Referential continuity and the coherence of discourse.

Two experiments were carried out to investigate the role of referential continuity in understanding discourse. In experiment I, a group of university students listened to s?ories and descriptive passages presented in three different versions: the original passages, versions in which trCze sentences occurred in a random order, and randomised versions in which referential continuity had been restored primarily by replacing pronouns and other terms with fuller and more appropriate noun phrases. The original stories were remembered better, and rated as more comprehensible, ihan the random versions, but the restoration of referential continuity ameliorated the effects of randomisation. Tite descriptive passages had little referential continuity from one sentence to the next, and as expected the effects of randomisation on co.mpre.hensibility and memory were negligible. In experiment 2, a group of skilled comprehenders and a group of less skilled comprehenders were selected from a population of 7-8-year-old children. The difference between the groups was known to be largely their inferential ability in reading texts. Both groups read a series of short stories presented in the same three versions as used in the previous experiment. .As predicted the ameliorating effects on memory of restoring referential cont’inuity in a randomised story were confined to the skilled group. The results are discussed in relation to the theories of story grammar, text microstructure, and mental models of discourse. *A number of the arguments against story grammars which we discuss in the introduction are part of an ‘oral tradition’. We know that similar points have been made by Emmon Bach and Keith Stcnning. Experiment 1 was carried out by the first author while he was supported by the Sloan Poundation at the Center for Cognitive Science at the University of Texas at Austin. Experiment 2 was carried out by the second author while she was supported by the Social Science Research Council at the Laboratory of Experimental Psychology at the University of Sussex. The third author was also supported by the Social Science Research Council. Reprint requests should be sent to P. N. Johnson-Lair& Centre for Research on Perception and Cognition, Laboratory of Experimental Psychology, University of Sussex, Brighton, BNl 9QG, England. 30 A. Garnham, J. Oakhid and P. N. Johnson-Laud What makes discourse story grammar comprises a st:t of rewrite rules, such as: STORY SETTING THEME PLOT RESOLUTION THEME (EVENT)* GOAL and, just as a sentence grammar assigns parse trees to sentences, so a story grammar assigns hierarchical tree structures to stories. But, unlike sentence grammars, grammars for stories .face four crucial problems. First, the terminal nodes in the structural tree for a story are assumed to be filled by propositions, but there are indefinitely many propositions which could occur in such categories as OUTCOME or ATTEMPT. There can be no finite listing of such propositions, and no story grammarian has ever provided an explicit set of principles for determining which propositions are members of which classes. It is doubtful whether any such principles will ever be forthcoming. Sentence grammars are indeed very different: their terminal nodes are filled by morphe:mes drawn from fixed classes, whose members can be specified by enumeration. Second, there are clearly sorrie constraints which must be placed on relations between the propositions expressed by the terminal nodes in a tree. Some of these constmints might have a ‘syrl;tactic’ flavour, but most of them concern continuity of content, and seem better classified as ‘semantic’. Thomdyke (1977) does not deal with this problem at all, so his grammar would generate a ‘story’ which comprises the beginning of Snow White, the middle portion of Cinderella, and the end of Sleeping Beauty. Rumelhart (1975) has attempted to place semantic constraints on the relations between the propositions He proposes rules such as: EPISODE EVENT REACTION =INITIATE (EVENT, REACTION) But he does not say how to determine when one event INITIATES another. Rejhwt tiat cm tinuity 3 1 Again, a finite listing of .a11 such pairs of events would be impossible, but, unless explicit criteria can be provided, there is no way of deciding whether a given tree is generated by the story grammar. Third, the majority of story grammars make use of a contcx~-ficc phrase structure grammar. Grammars of this type are more powerful than finitcstate grammars, and less powerful than tranformational grammars. Hence, a case must be made to justify their use rather than grammars of a stronger or weaker generative power. This issue has only recently been recognised by story grammarians, but it is important. Certain sorts of finite-state grammar directly correspond to the Markovian processes favoured by psycholinguists of the pre-Chomskyan era. Tllese processes can t4~: quite adequately carried out by the mechanism embodied in a ‘habit-family’ hierarchy (see Miller, 1967, who uses Michael Frayn’s delightful idea of a machine for writing newspaper stories to illustrate a finite-state grammar). It is sometimes said that a finite-state grammar fo,r stories would not suffice, because it could not accommodate s:if-embedding (Black and Wilensky, 1979). This claim must be stated wit!) care: on the one hand, finite-state devices can deal with any ,$hzite number of self-embeddings, and on the other, a finite-state device that permitted u1z.v sequence of constituents would plainly accommodate self-embedd.ing structures. A proof of the inadequacy of finite-state grammars for stories depends on showing (i) that an indefinitely large number of self-embeddings can occur within a single story, and (ii) that any such grammar which generated all legal selfembedded stories would also admit some non-stories (Levett, 1974). The proof is not easy, and it certainly has not been established for story grammars. Likewise, Black and Wilensky’s claim that a context-free grammar could not accommodate discontinuous elements is incorrect. If the node labels of a context-fr(ee grammar are construed asI complex symbols, rather than as unanalysable primitives, discontinuous ccnstituents can be handled quitlz readily (Stanley Peters, personal communication; Gazdar, in press). Fourth, story grammars are usually intended to deal with only a restrict4 set of stories. e.g., genre stories that possess a stereotyped and repetitive structure. But, if there is no independent way to specify the set of relevant stories, the whole exercise becomes dangerously circular: the grammar is intended to analyse just those stories that Tit its rules. Of course, actual story grammars appear very plausible, but this plausibility, we contend, derives largely from what users of the grammar bring to it, rather than from any explicit formulation of principles in the grammar. In many cases it is quite obvious when one event is the OUTCOME of another, but only explicitness about %he membership of such classes can give story grammars explanatory power. The virtue of story grammars is that they attempt to formalise intersentential relations which are used in understanding text. Their major empirical claim is that certain types of story have a specifiable structure, which is independent of their content, and which people know and use in the course of comprehension (Mandler and Johnson, 1980). However, there are many other forms of discourse that are entirely coherent, even though they fall outside the domain of stereotypical stories. Our main aim in the present paper is to investigate what makes texts in general coherent, since the same underl::ing principles may well do the work supposedly done by story grammars. There appears to be one overriding and necessary condition for the coherence of discourse: it should be possible fo construct a unitary representation that integrates all the information carried in its separate sentences. This condition in turn requires that the sentences make reference, either explicitly or implicitly, to referents in common. Discourse is sometimes about a single major topic to which the majority of sentences make reference, but perhaps more often a series of referents is introduced from one sentence to the next. The ease of establishing what these referents are depends on the referential continuity of the text. Ehrlich and JohnsonLaird (1980) have shown that a referential discontinuity-the occurrence of a sentence expressing a proposition that can only be integrated subsequently-has a disruptive effect on both comprehension and memory. Coherence also depends on abiding by the general co-operative principle that sentences should be structured so that they can be readily understood (see &ice, 1975). F or example, information that is taken for granted should either have been previously established or be readily inferable. I-Iowever, given that there are common referents, and that they can be recovered without difficulty, then how sensib’le a discourse seems will depend on the plausibility of the actions, states, ‘and relations, in which the referents participate, and on the order in which these events occur. If a text reports an unlikely sequence of singularly implausible events, then strictly speaking it is not incoherent, but rather the world it describes is bizarre. We accordingly distinguish between the coherence of a discourse and its plausibility. Coherence depends on common referents, referential continuity and general adherence to the Grizean co-operative principle. Plausibility depends on verisimilitude to known intentional, causal, and temporal relations between objects in the world. One of the main pieces of evidence on which psychological claims about story grammars are made is that jumbled versions of stories are more diff?cuIt to understand than the original stories. The usual explanation of this finding is that jumbling the sentences in a story breaks up its overall strucReferential continuity 33 twre, and hence it cannot be parsed by a story grammar. Thus, one of the reader’s main guides to understanding, a knowledge of the rules of story grammar, is no longer of any use in the processing of such passages, However, it is clear that jumbled versions of stories lack both continuity of reference and a pl2.usible sequence of events (insofar as they can be understood in the absence of referential continuity). Thus, for example, a yronoun in a jumbled text may entirely lack any plausible antecedent. There is evidence that it takes time to find an antecedent for an anaphoric expression (e.g., Garrod and Stanford, 1977; Haviland and Clark, 1974); if the task becomes impossible then readers will be una’,le to form an integrated representation, and their processing of subsequent text may be impaired by a prolonged search for possible antec,Jents. They may even realise that the story has been scrambled, and cease to treat it as a coherent text. In Experiment 1, we examined the comprehension and memory of randomised texts in which referential continuity had been restored. A preliminary study had confirmed that randomised stories of this revised variety were better remembered than unmodified randomisations, though not as well remembered as the original stories. Many descriptive passages have a more diffuse coherence than stories: there is little immediate referential continuity from one sentence to the next, but rather each sentence takes up a different aspect of the topic. We can therefore predict that randomisation should have r.elatively little effect on performance with such texts and, furthermore, that any effects should disappear when referential continuity is restored. Experiment 1 compared such descriptions with stories, and examined the comprehensibility and memorability of three versions of them: the original texts, versions with the sentences in random order: and randomised versions in which referential continuity had been restored primarily by replacing pronouns and other isolated dnaphoric expressions with more complete and more appropriate noun phrases.