Computational generation of limericks
French Studies, Queen's University, Canada
School of Computing, Queen's University, Canada
Keywords: natural language generation, humour
A program called VINCI from Queen's Employs computational means To make metre and rhyme In right to left time So limericks appear on your screens. The people attending this session We hope will have gained the impression That these powerful tools, With appropriate rules, Can give rise to poetic expression!
There is evidence that the generation of at least some classes of humour is a rule-governed activity, formalisable in terms of algorithms, and implementable computationally in order to generate actual humorous utterances. In earlier work, using the VINCI natural language generation environment, we have shown this to be the case for several subclasses of verbal humour, including Tom Swifties and several classes of riddles. (See Lessard & Levison, 1992, 1993, 1995, 1997). Others have done similar work (see for example Binsted & Ritchie, 1997). This does not mean that we suppose the specific algorithms used in generation to exist as internal mental representations in speakers, but only that the productions observed in humans admit of formalisation.
It must be admitted, however, that puns and riddles form instances of what is often called verbal humour. This is typically non-narrative, based on relatively simple paradigmatic lexical relations, and usually evaluated in terms of cleverness rather than funniness. (See Lessard, Levison & Venour, 2002, for a discussion of the distinction.) Instances of verbal humour are in many respects inherently simpler than more narrative structures like jokes. Limericks, on the other hand, presuppose at least a primitive narrative structure. As such, they provide a useful proving ground for more advanced work on both language generation (since it is necessary to take account of textual coherence) and humour generation (since narrative structure is involved).
Limericks may be characterised generally as five-line verses with the aabba rhyme scheme, one of a range of metrical foot structures, and some attempt at humour or at least cleverness. (For discussion, see Bibby, 1978). A typical example will illustrate the model:
There was an old man from Peru Who dreamt he was eating his shoe He awoke with a fright In the midst of the night And found it was perfectly true.
Taken as a subset of more general poetic language, limericks confront us with a number of practical problems, but also with a number of theoretical challenges when they are compared with prose-based natural language generation. For example, traditional approaches to text planning and generation tend to be driven by conceptual and syntactic rather than metrical structure. This is unsurprising when one considers that the object of most systems of the sort is to represent some encyclopedic or data-base-like description of some state of affairs in a discursive format. In such a context, the primary factors to be considered include the model of the user (see for example Paris, 1993), the overall structure of the text plan (see for example McKeown, 1985), paragraph-level phenomena (see for example Mann, 2002), and problems of reference (see for example Dale, 1992).
As a consequence, planning tends to be top-down (from universe of discourse, to dialogue model, to context, to overall text, to paragraph, to sentence) and many of the phenomena to be dealt with (anaphora, for example) can be examined in a left-to-right sequence. This has meant that conceptual tools such as phrase structure grammars or their variants have provided sufficient power to deal with the problems raised. Even phenomena such as freer word order may be captured in such systems by means such as the splitting apart of linear precedence and immediate dominance as in formalisms like Generalized Phrase Structure Grammar (Gazdar et alii, 1985).
The production of poetic text, on the other hand, requires that content be filtered by metrical and phonological factors such as the number of syllables, stress, and rhyme. A particularly interesting consequence of this fact is that generation must take account of the ends of lines in the production of the beginning and middle. For example, in the case of the limerick quoted above, we may represent the syllable structure as follows, where s represents an unstressed and S a stressed syllable:
There was an old man from Pe ru Who dreamt he was ea ting his shoe He a woke with a fright In the midst of the night And found it was per fect ly true s S s s S s s S
It can be seen that the first, second and fifth line contain an iamb followed by two anapests, while the third and fourth lines contain two anapests. In addition, rhyming lines (Peru-shoe-true) must end on the same stressed syllable nucleus and coda (to use the terminology of metrical phonology (see for example Hogg, 1987). Note that these are not necessarily lexical items: in the first, second and fifth lines, the sequences "his shoe" and "-ly true" rhyme with "Peru".
A full analysis of the semantic, syntactic, lexical and phonological relations found in the limerick would take more space than is available here. However, from the generative perspective, it is possible to imagine three mechanisms to satisfy this constellation of constraints:
1) exhaustive generation: in essence, if we define a set of templates, we can generate large numbers of instances of potential limericks and ask a human to select those few judged to be of good quality. (This is the monkeys typing Shakespeare model, or Borges' library.)
2) initial generation followed by tweaking: we may also begin by producing a sequence of lines and then selectively edit these to approach the target of an acceptable limerick, by means of the addition, removal or shifting of items. This may be how human poets work: it is how we produced the limericks at the head of this abstract. (ValÚry is said to have claimed that the first line of a poem was a gift of the gods and that the remainder was the product of the poet's craft and efforts.) Since the edit operations described above are well-known in the computational context, it is not impossible to imagine their implementation.
3) right-to-left generation: given the direction of the constraints in limericks, a possible computational model involves the selection of the right-most elements of a sequence of lines as a starting point and then within each line, the constraint-satisfaction generation of the intervening elements required to flesh out the entire limerick.
In fact, these right-to-left formal constraints must interact with top-down constraints which govern the semantic coherence of the text. In other words, in order for the limerick to be coherent, the narrative events which it includes must be consistent with the subject. These thematic constraints may be fairly loose, as in the Old Man from Peru poem, where there is no particular logic which attaches old men from that part of South America and the ingestion of footwear (although more generally, the poem respects the constraint that old men are humans and that humans eat food which is a subset of tangible objects, and that footwear is also a subset of tangible objects). On the other hand, they may be quite tight, as in the VINCI limericks, which capture a set of characteristics of the software (its origin, the architecture of the program). Compare as well Bibby's attempts to produce limericks appropriate to each of the Cambridge colleges.
The limerick thus represents the semantic expansion from one or more initial lexical choices, by means of the selection of traits or actions which are at least consistent with these. This presupposes a rich representation of the characteristics of each lexical item, including both high-level semantic traits such as that between humans, eating and objects, but also more encyclopedic information, such as the fact that VINCI is a program produced at Queen's. (See Peeters, 2000 for a discussion of some of the problems found at this 'lexicon-encyclopedia interface'). Note also that a sufficiently rich set of traits will provide the seed for variant limericks based on the same starting point. Consider for example the following example by Edward Lear:
There was an Old Man of Peru, Who watched his wife making a stew; But once by mistake, In a stove she did bake, That unfortunate Man of Peru.
At the formal level, problems arise both between lines (rhyme) and within lines (metrical structure). Consider the latter in the context of the initial old man from Peru poem. Let us assume that we have determined to use a metrical structure based on an iamb and two anapests and that we have selected the the place-name Peru. At this point, we have used up two of the eight available syllables. We also know that the next earliest syllable must be unaccented. The solution found here is "from". At this point, we know that we have a prepositional phrase "from Peru" which requires (among other choices) a noun phrase with the appropriate accent structure. This constraint is satisfied here by "an old man"). However, in order to verify that this noun phrase is appropriate, we must have knowledge of its metrical structure. For example, if the target were an iamb, we would need to seek, among other possibilities, a DET N structure. This application of constraints cascades right to left until the overall metrical target is met, within the semantic constraints already specified.
In the conference presentation, we will illustrate the implementation of the approach described above using the VINCI natural language generation environment (see http://www.cs.queensu.ca/CompLing for details). Briefly, VINCI allows for the initial specification (PRESELECTION) of a constellation of lexical items which may be constrained by semantics, morphological and syntactic characteristics, and phonology. Preselected items, as well as all their characteristics, are available to subsequent steps of the generation process. Among other things, this allows for multiple layers of preselection, in which a first item determines a set of rhyme constraints, as well as a set of semantically appropriate elements.
Within each line of the limerick, and between lines, the VINCI mechanism of GUARDED SYNTAX allows the control of subsequent steps of the overall generation to be conditioned by the framework existing at that point. (A simple example of guarded syntax: if a noun phrase node carries the attributes 'first person' or 'second person', then its children may only be instantiated by a pronoun, whereas if the parent node carries the attribute 'third person', then the children may be pronouns, full noun phrases, or proper names.) Applied to the production of a limerick, and assuming right-to-left generation, if the parent node of a tree contains a prepositional phrase which sums to an anapest, then the next left-most chunk of syntax must inherit this information and construct an appropriate item from a library of possible patterns.
One consequence of this model is that syntax is reduced from a single overarching tree to the sum of a number of possible micro-trees. Such an approach is comparable to the model of Tree Adjoining Grammars, in which insertion and development occurs at the level of lexical-based subtrees (see Joshi & Schabes, 1997). It is also somewhat similar to the Labelled Deductive System approach used in parsing by Kempson et alii, 2001. Another consequence is that by precluding tweaking and editing, the system itself is forced to deal with the interplay of linguistic levels and constraints, without human assistance, thus providing an acid test for both the model and the implementation. Of course, this leaves aside the question of evaluation, itself a thorny issue, since the criteria used are many, varied and probably fuzzy as well. In related work, we are examining the ability of humans to learn how to produce limericks, as well as their ability to evaluate them.
Bibby, Harold Cyril. (1978) The Art of the Limerick. Hamden, Conn. : Archon Books.
Binsted, Kim; Ritchie, Graeme. (1997) Computational rules for punning riddles. Humor, 10 (1), pp.25-76.
Dale, Robert. (1992) Generating referring expressions: constructing descriptions in a domain of objects and processes. Cambridge, Mass.: MIT Press.
Kempson, Ruth; Meyer-Viol, Wilfried; Gabbay, Dov (2001) Dynamic Syntax: the Flow of Language Understanding. London: Blackwell.
Gazdar, G.; Klein, E.; Pullum, G.; Sag, I. (1985) Generalized Phrase Structure Grammar. Cambridge, MA: Harvard University Press.
Hogg, Richard. (1987) Metrical phonology: a coursebook. Cambridge: Cambridge University Press.
Joshi, A.K.; Schabes, Y. (1997) Tree-Adjoining Grammars, in Handbook of Formal Languages, G. Rozenberg and A. Salomaa (eds.), Vol. 3, Berlin: Springer, pp. 69 - 124.
Lessard, Greg; Levison, Michael; Venour, Chris. (2002) Cleverness versus funniness. April Fool's Day Workshop on Computational Humour, Trento.
Lessard, Greg; Levison, Michael. (1997) Rule-governed wordplay and creativity. In Mind II: Computational Models of Creative Cognition, Dublic City University. http://www.compapp.dcu.ie/~tonyv/MIND/greg.html
Lessard, Greg; Levison, Michael. (1995) Linguistic and Cognitive Underpinnings of Verbal Humour. International Cognitive Linguistics Association Conference, Albuquerque.
Lessard, Greg; Levison, Michael. (1993) Computational Modelling of Riddling Strategies. ACH/ALLC Joint Annual Conference, Georgetown University, Washington, DC. (Extended abstract in conference proceedings, pp. 120-122.)
Lessard, Greg; Levison, Michael. (1992) Computational Modelling of Linguistic Humour: Tom Swifties. ALLC/ACH Joint Annual Conference, Christ Church, Oxford. (Extended abstract in conference proceedings, pp. 175-178.)
Mann, W. (2002) RST Web Site. http://www.sil.org/~mannb/rst/
McKeown, Kathleen. (1985) Text generation : using discourse strategies and focus constraints to generate natural language text. Cambridge: Cambridge University Press.
Paris, CÚcile. (1993) User modelling in text generation. Francis Pinter Publishers: London.
Peeters, Bert (2000) The Lexicon-encyclopedia interface. New York: Elsevier.