Detecting and Generating Ironic Comparisons: An Application of Creative Information Retrieval

Ironic utterances promise an expected meaning that never arrives, and deliver instead a meaning that exposes the failure of our expectations. Though they can appear contextually inappropriate, ironic statements succeed when they subvert their context of use, so it is the context rather than the utterance that is shown to be incongruous. Every ironic statement thus poses two related questions: the first, “what is unexpected about my meaning?” helps us answer the second, “what is unexpected about my context of use?”. Like metaphor, irony is not overtly marked, and relies instead on a listener’s understanding of stereotypical norms to unpack its true meaning. In this paper we consider how irony relies upon and subverts our stereotypical knowledge of a domain, and show how this knowledge can be exploited to both recognize and generate ironic similes for a topic. The Soul of a new cliché Samuel Goldwyn, the co-founder of MGM studios, famously summed up Hollywood’s attitude to creativity with the line “Let’s have some new clichés”. On the face of it, this seems like just another one of Goldwyn’s many memorable misstatements (like “include me out!”): after all, it’s hard to think of clichés as new, or as something that can be invented on demand. Yet, on closer analysis, one can find real insight in Goldwyn’s remark. Clichés are considered anathema to the creative process because they represent everything that is conventional and jaded about the status quo. However, clichés become tired thru overwork, and are overworked precisely because they prove themselves so useful in so many different contexts. Few writers set out to create a new cliché, but most would like their efforts to become as much a part of the fabric of our linguistic culture as the most tenacious of clichés. Copyright © 2012, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. One productive form of new cliché is the humorously pithy comparison, as in “as durable as a chocolate teapot” or “as useful as a screen door on a submarine”. Speakers recognize memorable comparisons when they hear them, and re-use them as eagerly as one retells a favorite joke. The most frequently reused comparisons can, in this way, acquire the clichéd status of a proverbial simile. When the folklorist Archer Taylor collected his corpus of proverbial similes in 1954, he observed not just a wide variety of humorous comparisons in American speech, but a wide variety of humorous forms for the same descriptive qualities, such as “durable” and “useful”. Speakers are clearly drawn to popular comparisons of proven value, but are equally fond of coining their own, in the hope that their witty new descriptions are widely reused by others in turn. This constant churn of re-invention keeps our language fresh, and ensures that ironic comparisons retain their ability to challenge and to entertain, even as others – such as “crazy like a fox!” and “as clear as mud!” – acquire an idiomatic status which makes them effortlessly understood. Stereotypes anchor a comparison in the realm of the familiar, where judgments about representativeness are made quickly and intuitively by cognitive processes that Kahneman (2011) dubs System 1. Conversely, similes help to perpetuate stereotypes by packaging them into such conveniently re-usable linguistic forms. However, ironic similes subvert the workings of System 1, and force us to engage a set of deliberative, analytic and non-intuitive processes that Kahneman dubs System 2. Whereas straight similes use stereotypes as exemplars of a given quality, ironic similes highlight the unexpected lack of this quality by instead using a withering counter-example, perhaps constructed from stereotypes that one intuitively associates with very different (and even antonymous) qualities. Humorously ironic comparisons often construct this counter-example by subverting the normal form of a stereotype. Thus, while we expect a vault to be the very model of a secure location, we might describe a weak and unsafe container as being “as secure as a chocolate vault”. The stereotypes “chocolate” and “vault” are combined here to produce a complex concept with internal incongruities (secure and strong versus brittle and soft) that lead to emergent qualities (such as not secure at all). Novel combinations such as “chocolate vault” require insights that can only come from System 2, to override those that are produced by System 1. Nonetheless, after encountering a variety of sardonic utterances that are anchored in a construct such as “chocolate X”, we might infer a general rule: If X is a stereotype of strength or solidity or durability or some related quality, then an X made of chocolate will have none of these qualities. Repeated exposure to the “chocolate X” pattern will help us acquire a meta-cliché that we can use, in turn, to generate novel instances of our own, such as “chocolate tank” and “chocolate castle”. Of course, we cannot simply stick the prefix “chocolate” in front of an arbitrary stereotype and hope for the best; rather we need a means of identifying meaningful combinations. An approach to linguistic creativity called CIR, Creative Information Retrieval, offers such a means (Veale, 2011). We describe in this paper how CIR allows us to harvest readymade phrases from a corpus such as the Google ngrams, and allows us to identify the most likely qualities of the complex concepts denoted by those phrases. In the next section we show how the stereotypical knowledge exploited by CIR is acquired from web-harvested similes, and consider the signs, and the likelihood, that a webharvested simile is in fact ironic. We then describe how this knowledge is operationalized in an informationretrieval guise as a set of powerful query operators, which can be used to retrieve similes, metaphors, analogies and other pithy descriptions from text corpora. Finally, we explore how this knowledge can be applied to a corpus of creative similes that has been annotated for irony, to automatically learn a collection of meta-clichés than can then be used to generate new examples of ironic similes. Stereotypes, Similes and Irony Similes – even ironic similes – exploit our mutual knowledge of stereotypical norms to communicate their meaning, even if these norms are ultimately subverted. As such, similes are also an ideal source of stereotypical knowledge for a tabula rasa agent such as a computer, for as Dickens notes in A Christmas Carol, “the wisdom of our ancestors is in the simile”. There is, however, an important caveat: such agents are liable to make the worst mistakes possible if they misread a speaker’s intention to be ironic. Veale & Hao (2007a) explored the use of similes as a source of stereotypical norms, and harvested tens of thousands of simile bodies (of the form “as X as a Y”) from the web. Lacking an automatic procedure for filtering straight similes from ironic similes, Veale & Hao filtered their simile-set manually, to obtain a collection of over 12,000 unique straight simile bodies (such as “as hot as an oven”) and almost 3000 unique ironic simile bodies (such as “as subtle as a sledgehammer”). Veale & Hao (2007b) then used the straight similes as a source of stereotypical norms (e.g. ovens are hot, jungles are humid, snow is soft) in a system for comprehending and generating metaphors. The unused ironic similes were later used as a test-set for an irony detection algorithm in Hao & Veale (2010). A simple question was posed of each simile in the combined set (straight + ironic): can a simile “as X as Y” be plausibly recast in a form that is highly likely to be non-ironic, or can it more plausibly be recast in a form that is highly likely to be ironic? For the idealized non-ironic form, the pattern “X Zs such as Xs” was chosen, since irony is very infrequently observed in this construction. Thus, the simile “as hot as a sauna” is recast as “hot Zs such as saunas”. For the idealized ironic form, the pattern “about as X as Y” was chosen, since “about” has been hypothesized to be a subtle marker of irony (Moon, 2008). Thus, the simile “as muscular as a paper-clip” is recast as “about as muscular as a paper-clip”. To test the plausibility of each recasting, simple web frequency for each form was calculated using the search engine Google. Though Veale & Hao describe a number of ancillary heuristics in their methodology, this simple recasting is the backbone of their approach, which achieves promising results overall: 87% of ironic similes are recognized with .63 precision, while 89% of straight similes are recognized with .97 precision. The F-score for classification of ironic similes is .73, for straight similes it is .93, and for irony/straight classification overall it is .88. Veale (2012) used the “about” construction to harvest a collection of irony-rich similes, and extended the form of the search to accept complex similes in which the vehicle is a multiword phrase (such as “a wet rag” or “a chocolate teapot”). This collection of 20,299 web similes was then hand-annotated for irony, so that it could be used as a goldset for future irony studies. The collection lends further support to the role of “about” and other markers of imprecision (such as “not exactly X” and “almost as X as”) in marking irony, since 76% of the newly harvested similes (such as “about as modern as a top-hatted chimneysweep”) are marked as ironic. Just 14% of these 20,299 “about” simile types use a vehicle with a single content-word, and a mere 3% (i.e., 676 types) are also found in the original harvesting process of Veale & Hao (2007a). So the “about” marker not only signals a speaker’s intention to be playful and ironic, it also signals a speaker’s intention to be humorous and perhaps creative. This finding suggests that when computers aim to be ironic in their descriptions, the use of the “about” construction will help ensure that their efforts are more readily received as well-formed irony. Creative Information Retrieval Ironic utterances can be viewed as congruous statements in incongruous contexts. As Sperber & Wilson put it in their echoic-mention theory of irony, an ironic utterance echoes one that would plausibly be made in a very different context, thus drawing an audience’s attention to a failure of expectations in the current context. One can ironically echo an utterance that is explicitly associated with another context (an explicit echo) or one can allude to a general expectation arising from another context (an implicit echo). Fludernik (2007) generalizes this notion of an implicit echo to include any pointed allusion to accepted beliefs, received wisdom, or stereotypical norms. In any case, to process irony a computer needs the ability to retrieve either the appropriate utterance or the appropriate norm. Veale (2011) describes a platform for the retrieval of creative language that can be used to achieve each of these ends. Creativity often arises from a process of appropriation, in which something is wrenched from its normative context of use and given new meaning in a new setting. In this vein, Duchamp popularized the notion of an artistic readymade when his Fountain – a signed urinal – was presented with some controversy at a Dada exhibition in 1917. We normally think of readymades as physical objects whose artistic merit derives wholly from their selection by an artist, but language is also rich in linguistic readymades. Just think of how many movies, songs, novels, poems, etc. allusively borrow utterances and phrases from each other. For example, the movie The Usual Suspects takes its name from a famous quote from the movie Casablanca, while the novel All The King’s Men takes it title from a famous nursery rhyme; this title has, in turn, inspired the title of Woodward and Bernstein’s book All The President’s Men. The well-formed phrases that one can extract from the Google database of n-grams can be viewed as a vast collection of linguistic readymades, insofar as each is a phrase that is wrenched from its original context of use on the web. To find the right readymade for a given task, one needs a set of retrieval tools that go beyond the normal range of query operators that are employed in standard IR. Veale (2011) describes a set of non-literal query operators that allow a user to search for n-grams on the basis of metaphorical similarity rather than literal string similarity. These operators are denoted @, ? and ^ and defined thus: @ is the stereotype operator, where @noun will match any stereotypical property of noun and @adj matches any noun denoting a concept that has the stereotypical property adj. For example, @razor will thus match the words “sharp”, “straight”, “clean” and “smooth”, while @sharp will match any of the words “razor”, “knife”, “sword”, and so on. ? is the neighborhood operator, where ?noun will match any other noun that is frequently clustered with noun, and ?adj will match any other adjective that is frequently used with adj in a complex description. The neighborhood for nouns is determined by gathering matches for the pattern “Xs and Ys” (a coordination of bare plurals) in the Google n-grams (such as “angels and demons”, “doctors and nurses”, “cups and saucers”, etc.), while the neighborhood for adjectives is determined by gathering matches for the pattern “as X and Y as” (a multi-property simile) on the web. For example, ?disaster will match any of the words “tragedy”, “catastrophe”, “calamity”, “misfortune”, “hardship”, “plague”, “famine” and so on, while ?tragic will match any of the words “sad”, “shocking”, “terrible”, “unfortunate”, “ridiculous” and so on. All matches for ?noun are ranked by the WordNet similarity of each match to noun, while matches for ?adj are ranked by the cooccurrence frequency of each match with adj on the web ^ is the category operator, where ^class will match any member of the pre-defined category class. Users can give names to their own ad-hoc categories, or designate names for the results of previous retrievals (for instance, a user might populate the category ^sharptool with words that match the CIR query @sharp & ^tool). The ^ operator can also be used to refer to WordNet categories; e.g., ^person matches any noun denoting a kind of person in WordNet. To these three operators we add a fourth: is the antonym operator, where –P will match any antonym of the adjective P. Antonyms are derived from the lexical resource WordNet. This, for example, -soft matches the word “hard”, and –strong matches the word “weak”. Importantly, these non-literal operators can be used in complex combinations. For instance, ?@adj will match any noun in the neighborhood of a noun / concept that exhibits the stereotypical property adj, while @-adj will match any noun/concept that is a stereotypical holder of any property denoted by an antonym of adj. Likewise, @?adj will match any noun / concept with a stereotypical property that is like adj, while @@noun will match any other noun / concept that shares at least one stereotypical property with noun. Veale (2011) describes how non-literal operators, when used to augment an IR system for retrieving Google ngrams, can provide a generic foundation for retrievaloriented linguistic creativity. For instance, CIR can be used to turn the readymade phrases of the Google n-grams into vehicles for creative comparison. For a topic X and a property P, straight similes of the form “X is as P as S” are easily generated by retrieving values for S ∈ (@P ∩ ??X). Likewise, the query “?P @P” will retrieve corpus-attested elaborations of stereotypes in @P to suggest similes of the form “X is as P as (?P @P)”. The simile “as cold as a fish” can thus be elaborated to yield “as cold as a wet fish”, “as cold as a dead haddock”, “as cold as a wet January”, “as cold as a frozen corpse”, and “as cold as a heartless robot”. Complex queries can retrieve more elaborate vehicles for creative similes, so “@P @P” (which retrieves “robot fish” and “snow storm” for cold) and “?P @P @P” (e.g. “creamy chocolate mousse” for rich) each retrieve ngrams that blend two different but overlapping stereotypes. By using the antonym operator -, ironic similes can also be generated for the P-ness of a topic X using the pattern “X is as P as (@-P ∩ ??X)”. In effect, (@-P ∩ ??X) finds counter-examples of P-ness that are comparable to X, rather than the stereotypical examples that we expect in the normative, non-ironic use of as-similes. However, adjectives can be ambiguous, and –P may not be an appropriate antonym for the intended sense of P. Modeling Stereotypical Behavior The effectiveness of CIR depends crucially on the matching ability of its operators @, ?, and ^, and on the range of @ in particular, since only this operator can match properties (adjectives) to stereotypes (nouns) and vice versa (?, and ^ each match words that reside in the same syntactic categories). The knowledge provided by @ is a model of common-sense stereotypical norms, which is precisely the kind of knowledge that is subverted by irony. However, the simile pattern “as X as Y” that Veale & Hao first used to harvest stereotypical associations from the web is limited to the retrieval of adjectival properties. It cannot harvest stereotypical behaviors, such as the fact that babies drool and cry, that politicians lie, or that dogs bark. To rectify this limitation and to simultaneously enlarge the stereotypical model underpinning CIR, we use a bottom-up, evidence-driven approach to identifying stereotypical properties and behaviors for a given topic. We first harvest all 3-word phrases from the Google 3grams that match the pattern . Here PROPERTY can match either an adjective in WordNet or an inflected verb that denotes a behavior, such as swaggering or armored. For adjectival cases, we generate the as-simile “as ADJ as a NOUN”, while for verbal behaviors we generate the like-simile “BEHAVIOR like a NOUN”. We then use the resulting simile as a web query to see how many times the simile occurs on the web. That is, as-similes are generated for adjectival properties and like-similes are generated for verbal behaviors, so e.g. the 3-gram “a reckless cowboy” yields the query “as reckless as a cowboy” and the 3-gram “a swaggering cowboy” yields the query “swaggering like a cowboy”. We next consider only the queries that return a non-zero result set; these correspond to stereotypical associations that have been attested by usage on the web. Nonetheless, this set still contains a great deal of noise, and we find many under-specified behaviors such as “walking like a drunk” or “talking like a baby”. We could use statistical techniques here to separate the most discriminating behaviors (e.g., “staggering like a drunk” and “babbling like a baby”) from the least informative (e.g., “walking like a sailor”). Yet, since this task only needs to be done once to create a precise and highly reusable lexico-conceptual resource, we do it manually to achieve the best results. It takes a matter of weeks to perform manual filtering, but the stereotype model that results from this effort is significantly larger than that produced by Veale & Hao (2007a). It contains 9,479 different stereotypes overall, and ascribes to each of these a selection of 7,898 different properties and behaviors. In all, the new resource contains over 75,000 unique noun-to-property associations. This compares very favorably with the 12,000+ associations in Veale & Hao’s original resource. Consider the term baby: the new resource suggests these 163 associations as being stereotypical of babies: {delicate, squalling, weeping, baptized, adopted, startled, attentive, blessed, teeny, rocked, adorable, whining, bundled, toothless, placid, expected, rescued, treasured, new, sleepy, indulged, slumbering, weaned, pure, supple, helpless, small, sleeping, animated, vulnerable, wailing, cradled, kicking, soft, rested, bellowing, blameless, grinning, screaming, orphaned, mute, cherished, reliant, thriving, loveable, guileless, sniveling, inexperienced, harmless, dribbling, unthreatening, nursed, angelic, bawling, beaming, tame, naked, spoiled, scared, weak, squirming, blubbering, contented, smiling, wiggling, mewling, blubbing, sniffling, overtired, dimpled, loving, dear, tired, powerless, bewildered, peaceful, distressed, naive, wee, soiled, sucking, fussy, gurgling, vaccinated, heartwarming, pouting, constipated, drooling, quiet, wiggly, lovable, bare, weaning, suckling, cute, bald, whimpering, tender, pampered, incontinent, fleshy, charming, dependent, artless, fussing, flabby, babbling, warm, giddy, crawling, snoozing, hairless, cuddled, sweet, sobbing, squealing, wrapped, tiny, cooing, swaddled, laughing, toddling, fragile, innocent, moaning, gentle, terrified, precious, cranky, giggling, confused, pink, cuddly, fat, ignorant, snoring, young, howling, screeching, shrieking, trusting, shivering, napping, resting, frightened, fresh, loved, demanding, chubby, adored, appealing, happy, relaxed, bumbling, wriggly, rocking, wriggling, conceived, clean, content, smooth, crying, submissive} Strategies of Ironic Subversion The use of antonym-based counter-examples is just one strategy that is available to the ironic speaker. We can express this strategy Santonym in CIR terms as follows: Santonym(P) ß ?-P @-P In other words, a set of ironic expressions for a property P can be found by retrieving all 2-gram phrases where the first word is an adjective that reinforces the idea of –P (any antonym of P) and the second word is a noun denoting a concept for which any element of –P is stereotypical. Thus, for example, the phrase “soggy pillow” is retrieved for the property hard, since soft & hard are antonyms in WordNet. Since we so often seek to impress with irony, our goal is not merely to communicate an implicit negation, but to communicate an implicit negation in the most imaginative, memorable and quotable words we can muster. A vivid juxtaposition of ideas can help us to achieve this goal. We can thus use the following variant of the Santonym strategy: Scombo(P) ß @-P @-P For example, since a wall is typically hard, and a good basis for ironic descriptions of softness, the phrases “brick wall”, “stone wall”, “steel wall”, “titanium wall”, “oak wall”, “granite wall” etc. are retrieved by Scombo(soft). Likewise, phrases like “marshmallow bunny” and “snow baby”– which can seem decidedly odd, and thus fresh and imaginative when considered out of their original context – are retrieved as ironic descriptions of hardness. In this way, strategies like Scombo embody Fishlov’s (1992) view of poetic similes, by providing more elaborate and more vivid mental images than a single stereotype alone could do. A computer that uses a database of readymade phrases to suggest possible word/idea combinations for creative descriptions is thus tapping into the collective imagination of many different speakers at once. We use the term “strategy” to denote a high-level approach to generating ironic descriptions, where each such approach is represented as a CIR query for retrieving matching phrases from a corpus such as the Google ngrams. For instance, we can define another strategy Sgroup that is defined via the following CIR query: Sgroup(P) ß (^group ∩ @P) “of ” @-P The CIR query for Sgroup matches any 3-gram phrase in which: the first word denotes both a group (like family or army) and a concept for which the given property P is stereotypical; the second word is the string literal “of”; and the third word is a noun that denotes a concept for which the antonym of the given property P is stereotypical. Under normal circumstances, any word matching (^group ∩ @P) would serve as an excellent example of P-ness, denoting a mass of P stereotypes in a coherent grouping. However, Sgroup finds phrases that subvert this stereotypical group, by populating the group with counter-examples of P-ness. Thus, “army of dreamers”, “army of civilians” and “army of irregulars” are all retrieved from the Google 3-grams as ironic vehicles for the property disciplined, while the 3grams “army of cowards”, “army of babies”, “army of ants”, “army of cripples”, “army of kittens”, “army of girls” and “army of worms” are retrieved for strong. Notice the way in which Sgroup creates descriptions that suggest P-ness right up to the last word, whereupon a final ironic reversal of meaning is delivered. Sgroup can thus be viewed as a more sophisticated ironic version of the adolescent strategy for generating sarcasm, where “Not!” is placed at the end of an otherwise affirmative utterance. To use the language of Attardo and Raskin’s GTVH, each strategy is effectively a logical mechanism that causes an incongruous combination of ideas to be juxtaposed, and to be subsequently resolved as an attempt at irony. In Santonym and Sgroup the incongruity derives from lexical knowledge of antonyms and the properties they denote. For example, consider the following strategy, dubbed Smaterial Smaterial(P) ß (^material ∩ @-P) @P The category ^material is populated with elements of the WordNet category ^substance for which 3-gram matches can be found for the CIR query “made of ^substance”, such as “chocolate”, “concrete”, “steel”, etc. Thus, though brick is a stereotypically hard object, Smaterial(hard) is populated with the 2-gram phrases “silk brick”, “snow brick” and “water brick”. To provide a garden-path effect, each of these vehicles can be re-formulated so that the ironic kicker is found at the end of the phrase, as in “a brick made of silk”, “a brick made of snow” and “a brick made of water”. As shown in Santonym, Scombo, Sgroup and Smaterial, each CIR query acts as a high-level rule in which the key elements are matched non-literally against a large corpus of phrasal n-grams. But, as in all high-level rules, each rule may give rise to unexpected, low-level exceptions. The more generic the strategy, the more exceptions it brooks. Consider this somewhat under-constrained strategy: Sproxy(P) ß ?-P @?-P This strategy is a variant of Santonym that uses neighbors of antonyms of P (that is, ?-P) rather than just antonyms of P, and evokes stereotypes of those neighborly oppositions (via @?-P). For instance, this strategy retrieves the 2-gram “meek lamb” as an ironic vehicle for the property savage. However, given the wide range of neighbors in ?-P, this strategy will significantly over-generate. Though it is well meaning, we can expect just a small proportion of the phrases suggested by Sproxy to achieve an ironic effect. Moreover, words in unexpected combinations often have specific nuances that we cannot predict from high-level rules alone, and specific combinations may not give rise to the predicted ironic effect. To address this problem, highlevel strategies – such as Sproxy and others considered above – should be implemented via lower-level tactics that are grounded in attested examples of irony, and which specifically employ words that have a proven record of communicating an ironic meaning.