Computational Models of Riddling Strategies

Greg Lessard,
French Studies, Queen's University

Michael Levison,
Computing and Information Science, Queen's University

1. Performance

Currently, there is considerable debate among linguists and cognitive scientists concerning the relative autonomy of language skills. On the one hand, proponents of an autonomous linguistics claim that the language faculty is to a significant extent independent of other cognitive mechanisms. It is pointed out that the range of errors made by children learning their first language is substantially smaller than the logically possible set of errors which they might make. This is taken for evidence of innate constraints on the possible structure of language, often referred to as Universal Grammar.

On the other hand, a number of cognitive scientists see language skills as one manifestation of highly general cognitive processes, and argue that any specificity of language may be accounted for by means of low-level input factors. Anderson (1983) is a good illustration of this approach. His ACT* system has been used to model language learning and some aspects of language behaviour using the same cognitive devices used for other activities.

In fact, the opposition between the two perspectives is not as clearcut as this. Most researchers would accept that language skills represent a complex interaction of general and specific cognitive devices. For example, on the cognitive level, Karmiloff-Smith (1992) provides examples to support her hypothesis that skill acquisition occurs over staged levels, beginning with (potentially built-in) analyzed procedural skills, moving to `reanalyzed' (and hence interpretable and decomposable) skills, and finally to procedural skills which are themselves expressible by means of explanations and therefore open to cognitive manipulation. From the linguistic perspective, attention to performance issues and corpus materials, with their inherent variation, has brought to the fore the distinction between the core and the periphery of grammar.

In this context, linguistic humour has an important role to play in providing us with detailed insights into the real-time processing of language, and the actual implementation of linguistic skills in a highly constrained cognitive context.

2. Humour

In what follows, we leave aside a certain number of perspectives on humour, including the psychoanalytical (why we make jokes, or what they mean: Freud 1976) and the social (what aspects of jokes are common to particular cultures, and what aspects transcend culture). We will also restrict ourselves to one side of the traditional distinction between situational and verbal humour (see for example Raskin 1985), adopting as the criterion here that verbal humour is bound up with the choice of lexical entries and syntactic structure, and includes all cases where the use of synonymy or paraphrase would destroy the humorous force of an utterance. Thus, the following is an example of verbal humour, since orange could not be replaced by say, citrus fruit.

     Knock, knock.
     Who's there?
     Banana.
     Banana who?
     Knock, knock.
     Who's there?
     Banana.
     Banana who?
     Knock, knock.
     Who's there?
     Orange.
     Orange who?
     Orange you glad I didn't say banana?

On the other hand, the following is an example of situational humour, since it is easy to imagine a large number of variant formulations.

     Question: How many X does it take to go ice-fishing?
     Answer: Ten: one to cut the hole and nine more to push the boat in.

In what follows, we take as our initial hypothesis that verbal humour is a rule-governed phenomenon, a hypothesis for which we presented evidence in earlier studies (Lessard 1988, Lessard and Levison 1992). The object of our approach is to model the rule systems which underly particular forms of humour. Furthermore, we consider that purely intuitive modelling of linguistic humour should be complemented by actual implementation in a computational environment. In other words, the computer must be capable of producing, on its own, examples of the humour type under study. Only in this manner can we be certain to have captured all components of the mechanism. At the same time, this forces us to deal with the complex role played by pragmatic, contextual and cognitive factors, which in the current stage of knowledge we are unable to model in the computer environment. It goes without saying as well that the requirements of modelling linguistic humour provide an acid test for a natural language generation environment.

3. Computational Environment

For our present purposes, we will use the following features present in the current version of the VINCI system (see Levison and Lessard 1992, 1995 for details of the system).

4. Previous Research

The study of riddles has tended to be cloistered in three sub-domains: children's literature, folklore studies, and the analysis of early literary texts (like the Exeter Book). At the same time, with a few exceptions to be discussed below, there has been a tendency to classify riddles by their provenance or their theme rather than by their linguistic traits.

The status of riddles has been the subject of some debate. Several points in particular merit discussion. As Pepicello and Green (1984) point out, a riddle must be seen as a two-part unit, composed of a question and an answer which are inextricably linked. Unlike the proverb, whose goal is to provide access to shared cultural knowledge, the riddle uses shared cultural and linguistic knowledge for the purpose of amusement. Rather than guiding the riddlee to wisdom, the riddler engages in play.

Riddles are by and large not meant to be solved by the riddlee, or at the least the solving is seen as an exceptional event. In other words, the search space defined by a riddle is intentionally underspecified. The pleasure of the riddle arises in the tension produced by the riddlee not finding the solution and the release produced by the riddler who gives the solution. In fact, there typically exist fixed phrases used to indicate to the riddler that the solution should be provided: I don't know . Given this, we focus not on solving riddles, but on their generation.

Such a change in perspective brings one additional benefit. From the perspective of generation, it is possible to see the riddle as turning on a particular linguistic entity capable of giving rise to two distinct series of links. Thus, in a riddle such as What has a bed but cannot sleep? A river, we note that the form bed forms the central point for two distinct series: (1) the part-whole relation involving river and bed, and (2) the verb-circumstance relation involving sleep and bed.

In earlier work on modelling Tom Swifties, Lessard (1988) proposed the term pivot to designate the central point of the humorous structure. Thus, given a Tom Swifty like I hate chemistry, Tom said acidly, the adverb acidly forms the pivot, with a semantic bridge to chemistry and a formal bridge to acid. We will adopt the same terminology here, using the term pivot for that element of the riddle around which the structure turns.

The productivity of riddle formulae has also been discussed. On the one hand, it is clear that the lexicon of a language contains a certain number of riddles known to a large number of the members of the linguistic community (or sub-community). Thus, What's black and white and read/red all over? A newspaper forms part of the basic cultural baggage of most North American speakers of English.

On the other hand, the existence of new riddle books demonstrates the capacity of speakers to generate heretofore unattested riddles. Furthermore, as we shall see below, many riddles can be seen as the product of a clearly defined set of rules, such that once the model is provided, an indefinitely large number of riddles may be created. Similarly, for any given riddle, once the rule is known, it is possible in at least some cases to provide an indefinitely large number of possible solutions. Given the riddle What is red on the outside and white on the inside?, a large number of potential answers are possible, including a lobster, a (red) book, a hand in a mitten, a gum, a cut of meat, and so on. Nevertheless, there exists at least one lexicalized answer, shared by the community: an apple. In actual usage, a riddler is within his or her rights to reject other potential answers in favour of this correct answer.

How are we to reconcile the existence of lexicalized riddles and the possibility of new forms? Maranda (1971, 136) proposes the existence of riddle-making rules, similar to grammatical rules, which would make it possible to predict all possible riddles in a culture. The unattested riddles formed on the basis of such rules would nevertheless be considered on the same footing as attested riddles. Pepicello and Green (1984, 83-4) criticize this proposal, pointing out that by definition, riddles exist as linked questions and answers, with the two forming a single unit.

In our opinion, both points of view miss the essential point that lexicalization and rule-governedness are both necessary for full-fledged riddles. On the one hand, only the existence of a predictable and rule- governed relation between question and answer allows the release of tension mentioned earlier, the sense of Why didn't I think of that?. An arbitrarily chosen answer to the riddle is not acceptable, as Pepicello and Green point out themselves (1984, 88) in excluding from the class of riddles such forms as Why did the elephant paint his toenails red? So he could hide in a cherry tree. At the same time, the fact that some riddles become lexicalized cannot be ignored. The difficulty is essentially the same faced by morphologists studying productivity and lexicalisation.

In what follows, we shall impose the following constraint. Rules for riddle-formation will be proposed only for those cases where multiple riddles are attested for the same framework. And in all cases, potential or newly generated riddles will be distinguished from attested riddles by the + symbol.

5. Corpus-Modelling Approach

In an attempt to isolate several representative mechanisms of riddling, we examined a selection of children's books devoted to the subject. We chose children's books in an attempt to find the simplest basic framework. Clearly, had we chosen more literary models, other problems would have come to light. Similarly, we limited the analysis to texts produced within the English language cultural paradigm, within the past 20 years.

On the basis of the corpus, we have isolated three distinct types of riddle structure. In each case, we will provide a corpus-based description of the structure isolated, together with examples. We have put into machine-readable form the data from the sources used.

6. Lexical Relation Riddles

The first class of riddles hinges on lexical relations between words, including synonymy, homonymy and meronymy. We will illustrate riddles of the sort using two distinct types: those based on the simultaneous presence of two homonyms, and those based on the presence of only one homonym.

6.1 Syntagmatic Homonym Riddles

Terban (1982) provides a collection of riddles whose solution is based on the repetition of two homonymous words. We will begin with these, since they have the advantage of putting explicitly into play all elements of the riddle.

6.1.1 Corpus

     Riddle: What do you call a less expensive bird?
     Solution: A cheaper cheeper.

     Riddle: What are groups of sailors on an ocean pleasure trip?
     Solution: Cruise crews.

     Riddle: What does the man who looks at oceans do all day?
     Solution: Sees seas.

     Riddle: Two ran the race but only...
     Solution: One won.

6.1.2 Analysis

Syntactic frames available for riddling include relatively frequent fixed forms such as What do you call..., What are..., but also relatively free forms, as in the last two examples.

Solutions present a wide range of syntactic diversity, including ADJ - N (cheaper cheeper), N - N (cruise crews), V - N ( sees seas) and PRON - V (one won). On the morphological level, we note that solutions may hinge not only on the abstract underlying form, but also on the final inflected form of some lexical items (won rather than win, crews rather than crew for example).

Finally, the relation between the riddle clues and the solution may be based on synonymy cheeper - bird, paraphrase less expensive -cheaper, or cyclicity two - one.

Examples such as those given above present the peculiarity of a `double pivot'. Thus, in the first example, cheeper possesses a link of synonymy with bird and a link of homonymy with cheaper, which in turn is linked by paraphrase with less expensive.

6.1.3 Modelling

In what follows, we shall model the structure ADJ N found above. A sense of the range of candidate forms can be established by means of an AWK search of the CUVOALD machine readable dictionary of English (Mitton 1992). Thus, the list of homonymous adjectives and nouns contains items such as the following (excluding cases where the same written form may function as either adjective or noun).

     attic | '&tIk | K6% | 2 |
     Attic | '&tIk | OA$ | 2 |
     earnest | '3nIst | L@%,OA% | 2 |
     Ernest | '3nIst | Nl% | 2 |
     Artie | 'AtI | Nl% | 2 |
     arty | 'AtI | OA% | 2 |
     oral | 'Or@l | K6$,OA% | 2 |
     aural | 'Or@l | OA% | 2 |
     shire | 'SaI@R | K6% | 1 |
     shyer | 'SaI@R | Or% | 2 |
     idle | 'aIdl | J2%,OB% | 22A,2C,15B |
     idol | 'aIdl | K6% | 2 |
     hour | 'aU@R | K6% | 2 |
     our | 'aU@R | OA* | 2 |
     baron | 'b&r@n | K6% | 2 |
     barren | 'b&r@n | OA% | 2 |
     boulder | 'b@Uld@R | K6% | 2 |
     bolder | 'b@Uld@R | Or% | 2 |
     bearer | 'be@r@R | K6% | 2 |
     barer | 'be@r@R | Or$ | 2 |
     better | 'bet@R | H0%,K6%,Or*,Pu* | 26A |
     bettor | 'bet@R | K6$ | 2 |
     bridle | 'braIdl | J2%,K6% | 22A,2C,6A |
     bridal | 'braIdl | K6$,Oq% | 2 |
     etc.

A computational model requires, however, additional data as well, in order to link each pivot with formally and semantically related items. Let us assume that we have lexical entries in the form of database records, which contain (in field 13) lexical pointers to adjectival homonyms (tagged by `homa') and synonyms (tagged by `syn'). To take a typical example, given the lexical entries bear and bare, the first will contain a lexical pointer to the second, which is its adjectival homonym, and also to the synonym bruin while the second will contain a lexical pointer to the synonym naked.

Then, within the VINCI system, we may write the syntax:

     ROOT = N[animate]/13=homa/13=syn
     %

     RIDDLE = TRANSFORMATION
     N : FRAME[name] 1/@13:homa/@13:syn 1/@13:syn PUNCT[question] ;
     %

     SOLUTION = TRANSFORMATION
     N : DET[indef] 1/@13:homa 1 ;
     %

     QUESTION = RIDDLE:ROOT
     %

     ANSWER = SOLUTION:ROOT
     %

The tree ROOT, in effect, selects any animate noun having both an adjectival homonym and a synonym (say bear). The transformation RIDDLE yields a question containing a synonym of the homonym (bear bare - naked) and a synonym of the noun (bear - bruin). These are inserted into a FRAME (basically, a lexical entry of the form What do you call a) and followed by a question mark. The transformation SOLUTION yields the homonym of the noun (bear -bare) and the noun itself (bear). Thus a typical result is:

     +Question : What do you call a naked bruin ?
     +Answer : a bare bear

6.2 Paradigmatic Homonym Riddles

Beisner (1983) provides a list of riddles based on the production of a list of clues, with the solution hinging on the identification of the referent which includes the characteristics provided. In many of the examples, the difficulty of the riddle hinges on a homonymous relation implicit in one or more of the clues, as in the following.

6.2.1 Corpus

     Q: What is it: / Has a mouth and does not speak, / Has a bed and does
        not sleep.
     A: River.

     Q: It has four legs and a foot / And can't walk. / It has a head and
        can't talk.
     A: Bed.

     Q: What has teeth and can't bite?
     A: Comb.

     Q: What is it / That has teeth / And can't eat?
     A: Rake

     Q: Four fingers and a thumb, / Yet flesh and blood I have none.
     A: Gloves.

6.2.2 Analysis

There are strong similarities between this class and the first one, in that both hinge on implicit lexical and semantic links. We must have:

In the examples presented above, thematic unity is provided by the trait `bodypart', shared by lexical items like mouth, teeth, legs, fingers etc. In what follows, we will attempt to model this subclass.

6.2.3 Modelling

As was the case earlier, modelling requires the traversal of a series of lexical pointers, in which each lexical item is linked to another. Here we assume that lexical entries contain (again in field 13) pointers to a noun homonym (tagged by `homn'), a holonym, that is, a larger entity of which another is a part (tagged by `hol') and a typical role filled by the the noun (tagged by `rol').

To take a typical example, the ROOT may select a noun like eye, which names a body part but which also has a noun homonym (eye of a potato) and a role (the ability of see, tagged by `rol'). The RIDDLE tranformation produces a FRAME having the form What has a followed by the noun homonym eye and then by another FRAME ( but can't) and then by the role played by the eye (see). The SOLUTION tranformation produces an indefinite article followed by a holonym of the noun homonym (potato), as follows:

     ROOT = N[bodypart]/13=homn/13=rol
     %

     RIDDLE = TRANSFORMATION
     N : FRAME[has] 1/@13:homn FRAME[cant] 1/@13:rol PUNCT[question] ;
     %

     SOLUTION = TRANSFORMATION
     N : DET[indef] 1/@13:homn/@13:hol ;
     %

     QUESTION = RIDDLE:ROOT
     %

     ANSWER = SOLUTION:ROOT
     %

On the basis of a lexicon containing items such as eye, see, potato, tongue, talk, shoe, this syntax produces appropriate riddles, as the following examples illustrate.

     +Question : What has an eye and can't see ?
     +Answer : a potato

     +Question : What has a tongue and can't talk ?
     +Answer : a shoe

7. Substring Riddles

Another set of riddles is based, not on existing lexical relations, but on purely formal string relations, either total or partial. We shall see that examples of the sort, which rely on a degree of processing which goes beyond existing lexical relations, make special demands of a generative environment.

7.1 Perfect Substring Riddles

We will begin with examples drawn from Most (1991) to illustrate the principle.

7.1.1 Corpus

     Q: Why did the musician find a dog in his trumpet?
     A: Because he always finds a pet in his trumpet.

     Q: What can you find thousands of at the beach?
     A: You can find thousands of grains of sand.

     Q: What kind of ghost invites you into its haunted house?
     A: A ghost that is a good host.

     Q: Why were the mice invited to the picnic?
     A: Because the mice brought the ice.

     Q: What happens when a greedy gorilla eats too many bananas?
     A: The greedy gorilla gets ill.

     Q: Why is the dragon trying to lift its tail off the ground?
     A: The dragon doesn't want to drag it around anymore.

7.1.2 Analysis

The linguistic relation underlying these examples is a simple orthographic substring relation. The substring does not necessarily correspond to an accented syllable of the superstring (see the first two examples), nor to the pronunciation of the superstring (see the second and third). Note also that the superstring found may be morphologically complex (plural, etc.), as in the case of mice. Substrings may be either nouns, adjectives or verbs.

The embedding riddle sentence can be syntactically quite complex, including anaphora (ghost and its) and gapping (What ... thousands of).

The pivot in all these riddles is the word shown in bold in the answer. Semantically, the embedding riddle sentence represents a projection of the pivot. Thus, pet gives rise to a sentence containing its hyponym dog, host gives rise to its corresponding verb invites, and sand gives rise to the metonymically related beach. The relation between ill and eats too many bananas is one of causality, while the relation between drag and lift is based on antonymy. In some cases, the relation is cognitively complex, as when ice gives rise to a sentence having to do with picnics, where ice is presumably used.

7.1.3 Modelling

Substring riddles such as those shown above form a potentially open set, in that any lexical item may form a pivot. In order to illustrate this, we chose the lexical item cat as a pivot. We then used AWK to search the CUVOALD in order to find singular nouns (code K6) containing the phonetic string [k&t] and the graphical string cat, excluding (by hand) items having a semantic link with the noun cat itself. The results are as follows:

     Magnificat | m&g'nIfIk&t | K6% | 4 |
     cataclysm | 'k&t@klIz@m | K6% | 4 |
     catafalque | 'k&t@f&lk | K6$ | 3 |
     cataleptic | ,k&t@'leptIk | K6$,OA$ | 4 |
     catalogue | 'k&t@l0g | H2%,K6% | 36A |
     catalyst | 'k&t@lIst | K6% | 3 |
     catamaran | ,k&t@m@'r&n | K6% | 4 |
     catapult | 'k&t@pVlt | H0%,K6% | 3 |
     cataract | 'k&t@r&kt | K6% | 3 |
     catcall | 'k&tkAl | I0%,K6% | 2 |
     caterpillar | 'k&t@pIl@R | K6% | 4 |
     caterwaul | 'k&t@wOl | I0%,K6% | 3 |
     scatter | 'sk&t@R | J0%,K6% | 22A,2C,6A,15A,15B |
     scatterbrain | 'sk&t@breIn | K6% | 3 |

Construction of an appropriate riddle in the VINCI formalism then requires the following steps. We begin by selecting a ROOT composed of an animate noun (cat in our example) having both a hyperonym ( animal) and a superstring (catamaran for example. The RIDDLE transformation produces a FRAME (What kind of) followed by the hyperonym (animal) and a verb which takes the its subject specification from the noun (thus 1!Nounsem.subj). Thus an animate noun will impose the choice of a verb selecting an animate subject. This is followed by an indefinite article, the superstring of the base noun (catamaran) and a question mark. The SOLUTION tranformation presents an indefinite article followed by the base noun.

     ROOT = N[animate]/13=hpr/13=spr
     %

     RIDDLE = TRANSFORMATION
     N : FRAME[kind] 1/@13:hpr V[1!Nounsem.subj] DET[indef] 1/@13:spr
         PUNCT[question] ;
     %

     SOLUTION = TRANSFORMATION
     N : DET[indef] 1 ;
     %

     QUESTION = RIDDLE:ROOT
     %

     ANSWER = SOLUTION:ROOT
     %

We provide some sample output of the specification:

     +Question : What kind of animal rides a catamaran ?
     +Answer : a cat

     +Question : What kind of animal reads a catalogue ?
     +Answer : a cat

     +Question : What kind of animal rides a catalogue ?
     +Answer : a cat

The third example illustrates a crucial weakness of the system. In its current version, it cannot place on the verb semantic information from a lexical item located by lexical pointer. Thus, given that cat points at catamaran, we cannot select a verb which takes a direct object which is a boat. Similarly, if catalogue is pointed to, we cannot specify a verb which takes reading matter as a direct object. More generally, this illustrates that human-generated verbal humour of the sort is remarkably modular in that it must be capable of re-inputting into the syntax lexical information located at several removes from the original lexical entries.

7.2 Partial String Match Riddles

In another type of string match riddle, drawn from Gounaud (1981), the range of pivots is formally limited to paronyms of mouse and mice.

7.2.1 Corpus

     Q: Name two mice genders.
     A: Mouseculine and feminine.

     Q: Where do mice go to get their prescriptions filled?
     A: A pharmousey.

     Q: Name a literary classic about French mice.
     A: Les Mouseerables.

     Q: Why do girl mice always beat boy mice in a race?
     A: Because mice guys always finish last.

7.2.2 Analysis

Formally, the relation in play is that of a partial string match between pivot and superstring, sometimes only on the phonological level, as in the second example, sometimes only on the orthographic level, as in the third, but usually on both levels, as in the rest. In most of the examples in the corpus, the match is based on the coexistence of [m] and [s] separated by one vowel other than the normal vowel of mouse. Note also that the superstring containing the pivot may be formed by a common noun (pharmousey, mouseculine), a proper name (Les Mouseerables) or a fixed locution (mice guys always finish last).

Formally, the question may represent, among other things, a command Name ... calling for the definition of the superstring word, a question based on the role of the superstring word Where do..., or a question based on a causal explanation.

Beyond the operations found in the first set of riddles, directed string match riddles require that the output word undergo a reverse manipulation in which the string partially matching mouse be replaced with the string mouse. Thus, masculine becomes mouseculine, pharmacy becomes pharmousey, etc.

7.2.3 Modelling

The attempt to model examples of the sort brings even more sharply into focus the complex interplay of levels required and in particular, the difficulty of transmitting information to the various modules involved in generation. The early stages of the process are relatively straightforward. Suppose we wish to produce riddles based on the noun horse. To illustrate the range of possibilities, we search the CUVOALD for a match on the initial [h] and the [s] of horse, with any intervening vowel. Items found must be nouns (K6). An AWK search gives the following results, excluding forms derived from horse.

     behest | bI'hest | K6% | 2 |
     field-hospital | fild-'h0spItl | K6$ | 4 |
     hacienda | ,h&sI'end@ | K6$ | 4 |
     hasp | hAsp | K6$ | 1 |
     hassle | 'h&sl | J2%,K6% | 22A,3A,6A |
     hassock | 'h&s@k | K6% | 2 |
     hearse | h3s | K6% | 1 |
     histogram | 'hIst@gr&m | K6% | 3 |
     hospice | 'h0spIs | K6% | 2 |
     hospital | 'h0spItl | K6% | 3 |
     hostage | 'h0stIdZ | K6% | 2 |
     hostel | 'h0stl | K6% | 2 |
     hosteller | 'h0st@l@R | K6% | 3 |
     husk | hVsk | H0$,K6% | 1 |
     hustler | 'hVsl@R | K6% | 2 |

Let us assume that we select the string hospital. To produce a riddle, we must on the one hand perform the complex string manipulation required to transform hospital into horspital, thus leaving the realm of attested lexical entries. In other words, the generation system must be capable of synthesizing new lexical entries by rule. In research elsewhere (Levison and Lessard 1995), we have taken the first steps in this direction, but much remains to be done.

On the semantic level, we must perform an operation which takes the semantic traits of hospital and uses them to construct an utterance (say, Where do horses go when they are ill?). However, as we saw earlier in the case of cat:catamaran, in the current VINCI system this is not possible.

Clearly, this class of riddles pushes computational modelling to its limits.

8. Cognitive Riddles

Sarnoff and Ruffins (1974) represents a very disparate collection of riddles, some of which bring into play cognitive factors. We illustrate the problem with a series of examples:

8.1 Corpus

     Q: Why does a cat, when it enters a room, look first to one side and
        then to the other?
     A: Because it can't look to both sides at the same time.

     Q: What do you call a woman who doesn't have all her fingers on one hand?
     A: Normal. Her fingers are divided between her two hands.

     Q: What did the boy octopus say to the girl octopus?
     A: I want to hold your hand, hand, hand, hand, hand, hand, hand, hand.

8.2 Analysis

In these examples, syntactic analysis is insufficient to distinguish the complexity of the riddle. We are confronted rather with pragmatic strategies which lead us down a sort of garden path. In the first example, the pragmatic expectation is that the cat's looking to both sides implies some goal-directed behaviour (to see enemies, mice, etc.), while in fact the riddle hinges on a much lower-level constraint (the impossibility of performing both operations simultaneously). The second example is similar. In the third, cultural factors are added. We must make intertextual reference to a Beatles' song, and multiply the last word based on our world knowledge of octopus physiology.

At the current state of research, generation of riddles of the sort is well beyond our capacities.

9. Conclusion

The examples we have examined here lead to a number of conclusions:

The fact that riddles operate not only on abstract lexical entries (mouse) but also on inflected forms provides an argument for making such units accessible to all levels of processing in generative modelling, and suggests that humans also have easy access to such forms.

Riddles bring into contact a number of distinct linguistic levels, ranging from the syntactic, to the lexical to the phonetic. Furthermore, these levels are not traversed in only a simple order of syntax -lexicon - morphology. Corpus data shows that humans have a high degree of plasticity in their manipulation of various types of linguistic information. Clearly, generative environment used to model humour must have a similar degree of plasticity, allowing not only syntax to drive lexicon, morphology, and phonology, but also the reverse.

We have seen that the VINCI environment deals with core riddles with no great difficulty. On the other hand, in its current version, it fails when faced either with riddles which require complex string processing at the level of the lexicon (processing which humans themselves find difficult), or when faced with riddles which require complex cognitive processing. At the same time, the border between cognitive processing and language use is most tenuous in the case of riddles.

There also remain many questions. For example, we have seen that the notion of the pivot, together with that of lexical pointer, gives us additional insight into riddle structure; however, at the same time, it forces us to wonder about the genetic aspect of this mechanism. This is perhaps the most challenging aspect of riddle research: to model the learning of the phenomenon in its transition from corpus of examples to rule-based device. We conclude with a peek into this process provided by Gounaud (1981):

References

Anderson, John (1983). The Architecture of Cognition, Cambridge, Mass.: Harvard University Press.

Beisner, Monica (1983). Monica Beisner's Book of Riddles, London: Jonathan Cape.

Freud, Sigmund (1976). Jokes and their Relation to the Unconscious, Harmondsworth: Penguin Books.

Gounaud, Karen (1981). A Very Mice Joke Book, Boston: Houghton Mifflin.

Karmiloff-Smith, Annette (1992) Beyond Modularity: a developmental perspective, Cambridge, Mass.: MIT Press.

Lessard, Greg (1988). "Tom Swifties: analyse sémantique et formelle." Paper presented at Canadian Linguistic Association Annual Meeting, Windsor, Ontario.

Lessard, Greg; and Levison, Michael (1992). "Computational Modelling of Linguistic Humour." Paper presented at ALLC/ACH92, Oxford.

Levison, Michael; and Lessard, Greg (1992). "A System for Natural Language Sentence Generation." Computers and the Humanities 26:43-58.

Levison, Michael; and Lessard, Greg (1995). "New Words from Old: A Formalism for Word-Formation." Computers and the Humanities 29:463-479.

Maranda, E.K. (1971). "The Logic of Riddles." In Structural analysis of oral tradition, edited by P. and E.K. Maranda, 189-234. Philadelphia: University of Philadelphia Press.

Mitton, Roger (1992). A description of a computer-usable dictionary file based on the Oxford Advanced Learner's dictionary of current English. Computer-readable file, Oxford Text Archive (ota.ox.ac.uk).

Most, Bernard (1991). Pets in Trumpets and Other Word-Play Riddles, New York: Harcourt Brace Jovanovich.

Pepicello, W.J.; and Green, Thomas (1984). The Language of Riddles: New Perspectives, Columbus: Ohio State University Press.

Raskin, Victor (1985). Semantic mechanisms of humor, Dordrecht: Reidel.

Sarnoff, Jane; and Ruffins, Reynold (1974). What? A Riddle Book, New York: Scribner.

Terban, Marvin (1982). Eight Ate: A Feast of Homonym Riddles, New York: Clarion Books.