Less Rhyme, More Reason: Knowledge-based Poetry Generation with Feeling, Insight and Wit

Linguistic creativity is a marriage of form and content in which each works together to convey our meanings with concision, resonance and wit. Though form clearly influences and shapes our content, the most deft formal trickery cannot compensate for a lack of real insight. Before computers can be truly creative with language, we must first imbue them with the ability to formulate meanings that are worthy of creative expression. This is especially true of computer-generated poetry. If readers are to recognize a poetic turn-of-phrase as more than a superficial manipulation of words, they must perceive and connect with the meanings and the intent behind the words. So it is not enough for a computer to merely generate poem-shaped texts; poems must be driven by conceits that build an affective worldview. This paper describes a conceit-driven approach to computational poetry, in which metaphors and blends are generated for a given topic and affective slant. Subtle inferences drawn from these metaphors and blends can then drive the process of poetry generation. In the same vein, we consider the problem of generating witty insights from the banal truisms of common-sense knowledge bases. Ode to a Keatsian Turn Poetic licence is much more than a licence to frill. Indeed, it is not so much a licence as a contract, one that allows a speaker to subvert the norms of both language and nature in exchange for communicating real insights about some relevant state of affairs. Of course, poetry has norms and conventions of its own, and these lend poems a range of recognizably “poetic” formal characteristics. When used effectively, formal devices such as alliteration, rhyme and cadence can mold our meanings into resonant and incisive forms. However, even the most poetic devices are just empty frills when used only to disguise the absence of real insight. Computer models of poem generation must model more than the frills of poetry, and must instead make these formal devices serve the larger goal of meaning creation. Nonetheless, is often said that we “eat with our eyes”, so that the stylish presentation of food can subtly influence our sense of taste. So it is with poetry: a pleasing form can do more than enhance our recall and comprehension of a meaning – it can also suggest a lasting and profound truth. Experiments by McGlone & Tofighbakhsh (1999, 2000) lend empirical support to this so-called Keats heuristic, the intuitive belief – named for Keats’ memorable line “Beauty is truth, truth beauty” – that a meaning which is rendered in an aesthetically-pleasing form is much more likely to be perceived as truthful than if it is rendered in a less poetic form. McGlone & Tofighbakhsh demonstrated this effect by searching a book of proverbs for uncommon aphorisms with internal rhyme – such as “woes unite foes” – and by using synonym substitution to generate non-rhyming (and thus less poetic) variants such as “troubles unite enemies”. While no significant differences were observed in subjects’ ease of comprehension for rhyming/non-rhyming forms, subjects did show a marked tendency to view the rhyming variants as more truthful expressions of the human condition than the corresponding non-rhyming forms. So a well-polished poetic form can lend even a modestly interesting observation the lustre of a profound insight. An automated approach to poetry generation can exploit this symbiosis of form and content in a number of useful ways. It might harvest interesting perspectives on a given topic from a text corpus, or it might search its stores of commonsense knowledge for modest insights to render in immodest poetic forms. We describe here a system that combines both of these approaches for meaningful poetry generation. As shown in the sections to follow, this system – named Stereotrope – uses corpus analysis to generate affective metaphors for a topic on which it is asked to wax poetic. Stereotrope can be asked to view a topic from a particular affective stance (e.g., view love negatively) or to elaborate on a familiar metaphor (e.g. love is a prison). In doing so, Stereotrope takes account of the feelings that different metaphors are likely to engender in an audience. These metaphors are further integrated to yield tight conceptual blends, which may in turn highlight emergent nuances of a viewpoint that are worthy of poetic expression (see Lakoff and Turner, 1989). Stereotrope uses a knowledge-base of conceptual norms to anchor its understanding of these metaphors and blends. While these norms are the stuff of banal clichés and stereotypes, such as that dogs chase cats and cops eat donuts. we also show how Stereotrope finds and exploits corpus evidence to recast these banalities as witty, incisive and poetic insights. Proceedings of the Fourth International Conference on Computational Creativity 2013 152 Mutual Knowledge: Norms and Stereotypes Samuel Johnson opined that “Knowledge is of two kinds. We know a subject ourselves, or we know where we can find information upon it.” Traditional approaches to the modelling of metaphor and other figurative devices have typically sought to imbue computers with the former (Fass, 1997). More recently, however, the latter kind has gained traction, with the use of the Web and text corpora to source large amounts of shallow knowledge as it is needed (e.g., Veale & Hao 2007a,b; Shutova 2010; Veale & Li, 2011). But the kind of knowledge demanded by knowledgehungry phenomena such as metaphor and blending is very different to the specialist “book” knowledge so beloved of Johnson. These demand knowledge of the quotidian world that we all tacitly share but rarely articulate in words, not even in the thoughtful definitions of Johnson’s dictionary. Similes open a rare window onto our shared expectations of the world. Thus, the as-as-similes “as hot as an oven”, “as dry as sand” and “as tough as leather” illuminate the expected properties of these objects, while the like-similes “crying like a baby”, “singing like an angel” and “swearing like a sailor” reflect intuitons of how these familiar entities are tacitly expected to behave. Veale & Hao (2007a,b) thus harvest large numbers of as-as-similes from the Web to build a rich stereotypical model of familiar ideas and their salient properties, while Özbal & Stock (2012) apply a similar approach on a smaller scale using Google’s query completion service. Fishelov (1992) argues convincingly that poetic and non-poetic similes are crafted from the same words and ideas. Poetic conceits use familiar ideas in non-obvious combinations, often with the aim of creating semantic tension. The simile-based model used here thus harvests almost 10,000 familiar stereotypes (drawing on a range of ~8,000 features) from both as-as and like-similes. Poems construct affective conceits, but as shown in Veale (2012b), the features of a stereotype can be affectively partitioned as needed into distinct pleasant and unpleasant perspectives. We are thus confident that a stereotype-based model of common-sense knowledge is equal to the task of generating and elaborating affective conceits for a poem. A stereotype-based model of common-sense knowledge requires both features and relations, with the latter showing how stereotypes relate to each other. It is not enough then to know that cops are tough and gritty, or that donuts are sweet and soft; our stereotypes of each should include the cliché that cops eat donuts, just as dogs chew bones and cats cough up furballs. Following Veale & Li (2011), we acquire inter-stereotype relationships from the Web, not by mining similes but by mining questions. As in Özbal & Stock (2012), we target query completions from a popular search service (Google), which offers a smaller, public proxy for a larger, zealously-guarded search query log. We harvest questions of the form “Why do Xs Ys”, and assume that since each relationship is presupposed by the question (so “why do bikers wear leathers” presupposes that everyone knows that bikers wear leathers), the triple of subject/relation/object captures a widely-held norm. In this way we harvest over 40,000 such norms from the Web. Generating Metaphors, N-Gram Style! The Google n-grams (Brants & Franz, 2006) is a rich source of popular metaphors of the form Target is Source, such as “politicians are crooks”, “Apple is a cult”, “racism is a disease” and “Steve Jobs is a god”. Let src(T) denote the set of stereotypes that are commonly used to describe a topic T, where commonality is defined as the presence of the corresponding metaphor in the Google n-grams. To find metaphors for proper-named entities, we also analyse n-grams of the form stereotype First [Middle] Last, such as “tyrant Adolf Hitler” and “boss Bill Gates”. Thus, e.g.: src(racism) = {problem, disease, joke, sin, poison, crime, ideology, weapon} src(Hitler) = {monster, criminal, tyrant, idiot, madman, vegetarian, racist, ...} Let typical(T) denote the set of properties and behaviors harvested for T from Web similes (see previous section), and let srcTypical(T) denote the aggregate set of properties and behaviors ascribable to T via the metaphors in src(T): (1) srcTypical (T) = M∈src(T) typical(M) We can generate conceits for a topic T by considering not just obvious metaphors for T, but metaphors of metaphors: (2) conceits(T) = src(T) ∪ M∈src(T) src(M) The features evoked by the conceit T as M are given by: (3) salient (T,M) = [srcTypical(T) ∪ typical(T)]