2.1 Words and Sentences

One way to compute the meaning of a text is to exploit the meanings of words and how
these words are syntactically composed to form a text. This inspired the development
of truth-conditional semantics or model-theoretic semantics in which the meaning of
a sentence is determined relative to a model, which can be taken to be an abstract
description of the world (Montague 1974; Tarski 1983). Lexical meaning and syntax
provide linguistic knowledge and play a crucial role in studying the behavior of semantic
phenomena bound at the sentence level (Bos 2011).

We illustrate the composition process by the effect intensiﬁers and downtoners
have on the evaluative expressions they modify. Many devices intensify by changing
the intensity of an evaluative word, whether by bringing it up or down. 例如,
adjectives may intensify or downtone the noun they accompany (例如, A deﬁnite success),
as adverbs do with adjectives (例如, A very dangerous trip) or verbs (例如, He behaved
badly). Examples (1) 和 (2), extracted from the CASOAR corpus (Benamara et al. 2016),
show a more complex case where the overall sentiment orientation is determined in a
bottom–up fashion.

(1) The actors are not good enough.

(2) This restaurant proposes good quality Greek cuisine in a warm atmosphere.

Moving from a subjectivity lexicon that encodes the meaning of sentiment-relevant
字 (like the adjectives good and warm), composition follows the syntactic tree up
to the main clause by combining pairs of sister nodes by means of a set of sentiment
composition rules. In Example (1), sentiment calculation has ﬁrst to deal with the
composition good enough that softens the positivity of the evaluation, which in turn
has to be composed with the negation (不是) that makes the overall opinion negative.
In Example (2), the sentence’s syntactic structure indicates that the atmosphere and the
cuisine have both a positive evaluation. For more discussions on sentiment composition,
the reader can refer to the Stanford Sentiment Treebank (索切尔等人. 2013).

The composition process assumes that the interpretation of a given word within a
sentence is ﬁxed or disambiguated before being combined, which makes it restrictive in
that it “precludes nonlinguistic information to go into the computation of meaning”
(Bunt 2001).3 的确, the meaning of a sentence is closely tied to the pragmatics of
how language is used, and thus to the meaning of the words themselves, which can
be assigned different possible readings in different situations (普斯特约夫斯基 1995; Lenci
2006). Consider the problem of lexical ambiguity. 例如, A sad movie expresses
a sentiment or feeling of grief, whereas Sad weather expresses an undesirable judgment
that can be paraphrased as The weather is bad. There are also ambiguities that are not
caused by lexical choice, but by the context in which the words occur. 例如, 这
adjective long may denote a negative sentiment in restaurant reviews (比照. 例子 (3))
but a positive sentiment in phone reviews (比照. 例子 (4)). The same adjective can also
be purely factual, as in Example (5).

3 See Janssen (2001) and Zimmermann (2013) for a discussion of the principle of compositionality.

666

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Benamara, Inkpen, and Taboada

Special Issue on Language in Social Media

(3) There is a long wait between courses.

(4) The smart phone has a long battery life.

(5)

It has rained for a long time.

The assumption that word meaning is a function of the contexts in which it occurs
within the sentence is at the center of the distributional semantics hypothesis (特尼
and Pantel 2010). Distributional models represent words by vectors build by extracting
co-occurrences statistics from large corpora, then use linear algebra as a computational
tool to project lexical vectors to phrase vectors. Vectorial representations are extremely
effective for computing semantic similarity between words, and more generally inves-
tigating the interplay between meaning and contexts (Lenci 2018).

The meaning of a sentence can also rely on other types of information, 例如
prosodic information in the case of spoken utterances; or punctuation, layout, and emo-
jis in the case of textual utterances. The latter is of particular importance when analyzing
社交媒体, as shown in Examples (6) 和 (7), where capitalization and character
repetition, 分别, emphasize the positive opinion towards the movie.

(6) This movie was AMAZING.

(7) This movie was amaaazzzzzing.

2.2 Beyond Sentences: Discourse Structure

Words and sentences do not occur in isolation, but both are always part of a coherent
and cohesive structure in which the discourse units are related to each other. Coherence
refers to the logical structure of the discourse, where every part of a text has a func-
的, a role to play, with respect to other parts in the text (Taboada and Mann 2006).
Coherence has to do with semantic or pragmatic relations among units to produce the
overall meaning of a discourse (Hobbs 1979; 曼和汤普森 1988; Grosz, Joshi,
and Weinstein 1995). The impression of coherence in text (that it is organized, 那它
hangs together) is also aided by cohesion, the linking of entities in discourse (Halliday
and Hasan 1976). Linking across entities happens through grammatical and lexical
connections such as anaphoric expressions and lexical relations (synonymy, meronymy,
hyponymy) appearing across sentences.

Theories of discourse interpretation typically account for meaning beyond the sen-
张力. 大致, two main approaches have been developed: dynamic semantics (Heim
1982; Kamp and Reyle 1993) and theories of discourse structure (Hobbs 1979; Grosz and
Sidner 1986; 曼和汤普森 1988; 亚瑟和拉斯卡里德斯 2003; Prasad, Webber,
和乔希 2014).

The ﬁrst approach extends model-theoretic semantics to account for the semantic
contribution that a sentence makes to a discourse in terms of a relation between an
input context prior to the sentence and an output one. Discourse context is therefore a
dynamic concept:

When a sentence S is interpreted within the discourse context K, the result of its
interpretation will be integrated into K. The updated context K(西德:48), which reﬂects the
contribution made by S as well as those made by the sentences preceding it, will then
be the discourse context for the next sentence. (Kamp and Reyle, 2010, 页 3)

667

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

In the second approach, theories of discourse structure derive meaning from the
rhetorical relations that link discourse units4 such as ELABORATION, EXPLANATION,
NARRATION, 等等. Discourse relations are important factors that make a dis-
course coherent. Coherence can be accounted for by positing relations between clauses,
句子, or speech acts (see the next section) that organize the writer’s intentions (和
explanations, elaborations, and contrasts, 例如) or explain speakers’ turns (例如,
answer to a question, acknowledgment of a proposal or an assertion, correction of an
assertion). A number of theories of relational coherence have been proposed, for written
text and dialogue, which make different assumptions about the kinds of relations (因此
yielding different taxonomies of discourse relations), or the resulting structure (a chain,
一棵树, or diversely constrained types of graphs that inﬂuence the interpretation process)
(see Asher and Lascarides 2003; Taboada and Mann 2006 for an overview).

Even if dynamic semantics and theories of discourse structure differ in their aims
and methods, they stress the need to model the cumulative nature of discourse interpre-
站, 即, the interpretation of a current discourse unit depends on the content of
the part of the discourse which precedes it. To illustrate the importance of discourse
structure and how constraints on coherent discourse determine lexical sense disam-
歧义, consider the following two short texts, taken respectively, from TripAdvisor
and Twitter.5

(8)

(9)

[This restaurant is not remarkable.]π1 [The dishes were correct]π2 [but side
dishes very average.]π3 [The wine was warm.]π4
I want to be an ecologist, but energy-saving light bulbs take more time to burst
these idiots moths.

例子 (8) shows that sentiment is a semantic scope phenomenon governed by
discourse structure (Polanyi and van den Berg 2011). In the ﬁrst sentence, the author in-
troduces the main topic of the discourse (This restaurant), expressing a negative opinion
towards it. This opinion is further elaborated in the discourse units π2 to π4, 哪里的
author comments on two aspects of the restaurant: the cuisine and wine. To infer the
ELABORATION relation that holds between π1 and (π2-π3) and between π1 and π4, 我们
need detailed lexical knowledge and probably domain knowledge as well (the fact that
cuisine and wine are part of a restaurant is implicit). π4 expresses a negative opinion
lexicalized by the adjective warm. The interpretation of the degree of subjectivity of this
adjective is a matter of context. The fact that π4 elaborates on π1 helps disambiguating
the sense of this adjective: one cannot elaborate positively on a topic that has been
previously assigned a negative opinion.

最后, 例子 (9) shows the importance of discursive contextual phenomena
at the sentence level: It is the contrast rhetorical relation triggered by the discourse
connective but that allows us to infer that the writer implicitly says that they are against
saving energy, even though they state the contrary in the ﬁrst sentence.

2.3 Beyond What Is Said

Full comprehension of a text also requires understanding more than what is linguis-
tically encoded, 那是, understanding beyond what is said. Approaches like speech act

4 Some theories do also provide a model-theoretic semantics for a discourse. 例如, the Structured
Discourse Representation Theory (亚瑟和拉斯卡里德斯 2003) incorporates, but also extends, dynamic
语义学.

5 This is a French tweet translated to English.

668

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Benamara, Inkpen, and Taboada

Special Issue on Language in Social Media

理论 (Austin 1962; Searle 1969) and convversational implicature (Grice 1975) make a
clear distinction between what is said by an utterance and what is implicated or performed in
a particular linguistic and social context or by saying something (Korta and Perry 2015).
Austin (1962) provided a framework for connecting the literal meaning of an utter-
ance with its intended meaning. He argued that every utterance has three layers of
意义: (我) a locutionary act that corresponds to the act of saying something with
字, (二) an illocutionary act, which conveys the speaker’s intended meaning on
the basis of the existence of a social practice, conventions, or “constitutive” rules in
doing things with words (like ordering, offering, warning, promising, ETC。), 和 (三、)
a perlocutionary act that reﬂects the listener’s perception of the speaker’s intended
意义, 那是, the effect a locutionary act has on the feelings, thoughts, or actions
of either the speaker or the listener (like inspiring, amusing, persuading, ETC。). 为了
例子, the illocutionary act of the utterance I am free next week, shall we meet on Friday?
is a suggestion, while its intended perlocutionary effect might be to invite the hearer to
ﬁx a particular day to meet. The illocutionary act is a central aspect of the speech-act
理论, developed later by Searle (1969).

Speech acts are the semantic/pragmatic counterpart of sentence types. The sen-
tences types afﬁrmative, interrogative, and exclamative correlate with the speech acts
of assertion, 问题, expression, and order. Speech acts are relevant in social media
and there is an emerging new interest in the computational community for speech acts
(看, 例如, the article by Joty and Mohiuddin in this special issue).

Whereas speech acts have traditionally been understood as unary properties of
expressions that convey propositions, Searle lists categories of speech acts like “an-
swers” that are clearly relational (an answer is an answer to a particular question).
Once one observes that some speech acts are relational, it is relatively straightforward
to see discourse relations like EXPLANATION and ELABORATION also as types of speech
行为. Unlike traditional speech acts, 然而, instances of discourse relations easily
embed under various operators (like modality), whereas it remains controversial as to
whether speech acts like assertion or requests embed.6

Speech acts are crucial in the analysis of some pragmatic phenomena such as
preferences and intentions that concern the future states of affairs or plans that one
wants to achieve. 例如, in the conversational thread for Example (10) (taken
from Twitter), the question–answer pair that links User’s A question to User’s B answer
helps to better capture User B’s intention towards eating organic food and not food
with additives or pesticides.

(10)

(User A) Do you prefer eating cakes with additives or fruits with pesticides?

(User B) Neither. I prefer to eat organic.

另一方面, Grice (1975) argued that communication between people was
also characterized by the process of intention recognition. He made a clear distinction
between what is said by an utterance (IE。, meaning out of context) and what is implied
or meant by an utterance (IE。, meaning in context). In his theory of conversational
implicature, Grice proposes that to capture the speaker’s meaning, the hearer needs
to rely on the meaning of the sentence uttered, contextual assumptions, and the Coop-
erative Principle, which speakers are expected to observe. The Cooperative Principle
states that speakers make contributions to the conversation that are cooperative, 并且是

6 See the work of Krifka (2002) for arguments that even standard speech acts embed to some degree.

669

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

expressed in four maxims that the communication participants are supposed to follow.
The maxims ask the speaker to say what they believe to be the truth (Quality), 成为
as informative as possible (Quantity), to say the utterance at the appropriate point
in the interaction (Relevance), and in the appropriate manner (Manner). The maxims
是, 从某种意义上说, ideals, and Grice provided examples of violations of these maxims for
various reasons. The violation of a maxim may result in the speaker conveying, 在
addition to the literal meaning of the utterance, a meaning that does not contribute to the
truth-conditional content of the utterance, which leads to conversational implicature.
Implicatures are thus inferences that can defeat literal and compositional meaning.
例子 (11) is a typical example of relevance violation: B conveys to A that they will
not be accepting A’s invitation for dinner, although they have not said so directly.

(11) A. Let’s have dinner tonight.

乙.

I have to ﬁnish my homework.

Grice makes the important assumptions that participants in a discourse are rational
agents and that they are governed by cooperative principles. 然而, 在某些情况下
involving non-literal readings or negotiation, agents do not always have rational com-
municative behavior.

Some contemporary researchers reject the distinction between literal and utterance
意义, arguing that what is said is always dependent on the context (Recanati 2004;
Korta and Perry 2015). The debate shared by literalists and contextualists on the frontier
between semantics and pragmatics is not the most important point here.7 What matters
for the purpose of this special issue is how to make computers capture the meaning of
a text when immersed in the context in which it is uttered.

In user-generated content such as product reviews, inference is often needed to cap-
ture implicit evaluation like the ones expressed in the movie reviews of Examples (12)
和 (13), taken from the CASOAR corpus. Even if there are no explicit subjective
字, everyone would expect a movie to be good when reading Example (12), and bad
after reading Example (13).

(12) This is a deﬁnite choice to be in my DVD collection.

(13)

I really want my money back.

Irony is another important pragmatic phenomenon that poses new challenges when
processing short texts. Irony can be deﬁned as an incongruity between the literal mean-
ing of an utterance and its intended meaning (Grice 1975; Sperber and Wilson 1981;
Utsumi 1996; Attardo 2000). In social media, such as Twitter, and mainly in English,
users apply speciﬁc hashtags (#irony, #sarcasm, #sarcastic) to help readers understand
that a message is ironic. This is shown in the tweet of Example (14), which clearly
expresses a negative opinion towards Nabilla, although there are two positive opinion
字 (classy and beautiful).

(14) #Nabilla a very classy and beautiful girl, not made over at all #irony

3. Context in Social Media

The interaction between the different sources of contextual information discussed so
far highlights a set of challenging issues in the semantics–pragmatics interface, not all

7 See McNally (2013) for an interesting discussion on that topic.

670

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Benamara, Inkpen, and Taboada

Special Issue on Language in Social Media

of which are solved and clear at the theoretical level. 此外, the NLP challenge
is how to take these insights about different types of context and make good use of
them in applications—in particular in applications that involve social media content. 在
this section, we review recent developments in processing social media language that
incorporate the role of context.

3.1 On the Role of Discourse Phenomena

Discourse structure in social media conversations (like Twitter multilogues, IE。, conver-
sations between users via the reply-to relation) differs in a number of aspects from that
of “classical” dialogues (IE。, human–human and human–machine spoken dialogues).
的确, some speciﬁc features such as Twitter @-mentions and hashtags may pose
some problems regarding the choice of the appropriate unit of analysis (句子,
discourse unit, ETC。) and level of the discourse structure these units should be embedded
(Sidarenka, Bisping, 和施泰德 2015). 此外, social media corpora are composed of
follow-up conversations, where topics are dynamic over conversation threads—that is,
not necessarily known in advance. 例如, posts on a forum or tweets are often
responses to earlier posts, and the lack of context makes it difﬁcult for machines to ﬁgure
出去, 例如, whether the post is in agreement or disagreement.

Discourse contextual phenomena in social media can be leveraged in several ways,

as discussed in the next sections.

3.1.1 Discourse Structure and Coherence Modeling. Although the analysis of discourse
structure for traditionally written text is now well established (林, 能, 和吴 2009;
埃尔诺等人. 2010; 冯和赫斯特 2014; Joty, Carenini, 和吴 2015), there is little
work on applying discourse theories to social media texts. Among them, Sidarenka,
Bisping, 和施泰德 (2015) study how coherence is achieved in social media conversa-
tions relying on Rhetorical Structure Theory (曼和汤普森 1988). They pro-
pose a scheme to manually annotate tweets according to Rhetorical Structure Theory
principles and found that up to 40% of German tweets are part of conversations, 和
that answer-relations create discourse trees. The analysis of Twitter-speciﬁc phenomena
reveals that URLs carry communicative content (such as Inform, Opening, Suggestion).
相似地, 话语关系 (such as Elaboration, Exempliﬁcation, 评估) 是
rarely explicit (仅有的 20% of the cases). They also observe that causal connectives are
frequent in Twitter: 1.7% of the tweets and 2.6% of the replies.

Following the entity grid coherent model (Barzilay and Lapata 2008), Joty, 阮,
and Mohiuddin (2018) also focus on the problem of coherence in asynchronous con-
诗篇. The authors propose a neural model to predict the underlying thread struc-
ture of fora conversations. The model has also been applied in reconstructing thread
结构.

最后, Perret et al. (2016) propose the ﬁrst discourse parser for multi-party chat
dialogues using integer linear programming. They investigate both treelike and non-
treelike full discourse structures, achieving an F-measure of 0.531. These results are
encouraging and open interesting future directions in discourse parsing of social media
conversations.

3.1.2 Argumentation Mining. Speciﬁc argumentative discourse relations are of particular
importance in social media. 的确, a user often not only reports facts, expresses opin-
离子, and engages with the reader, but also presents arguments in a certain order and
with certain organization. These arguments are structured in terms of a set of premises

671

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

that provide the evidence or the reasons for or against a conclusion. Tracking arguments
in text, also know as argumentation mining, consists of ﬁrst identifying arguments
(IE。, separating arguments from non-arguments), then their argumentative structure
(including the premises, 结论, and the connections between them such as the
argument and counter-argument relationships). Argumentation mining in Twitter has
been studied by Bosc, 敞篷车, 和维拉塔 (2016), who propose a binary classiﬁer to
argument identiﬁcation. Dusmanu, 敞篷车, 和维拉塔 (2017) go further by separating
personal opinions from actual facts, and detecting the source of such facts to allow for
provenance veriﬁcation.

Argumentation mining in social media has given rise to new tasks such as detecting
agreements and disagreement in conversations (艾伦, Carenini, 和吴 2014), counter-
factual recognition (Son et al. 2017), identiﬁcation of controversial topics (Addawood
and Bashir 2016), stance/rumor detection (Zubiaga et al. 2016), and fact-checking (Baly
等人. 2018). Argumentation and stancetaking are further discussed later in this special
问题 (比照. Cocarascu et al. and Kiesling et al., 分别).

3.1.3 Intention Detection. Another line of research concerns intention prediction.8 Analyz-
ing intentions in conversations is an old topic in natural language understanding, 在哪里
the goal is to detect what the speaker plans to pursue with their speech acts (Allen and
Perrault 1980). Compared with the Web search community, where predicting user inten-
tions from search queries and/or the user’s click behavior has been extensively studied
(Chen et al. 2002), there is little research that investigates how to extract intentions from
users’ free text.

The ﬁrst attempt was the use of indirect speech acts to detect e-mails requesting
行动 (科恩, Carvalho, and Mitchell 2004). E-mail intent detection is treated as a
binary classiﬁcation problem (request vs. nonrequest), leaving apart the difﬁcult de-
termination of the precise extent of the text that conveys this request. With the rise of
社交媒体, capturing intentions from user-generated content has become an emerging
research topic. Most approaches aim at assigning predeﬁned speech-act categories,
like ASSERTION, RECOMMENDATION, REQUEST, QUESTION, COMMENT. Methods vary
from supervised learning with bag-of-words representations to unsupervised models
exploiting surface features (例如, punctuations, emoticons), sentence-internal structure
(例如, parts of speech, dependency relations) (Zarisheva and Schefﬂer 2015; Vosoughi
and Roy 2016), or to a little extent, the conversational dependencies between sentences,
collapsing the set of user’s writings (tweets) into the same sequence (Joty and Hoque
2016).

3.1.4 Conversational Thread and Topic as Key Contextual Factors. Discourse analysis of social
media is a growing ﬁeld of interest in linguistics in general and in discourse analysis
尤其, with a signiﬁcant amount of the research published in journals such as
Discourse Studies or Journal of Pragmatics analyzing social media language, and even an
entire journal devoted to this ﬁeld (话语, Context & 媒体, published by Elsevier).
Although the study of discourse and context in computational linguistics is perhaps
not central, leveraging the context provided by the conversation thread and topic has
recently been the center of many NLP applications. Perhaps the best example comes
from sentiment analysis where conversations are used to enhance the performance of
polarity detection. 的确, although neighboring tweets tend to share similar polarity,

8 We use the term intention as a broader term that covers desires, 计划, 目标, and preferences.

672

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Benamara, Inkpen, and Taboada

Special Issue on Language in Social Media

the polarity orientation of the root (IE。, the original post/tweet) is usually shifted during
the reply process (黄, 曹, and Dong 2016). Vanzo, Croce, and Basili (2014) 模型
polarity detection as a sequential classiﬁcation task over streams of tweets about the
same topic and observe an improvement of about 20% in F1 measure compared with
approaches that do not account for the history of preceding posts. Ren et al. (2016)
incorporate word embedding vectors extracted from both the current tweet’s content
and the conversation context into a neural network, and measure the role of context
based on history tweets of the same author, which can serve as a prior for a tweet’s
情绪. The context-based neural model gains more that 10% in macro F-measure.

Figurative language processing is another area of research where conversation plays
a crucial role. With social media texts being very short, it is often difﬁcult to recognize
sarcasm or irony on the basis of the content of an utterance taken in isolation. 因此,
the context provided by the preceding messages can help in detecting the incongruity
between the literal meaning of an utterance and its intended meaning. Several ap-
proaches have been proposed to leverage such context, like Bamman and Smith (2015),
who explore the properties of the author (例如, proﬁle information and historical salient
条款), the audience (author/addressee topics), and the immediate communicative
环境 (previous tweets); and Wallace, Choe, and Charniak (2015), who exploit
signals extracted from the conversational threads to which the comments belong. 为了
a general discussion of context-based approaches to irony/sarcasm detection, we refer
the reader to Joshi, Bhattacharyya, and Carman (2017).

Topic prediction can also beneﬁt from document/posts sequential structure. 为了
例子, 戈什等人. (2016) recently propose Contextual Long-Short Term Memory
(CLSTM), a new sequence learning model that extends the recurrent neural network
LSTM by incorporating contextual features. CLSTM has been used for sentence topic
prediction: Given the words and the topic of the current sentence, predict the topic of
the next sentence.

3.2 On the Role of Other Contextual Phenomena

In addition to the discursive contextual phenomena that are mainly driven from posts’
conversation structure, there are many other types of context that can be combined with
linguistic content. Among them, we focus now on demographic information and social
network structure.

3.2.1 Demographic Information. This refers to author-related information like age, 性别,
种族, 收入, 地点, political orientation, and other demographic categories. 二
lines of research have recently gained relevance in the NLP community to derive demo-
graphic information from texts: author proﬁling and author identiﬁcation (Rosso et al.
2018; 斯塔马塔托斯等人. 2018). In the ﬁrst task, information such as the author’s age and
gender can be predicted, as authors who share similar demographic traits also share
similar linguistic patterns. In the second task, given a group of potential authors, 这
goal is to determine the right one (also known as authorship attribution). Whereas most
approaches mainly rely on lexical features derived from the linguistic content of the
message alone, recent approaches propose to account for discourse structure (Wanner
and Soler 2017).

When available, author-related information has been extensively used in different
NLP tasks, including sentiment/emotion analysis. 例如, several studies have
found strong correlations between the expression of subjectivity and gender (为了考试-
普莱, some subjective words will be used by men, but never by women, and vice versa),

673

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

and leverage these correlations for gender identiﬁcation (Burger et al. 2011; Volkova
and Bachrach 2016). Stylometric and personality features of users have also been used
for sarcasm detection (Hazarika et al. 2018).

Detecting the location of the social media users provides another type of demo-
graphic information useful in various applications. This information can be directly
available from user proﬁles or other meta-data (such as GPS information for posted
消息). When it is not available, it can be predicted based on the network structure
(“you are where your friends are”) or relations between those who follow and those
who are followed (Rout et al. 2013) or based on the content of the posted messages.
The latter content-based approaches extract information about the use of language, 这
main topics discussed, the named entities mentioned frequently, 等等. (Eisenstein
等人. 2010; Han, 厨师, and Baldwin 2012; Liu and Inkpen 2015). The accuracy of these
methods is not high, but it can be improved by combining content-based approaches
with the contextual information provided by the network structure and other location-
indicative meta-data.

3.2.2 Social Network Structure. In social media, social relationships between users enable
grouping users into speciﬁc communities. A community is often not identiﬁed in ad-
vance, but its users are expected to share common goals: circles of friends, members,
groups of topically related conversations, 等等. Drawing from the assumption
that users connected in the social network (例如, via followers, mentions, reply-to) 或者
that belong to the same community may have similar subjective orientations, several
studies show that users’ social relationships can enhance sentiment analysis (Tan et al.
2011). 例如, 黄, 辛格, and Atrey (2014) showed that modeling the social
network structure improves accuracy when detecting cyber-bullying messages.

4. Overview of the Articles in this Special Issue

This issue aimed to study how the treatment of linguistic phenomena, 尤其
at the discourse level, can beneﬁt NLP-based social media systems, and help such
systems advance beyond representations that include only bags of words or bags of
句子. Discourse and pragmatic information can also help move beyond sentence-
level approaches that typically account for local contextual phenomena relying on
dedicated lexicons and shallow or deep syntactic parsing. 更重要的是, the aim
of this issue is to show that incorporating linguistic insights, discourse information, 和
other contextual phenomena, in combination with the statistical exploitation of data,
can result in an improvement over approaches that take advantage of only one of those
perspectives.

We received a total of 15 submissions, reﬂecting a signiﬁcant interest in these phe-
nomena in the computational linguistics community. After a rigorous review process,
we selected six articles, covering various aspects of the topic. The selected articles
address deep issues in linguistics, computational linguistics, and social science. 这
special issue is structured around three main themes, according to the type of context
considered in each article:

Social context: The focus here is on the social and relational meaning in
online conversations from a theoretical point of view (Kiesling et al.).

Conversation turns and common-sense knowledge: 这里, we group papers that
study phenomena for which people make inferences in their everyday use

•

674

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Benamara, Inkpen, and Taboada

Special Issue on Language in Social Media

语言的, focusing on inferences that are drawn when searching for the
ﬁgurative meaning of an utterance (Ghosh et al.; Van Hee et al.).

•

Conversational context: The third part focuses on the role of discourse
phenomena in processing social media conversations, including topicality
(李等人。), speech acts (Joty and Mohiuddin), and argumentation
(Cocarascu and Toni).

The rest of this section provides a brief introduction to each of the six accepted

文件.

The article by Kiesling et al. (“Interactional Stancetaking in Online Forums”) investi-
gates thread structure and linguistic properties of stancetaking from the online platform
红迪网. Stancetaking captures the speaker’s (or writer’s) relationship to the topic of
讨论, the interlocutor, or audience, and the talk (or writing) 本身. 作者
ﬁrst propose a new data set where conversation threads are annotated according to
three linked stance dimensions: 影响, 投资, and alignment. These dimensions
are then predicted relying on lexical features. The quantitative and qualitative results
of this study show that stance utterances tend to pattern in coherent conversational
threads.

李等人. (“A Joint Model of Conversational Discourses and Latent Topics on
Microblogs”) extract topics from microblog messages, a challenging task given the
data sparsity in short messages that often lack structure and context. To address this
问题, the authors represent microblog messages as conversation trees based on their
reposting and replying relations, and propose an unsupervised model that jointly learns
word distributions to identify the different functions of conversational discourse and
various latent topics to represent content-speciﬁc information embedded in microblog
消息. Their experiments show that the proposed joint model on topic coherence
outperform state-of-the-art models. The output from the joint model is then used for
microblog summarization: By additionally capturing word distributions for different
sentiment polarities, the jointly modeled discourse and topic representations can effec-
tively indicate summary-worthy content in microblog conversations.

The article by Ghosh et al. (“Sarcasm Analysis Using Conversation Context”) stud-
ies the role of conversation to detect sarcasm in tweets and discussion forums. 这
context considered here concerns the current turn as well as the prior and the succeeding
一 (when available). In order to show to what extent modeling of conversation context
helps in sarcasm detection, the authors investigate both classical learning models with
linguistically motivated discrete features and several types of LSTM networks (condi-
tional LSTM network, LSTM networks with sentence-level attention). The models were
tested on different corpus genre data sets and the results show that attention models
achieve signiﬁcant improvement when using the prior turn as context for all the data
套. To better measure the difﬁculty of the task, the authors perform a qualitative
analysis of attention weights produced by the LSTM models and discuss the results
compared with human performance on the task.

In the article by Van Hee et al. (“We Usually Don’t Like Going to the Dentist: 使用
Common Sense to Detect Irony on Twitter”), the role of context in ﬁgurative language
detection is also explored. Compared with Ghosh et al., who focus on conversational
语境, Van Hee et al. target common sense and connotative knowledge and propose
to model implicit or prototypical sentiment (例如, “ﬂight delays,” “going to the dentist”
generally convey negative sentiment) in the framework of automatic irony detection
in tweets. Their approach uses a support vector machine classiﬁer relying on lexical,

675

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

句法的, and semantic features, with a particular focus on lexical and semantic features
that have been extended with language model features and word cluster informa-
的. The results show that applying sentiment analysis using SenticNet and real-time
crawled tweets is a viable method to determine the implicit sentiment related to that
concept or situation.

Cocarascu and Toni (“Combining Deep Learning and Argumentative Reasoning
for the Analysis of Social Media Textual Content Using Small Data Sets”) propose a
method to check whether news headlines support statements from tweets, to allow for
fact-checking. Their deep learning method extracts argumentative relations of attack
and support. Then they use the proposed method to extract bipolar argumentation
frameworks from reviews, to help detect whether they are deceptive. They show ex-
perimentally that the method performs well in both settings. 尤其, in the case of
deception detection, the method contributes a novel argumentative feature that, 什么时候
used in combination with other features in standard supervised classiﬁers, outperforms
the latter even on small data sets.

The last article in this special issue, by Joty and Mohiuddin (“Modeling Speech
Acts in Asynchronous Conversations: A Neural-CRF Approach”), presents a method
for speech act recognition, a problem that has long been a concern in the spoken
dialogue research community, and one that poses particular problems in online social
media communication, which tends to be asynchronous. Joty and Mohiuddin train
LSTM-RNNs using conversational word embeddings. This is a signiﬁcant result, 像他们
show that word embeddings trained on a related domain improve the performance
of the system. The contribution of this article is to incorporate context in the form of
dependencies across sentences. It is clear from the literature that conversation structure
is relevant when interpreting speech acts. The authors propose to model it as a graph
结构, given the nonlinear nature of asynchronous conversation. 此外. Joty
and Mohiuddin work from the hypothesis that, when representing sentence meaning,
word order is important, and should be preserved. Although this does not seem like
a revolutionary concept, word order is often disregarded in “classic” machine learning
方法, and in modern vector representations of text.

5. Conclusions and Future Directions

We hope that this special issue contributes to a deeper understanding of the role of
different types of context and their interaction to process social media data from the
perspective of discourse interpretation. We believe that we are entering a new age of
mining social media data, one that extracts information not just from individual words,
短语, and tags, but also uses information from discourse and the wider context. 最多
of the “big data” revolution in social media analysis has examined words in isolation—
a bag-of-words approach. We believe it is possible to investigate big data, 和社会的
media data in general, by exploiting contextual information.

To achieve that purpose, we need to ﬁrst develop tools to automatically determine
the structure of discourse, including discourse relations, argumentation, and threads
in conversations such as those found in Twitter and other social media. This is an
interdisciplinary enterprise that needs to address deep issues in both linguistics and
computational linguistics, including the analysis of the discursive properties of social
media content and the empirical study of how these properties are deployed in different
corpus genres through corpus annotation. We need to propose new solutions in various
use cases including sentiment analysis, detection of offensive content, and intention

676

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Benamara, Inkpen, and Taboada

Special Issue on Language in Social Media

detection. These solutions need to be reliable enough in order to prove their effective-
ness against shallow bag of words approaches.

Another direction of research that we encourage is to further explore the interac-
tions between content and extra-linguistic or extra-textual features, in particular time,
地方, author proﬁles, demographic information, conversation thread, and network
结构.

致谢
We would like to thank all the authors who
submitted articles and all the reviewers for
their time and effort. We also greatly thank
the journal editors, Paola Merlo and Hwee
Tou Ng, for their guidance and support
during the entire process.

参考
Addawood, Aseel and Masooda Bashir. 2016.

“What is your evidence?” A study of
controversial topics on social media.
In Proceedings of the Third Workshop on
Argument Mining, ArgMining 2016,
pages 1–11, 柏林, 德国.

Aiello, Luca Maria, Georgios Petkos,

Carlos J. Mart´ın, David Corney, Symeon
Papadopoulos, Ryan Skraba, Ayse G ¨oker,
Ioannis Kompatsiaris, and Alejandro
Jaimes. 2013. Sensing trending topics in
推特. IEEE Transaction of Multimedia,
15(6):1268–1282.

艾伦, J. F. 和C. 右. Perrault. 1980.

分析话语中的意图. 人工
智力, 15(3):143–178.

艾伦, Kelsey, Giuseppe Carenini, 和
Raymond T. 的. 2014. Detecting
disagreement in conversations using
pseudo-monologic rhetorical structure.
In Proceedings of the Conference on Empirical
Methods in Natural Language Processing,
EMNLP 2014, pages 1169–1180, Doha.
亚瑟, 尼古拉斯和亚历克斯·拉斯卡里德斯. 2003.

对话逻辑. 剑桥
大学出版社.

Attardo, Salvatore. 2000. Irony as relevant
inappropriateness. Journal of Pragmatics,
32(6):793–826.

Austin, John Langshaw. 1962. How to Do

Things with Words. 牛津.

巴赫, Kent. 1997. The semantics-pragmatics
distinction: What it is and why it matters.
VS Verlag f ¨ur Sozialwissenschaften.
pages 33–50.

Baly, Ramy, Mitra Mohtarami, James R.
Glass, Llu´ıs M`arquez, Alessandro
Moschitti, and Preslav Nakov. 2018.
Integrating stance detection and
fact checking in a uniﬁed corpus. 在
Proceedings of the Conference of the North

American Chapter of the Association for
计算语言学: 人类
语言技术, pages 21–27,
New Orleans, 这.

Bamman, David and Noah A. 史密斯. 2015.
Contextualized sarcasm detection on
推特. 国际会议录
Conference on Web and Social Media,
ICWSM 2015, pages 574–577, 牛津,
英国.

Barzilay, Regina and Mirella Lapata. 2008.

Modeling local coherence: An entity-based
方法. 计算语言学,
34(1):1–34.

Benamara, Farah, Nicholas Asher, Yannick
Mathieu, Vladimir Popescu, and Baptiste
Chardon. 2016. Evaluation in Discourse:
a Corpus-Based Study. Dialogue and
话语, 7(1):1–49.

Benamara, Farah, Maite Taboada, 和
Yannick Mathieu. 2017. Evaluative
Language Beyond Bags of Words:
Linguistic Insights and Computational
应用领域. 计算语言学,
43(1):201–264.

Bos, Johan. 2011. A survey of computational
语义学: Representation, inference and
knowledge in wide-coverage text
理解. Language and Linguistics
Compass, 5(6):336–366.

Bosc, 汤姆, Elena Cabrio, and Serena Villata.
2016. Tweeties squabbling: Positive and
negative results in applying argument
mining on social media. 在诉讼程序中
Computational Models of Argument, COMMA
2016, pages 21–32, 波茨坦.

Bunt, Harry. 2001. From lexical item to

discourse meaning: Computational and
representational tools. In Computing
意义, 体积 77 of Studies in Linguistics
and Philosophy. Springer Netherlands,
第 1–10 页.

Bunt, Harry and Bill Black. 2000. The ABC

of computational pragmatics. 约翰
Benjamins, pages 1–46.

Burger, John D., John Henderson, 乔治

Kim, and Guido Zarrella. 2011.
Discriminating gender on Twitter. 在
诉讼程序 2011 会议
自然语言的经验方法
加工, pages 1301–1309, 爱丁堡.

677

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

陈, 郑, Fan Lin, Huan Liu, Yin Liu,

Grosz, 乙. J。, Aravind K. Joshi, and Scott

Wei-Ying Ma, and Liu Wenyin. 2002. User
intention modeling in Web applications
using data mining. World Wide Web,
5(3):181–191.

科恩, William W., Vitor R. Carvalho, 和

Tom M. 米切尔. 2004. Learning to classify
email into “speech acts.” In Dekang Lin
and Dekai Wu, 编辑, 诉讼程序
Conference on Empirical Methods in Natural
Langugage Processing, EMNLP 2004,
pages 309–316, 巴塞罗那.

Deitrick, William and Wei Hu. 2013.

Mutually enhancing community detection
and sentiment analysis on twitter
网络. Journal of Data Analysis and
Information Processing, 1(3):19–29.

Dusmanu, Mihai, Elena Cabrio, 和瑟琳娜

维拉塔. 2017. Argument mining on Twitter:
Arguments, facts and sources. 在
诉讼程序 2017 会议
自然语言的经验方法
加工, EMNLP 2017, pages 2317–2322,
哥本哈根, 丹麦.

Eisenstein, 雅各布, Brendan O’Connor,

诺亚A. 史密斯, and Eric P. Xing. 2010.
A latent variable model for geographic
lexical variation. 在诉讼程序中 2010
Conference on Empirical Methods in Natural
语言处理, pages 1277–1287,
剑桥, 嘛.

Farzindar, Atefeh and Diana Inkpen.

2017. Natural Language Processing for
Social Media. 摩根 & Claypool
出版商.

冯, Vanessa Wei and Graeme Hirst. 2014.
A linear-time bottom-up discourse parser
with constraints and post-editing. 在
Proceedings of the 52nd Annual Meeting of the
计算语言学协会
(体积 1: Long Papers), pages 511–521,
巴尔的摩, 医学博士.

戈什, Aniruddha, Guofu Li, Tony Veale,
Paolo Rosso, Ekaterina Shutova, John A.
孩子们, and Antonio Reyes. 2015.
Semeval-2015 task 11: Sentiment analysis
of ﬁgurative language in Twitter.
In Proceedings of the 9th International
Workshop on Semantic Evaluation,
SemEval@NAACL-HLT 2015,
pages 470–478, 丹佛, 一氧化碳.

戈什, Shalini, Oriol Vinyals, Brian Strope,

Scott Roy, Tom Dean, and Larry P.
Heck. 2016. Contextual LSTM (CLSTM)
models for large scale NLP tasks. CoRR,
abs/1602.06291.

Grice, H. 保罗. 1975. Logic and conversation.
In Peter Cole and Jerry L. 摩根, 编辑,
Speech Acts. Syntax and Semantics, 体积 3,
学术出版社, pages 41–58.

678

Weinstein. 1995. Centering: A framework
for modelling the local coherence of
话语. 计算语言学,
21(2):203–225.

Grosz, Barbara J. and Candace L. Sidner.
1986. Attention, 意图, 和
structure of discourse. 计算型
语言学, 12(3):175–204.

Halliday, Alexander Kirkwood and Ruqaiya
哈桑. 1976. Cohesion in English. 劳特利奇.

Han, Bo, Paul Cook, and Timothy Baldwin.
2012. Geolocation prediction in social
media data by ﬁnding location indicative
字. COLING 论文集 2012,
pages 1045–1062, Mumbai.

Hazarika, Devamanyu, Soujanya Poria,
Sruthi Gorantla, Erik Cambria, 罗杰
Zimmermann, and Rada Mihalcea.
2018. CASCADE: Contextual sarcasm
detection in online discussion forums.
In Proceedings of the 27th International
Conference on Computational Linguistics,
前交叉韧带 2018, pages 1837–1848, 圣达菲, NM.

Heim, Irene. 1982. The Semantics of Deﬁnite
and Indeﬁnite Noun Phrases. 博士. 论文,
University of Massachusetts.

Hernault, H。, H. Prendinger, D. duVerle, 和

中号. Ishizuka. 2010. Hilda: A discourse
parser using support vector machine
分类. Dialogue and Discourse,
1(3):1–33.

Hobbs, Jerry. 1979. Coherence and

coreference. 认知科学, 3(8):67–90.
黄, Minlie, Yujie Cao, and Chao Dong.

2016. Modeling rich contexts for sentiment
classiﬁcation with LSTM. CoRR,
abs/1605.01478.

黄, Qianjia, Vivek Kumar Singh, 和
Pradeep Kumar Atrey. 2014. Cyber
bullying detection using social and textual
分析. In Proceedings of the 3rd
International Workshop on Socially-Aware
Multimedia, SAM ’14, pages 3–6,
纽约, 纽约.

Inkpen, Diana, Ji Liu, Atefeh Farzindar,

Farzaneh Kazemi, and Diman Ghazi. 2015.
Detecting and disambiguating locations
mentioned in Twitter messages. 在
Computational Linguistics and Intelligent Text
加工, CICLing, pages 321–332, Cairo.
Janssen, Theo M. V. 2001. Frege, contextuality

and compositionality. 杂志
Logic, Language and Information,
10(1):115–136.

Jaszczolt, K. 中号. 2012. Semantics and

pragmatics: The boundary issue. In K. von
Heusinger, 磷. Portner, 和C. Maienborn,
编辑, 语义学: An International
Handbook of Natural Language Meaning,

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Benamara, Inkpen, and Taboada

Special Issue on Language in Social Media

Mouton de Gruyter, 柏林,
pages 306–332.

Joshi, Aditya, Pushpak Bhattacharyya, 和

Mark J. Carman. 2017. Automatic sarcasm
detection: 一项调查. ACM Computing
Surveys, 50(5):1–22.

Joty, Shaﬁq, Giuseppe Carenini, 和

Raymond Ng. 2015. CODRA: A novel
discriminative framework for rhetorical
分析. 计算语言学,
41(3):385–435.

Joty, Shafiq R. and Enamul Hoque. 2016. Speech

act modeling of written asynchronous
conversations with task-speciﬁc
embeddings and conditional structured
型号. In Proceedings of the 54th Annual
Meeting of the Association for Computational
语言学 (体积 1: Long Papers),
pages 1746–1756, 柏林.

Joty, Shaﬁq R., Dat Tien Nguyen, 和

Muhammad Tasnim Mohiuddin. 2018.
Coherence modeling of asynchronous
conversations: A neural entity grid
方法. In Proceedings of the 56th Annual
Meeting of the Association for Computational
语言学, 前交叉韧带 2018, pages 558–568,
墨尔本.

Kamp, Hans and Uwe Reyle. 1993. 从

Discourse to Logic. 多德雷赫特.

Karoui, Jihen, Farah Benamara, V´eronique
Moriceau, Nathalie Aussenac-Gilles, 和
Lamia Hadrich-Belguith. 2015. Towards a
contextual pragmatic model to detect
irony in tweets. In Proceedings of the 53rd
Annual Meeting of the Association for
Computational Linguistics and the 7th
International Joint Conference on Natural
语言处理 (体积 2: Short
文件), pages 644–650, 北京, 中国.

Korta, Kepa and John Perry. 2015.

Pragmatics. In Edward N. Zalta, 编辑,
The Stanford Encyclopedia of Philosophy,
Metaphysics Research Lab, 斯坦福大学
大学. https://plato.stanford.edu/
archives/win2015/entries/pragmatics/.

Krifka, Manfred. 2002. Embedded speech
行为. In Proceedings of the Workshop In the
Mood, 法兰克福.

Lenci, Alessandro. 2006. The lexicon and the

boundaries of compositionality. Acta
Philosophica Fennica, 78:303–320.

Lenci, Alessandro. 2018. Distributional

models of word meaning. Annual Review of
语言学, 4(1):151–171.

林, Ziheng, Min-Yen Kan, and Hwee Tou

的. 2009. Recognizing implicit discourse
relations in the Penn discourse treebank. 在
诉讼程序 2009 会议
自然语言的经验方法
加工, pages 343–351, 新加坡.

刘, Ji and Diana Inkpen. 2015. Estimating

user location in social media with stacked
denoising auto-encoders. 在诉讼程序中
the 1st Workshop on Vector Space Modeling
for Natural Language Processing,
pages 201–210, 丹佛, 一氧化碳.

Londhe, Nikhil, Rohini K. Srihari, 和
Vishrawas Gopalakrishnan. 2016.
Time-independent and
language-independent extraction of
multiword expressions from Twitter.
In 26th International Conference on
计算语言学, 科林,
pages 2269–2278, 大阪.

Mann, William C. and Sandra A. 汤普森.
1988. Rhetorical Structure Theory: 走向
a functional theory of text organization.
Text, 8(3):243–281.

McNally, Louise. 2013. Semantics and

pragmatics. Wiley Interdisciplinary Reviews:
认知科学, 4:285–297.

Montague, 理查德. 1974. English as a formal

语言. In Richmond H. Thomason,
编辑, Formal Philosophy: Selected Papers of
Richard Montague, 耶鲁大学出版社,
新天堂, CT, pages 188–222.
Mukherjee, Subhabrata and Pushpak

Bhattacharyya. 2012. Sentiment analysis in
Twitter with lightweight discourse
分析. In Proceedings of International
Conference on Computational Linguistics,
科林 2012, pages 1847–1864, Mumbai.

Perret, J´er´emy, Stergos D. 阿凡特诺斯,

Nicholas Asher, and Mathieu Morey. 2016.
Integer linear programming for discourse
解析. 在诉讼程序中 2016
Conference of the North American Chapter of
the Association for Computational Linguistics:
人类语言技术,
pages 99–109, 圣地亚哥, CA.

Persing, Isaac and Vincent Ng. 2014. Vote
prediction on comments in social polls.
在诉讼程序中 2014 会议
自然语言的经验方法
加工 (EMNLP), pages 1127–1138,
Doha.

Polanyi, Livia and Martin van den Berg.

2011. Discourse structure and sentiment.
In Data Mining Workshops (ICDMW),
pages 97–102, Vancouver.

Prasad, Rashmi, Bonnie Webber, 和

Aravind Joshi. 2014. Reﬂections on the
Penn Discourse Treebank, comparable
语料库, and complementary annotation.
计算语言学, 40(4):921–950.

普斯特约夫斯基, James. 1995. The Generative

Lexicon. 与新闻界.

Recanati, Franc¸ois. 2004. Literal Meaning.

Literal Meaning. 剑桥大学
按.

679

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

Recanati, Franc¸ois. 2008. Pragmatics and
语义学. Blackwell Publishing LTD.
pages 442–462.

Ren, Yafeng, Yue Zhang, Meishan Zhang,

and Donghong Ji. 2016. Context-sensitive
twitter sentiment classiﬁcation using
neural network. 在诉讼程序中
Thirtieth AAAI Conference on Artiﬁcial
智力, AAAI 2016, pages 215–221,
Phoenix, AZ.

Rosso, 保罗, Francisco M. Rangel Pardo,
Iraz ´u Hernandez-Farias, Leticia C.
Cagnina, Wajdi Zaghouani, and Anis
Charﬁ. 2018. A survey on author proﬁling,
deception, and irony detection for the
Arabic language. Language and Linguistics
Compass, 12(4):1–20.

溃败, Dominic, 卡琳娜·邦切娃, Daniel
Preotiuc-Pietro, and Trevor Cohn. 2013.
Where’s @wally?: A classiﬁcation
approach to geolocating users based on
their social ties. In HyperText and Social
媒体 2013, pages 11–20, 巴黎.

Saloot, Mohammad Arshi, Norisma Idris,
AiTi Aw, and Dirk Thorleuchter. 2016.
Twitter corpus creation: The case of a
Malay chat-style-text corpus (MCC).
Digital Scholarship in the Humanities,
31(2):227–243.

Searle, 约翰·R. 1969. Speech Acts: An Essay in
the Philosophy of Language. 剑桥
大学出版社.

Sidarenka, Uladzimir, Matthias Bisping, 和
Manfred Stede. 2015. Applying Rhetorical
Structure Theory to Twitter conversations.
In Proceedings of the Workshop on
Identiﬁcation and Annotation of Discourse
Relations in Spoken Language (DiSpol),
pages 1–2, Saarbr ¨ucken.

Socher, 理查德, Alex Perelygin, Jean Wu,
Jason Chuang, Christopher D. 曼宁,
安德鲁·Y. 的, and Christopher Potts.
2013. Recursive deep models for semantic
compositionality over a sentiment
treebank. 在诉讼程序中 2013
Conference on Empirical Methods in Natural
语言处理, EMNLP 2013,
pages 1631–1642, Seattle, WA.

儿子, Youngseo, Anneke Buffone, Joe Raso,
Allegra Larche, Anthony Janocko, Kevin
Zembroski, H. Andrew Schwartz, and Lyle
Ungar. 2017. Recognizing counterfactual
thinking in social media texts. 在
Proceedings of the 55th Annual Meeting of the
计算语言学协会,
前交叉韧带 2017, pages 654–658, Vancouver.

Sperber, Dan and Deirdre Wilson. 1981. Irony
and the use-mention distinction. Radical
Pragmatics, 49:295–318.

680

Stamatatos, Efstathios, Francisco M. Rangel
Pardo, Michael Tschuggnall, Benno Stein,
Mike Kestemont, Paolo Rosso, and Martin
Potthast. 2018. Overview of PAN 2018 –
author identiﬁcation, author proﬁling, 和
author obfuscation. In CLEF 2018, 体积
11018 计算机科学讲义,
pages 267–285, 施普林格.

Taboada, Maite and William C. Mann. 2006.
Rhetorical structure theory: Looking back
and moving ahead. Discourse Studies,
8(3):423–459.

Tan, Chenhao, Lillian Lee, Jie Tang, 长的
Jiang, Ming Zhou, and Ping Li. 2011.
User-level sentiment analysis
incorporating social networks. 在
Proceedings of the 17th ACM International
Conference on Knowledge Discovery and Data
Mining, SIGKDD, pages 1397–1405,
圣地亚哥, CA.

Tarski, Alfred. 1983. Logic, 语义学,

metamathematics: Papers from 1923 到
1938. Hackett Publishing Company,
印第安纳波利斯, 在.

特尼, Peter D. and Patrick Pantel. 2010.

From frequency to meaning: Vector space
models of semantics. Journal of Artiﬁcial
Intelligent Research, 37(1):141–188.

Utsumi, Akira. 1996. A uniﬁed theory of

irony and its computational formalization.
In Proceedings of the International Conference
on Computational Linguistics, 科林,
pages 962–967, 哥本哈根.

Vanzo, 安德里亚, Danilo Croce, and Roberto
Basili. 2014. A context-based model for
sentiment analysis in Twitter. 在
Proceedings of the 25th International
Conference on Computational Linguistics,
科林 2014, pages 2345–2354,
都柏林.

Volkova, Svitlana and Yoram Bachrach. 2016.
Inferring perceived demographics from
user emotional tone and user-environment
emotional contrast. 在诉讼程序中
54th Annual Meeting of the Association for
计算语言学, 前交叉韧带 2016,
pages 1567–1578, 柏林.

Volkova, Svitlana, Glen Coppersmith, 和

Benjamin Van Durme. 2014. Inferring user
political preferences from streaming
通讯. In Proceedings of the 52nd
Annual Meeting of the Association for
计算语言学, 前交叉韧带 2014,
pages 186–196, 巴尔的摩, 医学博士.

Vosoughi, Soroush and Deb Roy. 2016. Tweet
行为: A speech act classiﬁer for Twitter. 在
Proceedings of International AAAI Conference
on Web and Social Media, ICWSM 2016,
pages 711–715, Cologne.

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Benamara, Inkpen, and Taboada

Special Issue on Language in Social Media

华莱士, Byron C., Do Kook Choe, 和

Eugene Charniak. 2015. Sparse,
contextually informed models for irony
detection: Exploiting user communities,
entities and sentiment. In Proceedings
of the 53rd Annual Meeting of the
计算语言学协会
and the 7th International Joint Conference
on Natural Language Processing,
ACL-IJCNLP, pages 1035–1044,
北京.

Wanner, Leo and Juan Soler. 2017. 上
relevance of syntactic and discourse
features for author proﬁling and
识别. In EACL 2017,
pages 681–687, Valencia.

Webber, Bonnie, Markus Egg, and Valia

Kordoni. 2012. Discourse structure and
language technology. 自然语言
Engineering, 18(4):437–490.

西方, 罗伯特, Hristo S. Paskov, Jure

Leskovec, and Christopher Potts. 2014.
Exploiting social network structure for
person-to-person sentiment analysis.

Transactions of the Association of
计算语言学 (处理),
2:297–310.

Zarisheva, Elina and Tatjana Schefﬂer. 2015.

Dialog act annotation for Twitter
conversations. In Proceedings of the 16th
Annual Meeting of the Special Interest
Group on Discourse and Dialogue, SIGDIAL
2017, pages 114–123, Prague.

Zimmermann, 时间. 乙. 2013. The Oxford
handbook of compositionality. 在
Wolfram Hinzen, Edouard Machery and
Markus Werning, 编辑, Compositionality
Problems and How to Solve Them, 牛津
大学出版社, pages 81–106.

Zubiaga, Arkaitz, Elena Kochkina, Maria

Liakata, Rob Procter, and Michal Lukasik.
2016. Stance classiﬁcation in rumours
as a sequential task exploiting the tree
structure of social media conversations.
In Proceedings of the 26th International
Conference on Computational Linguistics:
技术论文, 科林 2016,
pages 2438–2448, 大阪.

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

681

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
6
3
1
8
0
9
9
0
3
/
C
哦

我
我

_
A
_
0
0
3
3
3
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

682
下载pdf