2 Linguistic theories of irony
2.1 Echoic mention theory
2.2 Pretense theory
2.3 Allusional pretense theory
2.4 Unified theory
3 Automatic detection of irony
3.1 “Yeah Right”: Sarcasm Recognition for Spoken Dialogue Systems .
3.2 Lexical Influences on the Perception of Sarcasm
3.3 Using LSA to detect Irony
3.4 Clues for Detecting Irony in User-Generated Contents: Oh...!! It’s “so easy” ;-)
3.5 Detecting Ironic Intent in Creative Comparisons
3.6 Automatic Satire Detection: Are You Having a Laugh?
3.7 Semi-Supervised Recognition of Sarcastic Sentences in Twitter and Amazon
4 Future Work
“Mr. Play It Safe was afraid to fly
He packed his suitcase and kissed his kids good-bye
He waited his whole damn life to take that flight
And as the plane crashed down he thought
‘Well isn’t this nice...’
And isn’t it ironic...don’t you think”
The situation described by Morissette (1995) is intuitively ironic since one assumes flying to be safe, thus the man’s fear of flying to be unjustified. When the man dies during his first flight, this is far from expectation, which gives rise to the situational irony. When the plane crashes down, he recognizes the irony in this situation and thinks “Well isn’t this nice?”. This thought is perfectly utterable to an audience which makes it an instance of verbal irony.
Other forms of irony are socratic irony which is the act of Sokrates asking simple questions in order to destroy another one’s argument and irony of fate in which one acts solely in order to escape a foretold fate and only this acting then makes the un- wished fate possible.
This course work focuses on verbal irony, so it is only concerned with ironic utter- ances as the one above. Intuitively, such utterances can be characterized as meaning something different than what is actually said and expressing an opposition of what should be and what is. Sarcasm is a known particularly mocking variety of ironic ut- terance which is widely used in satire. Since an exact differentiation between irony and sarcasm is not aimed at in this work and may in general be a subjective and not solely objective matter, from now on, the expressions irony and sarcasm will be used as syn- onyms to verbal irony.
Human communication often involves the use of irony. In many cases, it is far from ob- vious if an utterance is meant ironical or not. Context and world knowledge are needed to discriminate literal from ironic intent. Linguists have worked on describing the nature of irony and come up with some ideas which reflect the intuitive understanding of irony. Parallely, computational linguists are confronted with the challenge of automatically detecting irony. When an utterance contains irony, the only chance of getting the in- tent, is understanding and interpreting the irony in it. If you don’t get the irony in the introductory example, you would actually end up with the belief that the man thought the situation to be nice or the man to be happy, which is more like the opposite of the true intent.
I review different theories of irony in chapter 2. Chapter 3 describes the state-of- the-art of automatic irony detection. Chapter 3 covers the importance of corpus study for future research and proposes a fusion between theory, corpus study and automatic detection.
2 Linguistic theories of irony
Since Grice’s theory of pragmatics (Grice, 1975), linguists have developed theories of irony. Approaching the topic from different perspectives, these theories have struggled with a multitude of questions:
What does it mean for an utterance to be literal or ironic?
In how far is an ironic utterance different from its literal counterpart?
How does a speaker / writer indicate his ironical intent?
And how may it be recognized by the listener / reader?
(1) I’m a lucky man.
Imagine (1) being uttered by a man who just lost all his money in a casino in Las Vegas. Following Grice (1975), this utterance is ironic because its literal meaning is obviously false, since loosing all your money in a casino contradicts being lucky. By violating the Gricean maxim of quality (“Do not say what you believe to be false”), the speaker signals the listener that the literal meaning of (1) must be rejected. By recognizing this maxim violation the listener knows he must adopt the contrary meaning for (1), namely that the speaker is not a lucky man, but only speaking ironically.
As Juez (1995) shows, ironic comments may not only violate the Gricean maxim of quality but also the other maxims. She argues, the maxim of quantity may be violated by understatements in general, since understating involves saying too few. The maxim of relevance is violated by intransparent euphemisms. Finally, Juez argues, the maxim of manner to be violated by her example utterance “She’s not of the most helpful variety” which is neither brief nor clear.
(2) Am I not a lucky man?
More critics on Gricean account of irony concern the reinterpretation of ironical utter- ances by assuming the opposite of the literal meaning. Sperber and Wilson (1981) note at least three problems which arise from it: First, if the meaning of an ironic utterance is simply the opposite of the literal meaning, there is no reason why people should ever use the ironic form at all as understanding irony may need more effort than understanding literal intent.
Second, the rhetorical question (2) has quite similar (communicative and humorous) ef- fects when being uttered in the same context as (1). Still, there is neither an “opposite” of (2) which could count as Gricean ironic reinterpretation nor do rhetorical questions count as ironic utterances per se. And third, if there is a special tone of voice (Grice, 1975) that lets us detect irony if the speaker’s beliefs are unknown, then how can irony be detected if there is no such special tone of voice, e.g. in written text?
In this section, I will review four different post-Gricean linguistic theories of irony. These theories try solve problems which were inherent in Grice’s theory and also ask and answer new questions like those indicated above. The main ideas, improvements on the classic approach but also prevailing problems will be analyzed. Some aspects of the theories may be helpful for the task of identifying ironic utterances automatically. I will comment on this “helpfulness for computational linguists” for each theory.
2.1 Echoic mention theory
A basic question for all theories of irony is what should count as irony. Traditionally, the so-called figurative utterances are classified as stylistic devices such as e.g. irony, rhetorical questions, metonymy, hyperbole. Sperber and Wilson (1981) judge these de- vices to be defined quite fuzzy by some set of their representatives which are themselves defined rather fuzzy. Instead of trying to stick to this abstract historical classification, they propose to group utterances by the effects they produce and commonly underlying psychological mechanisms.
(3) I’m the luckiest person on earth.
Following this approach, the statement (1), the rhetorical question (2) and the hy- perbole (3) may all be classified as ironic, since all three may be described as “gallows humor” and the commonly underlying psychology intuitively is the speaker mocking himself in order to improve his ability to cope with his disastrous situation compared to more “basic” emotional reactions like bursting into tears or screaming out loudly.
All to be reviewed theories agree with Sperber and Wilson (henceforth SW) on that point of view, namely their theories of irony are not concerned with finding similarities between irony discriminated from hyperbole, rhetorical questions etc. but instead try to describe mechanisms involved in irony and deduce from that which example utterances may be called ironic and which not. That is, the upcoming theories try to describe irony instead of prescribing it.
(4) How I love winning!
In their paper “Irony and the use-mention distinction”, Sperber and Wilson (1981) introduce their theory of irony as an instance of echoic mention. When a speaker utters an ironical statement, the authors argue, he does not use this utterance, but mention it. The mentioned proposition is an indirect quoation of a concrete utterance or a thought. Someone using (4) refers to the achievement of winning, while someone mentioning (4) refers to the linguistic expression of (4) itself.
Example (4) could either echo the speaker’s or another person’s literally intended ut- terance that he loved winning. According to SW, it is possible that there is no concrete antecedence for an ironic utterance. (4) could merely quote a thought or an opinion, like the social norm that one generally likes (or should like) to win.
SW argue that a speaker may echo a remark in order “to suggest that he finds it untrue, inappropriate or irrelevant” (Sperber and Wilson, 1981, p. 307). Thus, this theory distinguishes an ironic utterance from the opposite of a corresponding literal utterance. The use of irony here has a clear psycholinguistic function.
Clark and Gerrig (1984) point out a problem with this account, namely that for every ironic utterance an antecedent has to be found and sometimes there simply may not exist one to be echoed. Clark and Gerrig give the example of highly creative literature (for instance, Jonathan Swift - A Modest Proposal), which produces original ideas for which it is very unplausible that there be an antecedent which they simply echo. To deal with this problem, Wilson and Sperber (1992) generalized their account from echoic mention to echoic interpretation. They no longer analyze irony as mention of an attributed thought or utterance, but as interpretation of an antecedent utterance, thought or opinion. This is supposed to - inter alia - solve the problem of original ironic utterances, since they can now treated as interpretation or implication of a previous existing utterance, thought or opinion.
I think, the concept of an echoic interpretation is a more harmful than helpful extension of the echoic mention. First, it is too general, because for virtually every utterance you may find an antecedent utterance, idea, thought or opinion which you can interpret creativelly to end with the utterance you started with. Second, it is not even necessary to extend the concept of echoic mention to interpretation which I interpret as being as general as the former. This is because the antecedents being echoed by the ironic utterances are quite arbitrary. For instance, the echoed utterance is allowed to be from some imaginary person (Sperber and Wilson, 1981, p. 309-310). That way you can always think of an imaginary person who uttered the sentence or thought before and thus can successfully identify all cases of irony as cases of echoic mention. You do not even need to look for an echoic interpretation, since all irony is echoic mention.
From a computational perspective, the search for an appropriate antecedent is infeasi- ble, since it can generally be everywhere or nowhere. Everywhere if the ironic utterance is a very common one with possibly a multitude of interpretations. Or nowhere, when an ironic utterance does not echo an actual utterance, but merely a thought, belief or an opinion. To have at least a chance to grasp irony of the latter case computationally, points to the usage (and thus construction) of large, hence expensive knowledge ontolo- gies. Searching for echoic mentions may therefore be a helpful tool to recall at least a small amount of all irony, but search should be restricted to concrete linguistic echoes to keep costs at a resonable level.
2.2 Pretense theory
As an alternative to the echoic mention theory, Clark and Gerrig (1984) propose the pre- tense theory of irony. When speaking ironically, the theory goes, the speaker pretends to be someone else. When uttering (4), the speaker may pretend to be someone who did not get the game’s rules he just lost. By pretending to be such a person, the speaker ridicules the person he pretends to be. Since pretending to be another person is close to imitation, the pretense may (but does not have to) be indicated by a special tone of voice characteristical for the pretended (type of) person. While noted by several theorists, the special tone of voice was empirically investigated by Gibbs (2000) and Traum and Narayanan (2006).
So, when uttering (4), the speaker may pretend being some person who did not get the game’s rules and just lost. Opposing to the echoic mention theory, in the pretense theory there does not need to be any antecedent to echo. The sentence does not need to have been uttered before or reflect a common opinion or thought. The only common ground there should be, Clark and Gerrig (henceforth CG) argue, is some common knowledge in the audience. The audience may only recognize the speaker’s ironic intent if they can infer which person or stereotype the pretended speaker is ought to be. At the very least, the audience has to figure out that the speaker is pretending at all.
All ironic mentions can be translated into pretense.
Moreover, CG argue, the speaker does not only pretend to be some other speaker, but also to talk to a pretended audience, namely an audience who would take the pretense serious. This way, the speaker not only expresses his (usually contemptuous) attitude towards the pretended speaker and his utterance but also towards the pretended audi- ence. The authors call these, the “different types of victims” of irony. One victim is the pretended speaker of the utterance, while the other victim is the audience taking the utterance serious. The echoic mention theory does not differ between these two types of victims, CG argue. Furthermore, CG explain, echoic mention is just a special type since all echoic mention can be translated into pretense, namely the pretense of being that person who the antecedent originates from.
The authors pronounce the game theoretic aspect of irony, namely that the speaker deceives the audience which is consequencely prompted to discover the truth behind the pretense. Since speakers pretend to be someone else, the pretense theory gives an intuitive explanation for the special tone of voice of ironic utterances.
2.3 Allusional pretense theory
Kumon-Nakamura et al. (1995) propose the allusional pretense theory which is connected to the classic Gricean theory and the one of echoic mention. The allusional pretense the- ory explains, how utterances of any major speech act categories (assertives, directives, commissives, expressives, declarations) can convey irony if the utterance is “used to al- lude to a failed expecation or violated norm” (Kumon-Nakamura et al., 1995, p. 89). An ironic utterance may be identified by an inherent pragmatic insincerity which is either a violation of a Gricean maxim or a felicity condition for speech acts.
(5) You sure know a lot.
Examples (2)-(4) violate a felicity condition for assertives, namely the one that the assertion must be true. They allude to the failed expectation of winning the game or more generally to the violated norm of winning games. The same allusion is observed for example (5) which violates the felicity condition for questions that you expect an answer when you ask something. This insincerity is inherent in all rhetorical questions.
Kumon-Nakamura (henceforth KN) formulate for all major speech acts how they may be pragmatically insincere, i.e. which violations of felicity conditions may appear. They also note that their concept of allusion subsumes those of echoic mention since mentioning something is the directest way of alluding to it. Similarly the necessity of pragmatic insincerity for ironic utterances generalizes Grice’s requirement of a violation of the maxim of quality, but with their approach a violation of any maxim or felicity condition may suffice, it need not be the maxim of quality.
KN give a plausible explanation for the asymmetry of the statement-meaning- relation, i.e. why ironic utterances are much more likely to be formulated positively and meant negative than the other way around. This is because conventional expectations are more likely to be positive than negative.
(6a) That tastes good!
(6b) That tastes bad!
Consider (6a) being uttered when eating at a friend’s place. (6a) is much more likely than (6b) to convey irony, since the expectation for food is to taste good or since when eating someone else’s food, a polite comment is normal. Since the expectation is normally one of good taste, (6a) is much more likely to allude to a failed expectation and thus to convey irony than (6b), since there is no norm of bad food.
However, KN note, that (6b) is still a possible ironic quotation. Someone else may have uttered before dinner that the food will taste bad. Now, that they found out the food to be actually very delicious, quoting this wrong assumption would clearly convey ironic and mock the quoted person.
KN perform several experiments in their paper which they interpret as evidence for their claim that ironic utterances allude to failed expectations and regarding the asym- metry of the statement-meaning-relation. There is an asymmetry in the politeness- meaning-relation, since overpolite utterances are more often interpreted as ironic in their experiments than underpolite ones. However, they experimentally show that ex- pectations need not to be positive per se. Listeneres also detected an ironic intent when negative expectations were failed.
2.4 Unified theory
Utsumi (1996) proposes the unified theory of irony in which he formally defines an ironic utterance. He defines an ironic environment and an implicit display of ironic en- vironment and then arranges those utterances as ironic which implicitly display an ironic environment. The ironic environment is formalized through situation theory and action theory while the implicit display is defined in terms of standard pragmatic theory.
Utsumi agrees with Kumon-Nakamura et al. (1995) in that an ironic utterance alludes to a failed expectation, but Utsumi elaborates on how the speaker’s expectation can fail. Either, the speaker expects to fulfil his expectation by performing an action A or the expectation is fulfilled by doing nothing. In the first case, the expectation can fail if an action A’ is performed which hinders A or if A is not perfromed. In the latter case the expectation can fail if an action is performed or it can fail accidentally. However, any of these ways of failed expectations creates an ironic environment.
For an utterance U to implicitly display an ironic environment, three requirements have to be fulfilled. First and second, as Kumon-Nakamura et al. (1995) put it before, U has to allude to the speaker’s expectation and to contain pragmatic insincerity. And third, the utterance has to imply the speaker’s emotional attitude towards the failure of the speaker’s expectation. Note, that it suffices for an implicit display when two out of three requirements are fulfilled. Utsumi argues, this explains why an utterance can be understood as ironic although the speaker did not intend it to be understood as irony (Kumon-Nakamura et al., 1995, p. 966).
I will elaborate on Utsumi’s example 2 (Kumon-Nakamura et al., 1995, p. 963) where a mother asks her son to clean up his room which he does not do satisfactory. She then utters that she “love[s] children who keep their rooms clean”. Since the mother’s expectation of a clean room can be fulfilled by an action which was not performed by the son, the situation creates an ironic environment. It also implicitly displays this ironic environment since all three conditions for an implicit display are fulfilled: The mother alludes to her expectation of a clean room, she implies her emotional attitude towards the expectation (she loves clean rooms, implying she hates messy rooms) and she is being pragmatically insincere by violating the Gricean maxim of relevance with her utterance.
In the way, Utsumi elaborates on the different kinds of failing an expectation, his ac- count is an improvement over Kumon-Nakamura et al. (1995) (although the latter gives psycholinguistic evidence for his claims, see chapter 2.4), but what really sets his account apart from other theories of irony, is the intense formalism he employs. To explain the just elaborated example 2, he needs formulae which describe the instantiated causal relations, i.e. the causal relations needed to build a formal argument along the lines of my verbal argument for example 2.