|
|
An Introduction to Narrative Generators: How Computers Create Works of Fiction Rafael Pérez y Pérez and Mike Sharples Oxford University Press, 2023 |
Reviewed: David Longman, 11 August 2023
This is a book about stories, how they are creatively developed from various textual elements and specifically how computers can help us understand how we do such a marvellous thing. It is a follow-on from the authors’ previous book ‘Story Machines’ which takes a broader and more historical approach to the topic of computational machines – digital or otherwise – which can generate coherent texts that human readers might recognise as stories. ‘Narrative Generators’ is more focused on detailed explanations of how computers can be made to produce stories from given elements such as characters, events, actions, plots and even themes in time-based sequences following a traditional format of beginning, middle and end.
However, this is not a book of literary theory in the usual sense and it does not approach its subject from the perspective of semantics, linguistics, poetics or aesthetics although all these appear here in some form. Instead it aims to describe how stories are built up from simpler components of text or data using procedural rules to combine them. Such techniques are not restricted to fiction and have already been in use for some time, often invisibly, to automate aspects of journalism, e.g. the production of news, some of which is genuine but some that is itself a form of fiction or ‘fake news’.
For these auth;ors the rationale for building narrative generators (or story machines) rests on the idea that we “live through stories” or narratives about ourselves, our lives and the worlds we occupy. Stories suggest not merely the act of imaginative creation – though that is the focus of this book – but a fundamental way in which our minds work to make sense of our experience of the world, a fundamental feature of human thought and cognition through which we shape our lives intellectually, emotionally and socially.
Understanding this process is a key reason for the scientific and computational study of how stories are made. We still know too little about how ‘narrative generation’ is formulated in human thinking and here the authors remind us of the important role that the science of computation has to play in furthering our knowledge about the workings of the human mind:
“ … Computer science, and in particular artificial intelligence, is a powerful tool that can contribute to this endeavour. This book describes how computer programs can generate narratives and how studies of computational narrative can illuminate how humans tell stories.” (p2)
Here then is the scientific position for AI as a tool for understanding ideas about how we think, a stance that harks back to the very origins of Artificial Intelligence as an area of scientific study: building models to test hypotheses and theories of minds in order to understand them. This endeavour requires exhaustive detail in the construction of models, of course, but also and, importantly, models of story making which should produce outputs that are coherent, aesthetically appropriate or intelligible to an audience. They must ‘work’ to some extent and be recognisable to the human reader as stories which even to a small degree must reach back into human experience and traditions, they have to mean something.
In this regard the book’s subtitle, “How Computers Create Works of Fiction” is an important qualifier because, by exploring fiction specifically, it avoids the many problems of truth, bias and, to a large extent, the issue of informational garbage that such machines may produce. Fiction, after all, is essentially not bound by the logic of truth but ruled mostly by cultural expectations. A word salad produced in the name of fiction is just that and it can be ignored or admitted as the conventions of taste, aesthetics, and perhaps marketing, might allow. (Fiction generates its own ‘fake news’ of which The Protocols of the Elders of Zion is a famous, and profoundly appalling, example). None of this directly concerns the purpose of this book but the computational work described here can help to explain how such content is manufactured particularly in the present age of automated ‘intelligence’.
This point should be emphasised. The aims of the examples and techniques described here are not merely to devise cognitive machines that can imitate the creativity that lies behind story making and thereby add to the countless screeds of deathless prose that advance across our cultural landscapes (we have GPTs for that!). This is a book that illustrates a search for an understanding, an explanation, of important aspects of human creativity through which we shape our lives intellectually, emotionally and socially. It is first and foremost a book of experimental science, a testing ground for understanding theories and hypotheses about how our creative minds work.
The book takes a specialised approach in two ways. First, it does not assume advanced knowledge of computer programming (although most of the examples under discussion have been created with sophisticated programming techniques) but second, it does presume that the reader is able interpret dynamically what are essentially static accounts of how such computational machines can be constructed. For this a basic understanding of how computer programming works is helpful. An obvious analogy might be to suggest it is rather like trying to understand the experience of driving a car by reading a Haynes manual!
In this respect ‘Narrative Machines’ at times requires considerable acts of imagination and points up the challenges that this book can present to the unaccustomed reader because computational models such as these can only be tested and their effects seen by running them programmatically. This is, if you like, the key component of how the ‘computational method’ does it work.
However, logical thinking, a constructionist turn of mind combined and an interest in the general aims of computational science is enough to grasp the modelling that is going on here. Note that there are no claims here about ‘machine intelligence’ and scepticism towards that particular trope is entirely consistent with the idea that computational machines can generate insights into the mental processes that underlie imaginative writing.
The digital transformation of social communication and information practices across the entire spectrum of everyday genres including financial, commercial, scientific, political and artistic has been gathering pace for some time, perhaps longer than many of us realise, and the kind of techniques discussed in this book have already effected significant changes in our use of email, spell checkers, search engines and word processors. It is no longer a pun to refer to some of these changes as transformational because that very term now appears in the descriptive name of one of the most prominent of the current crop of digital artefacts Generative Pre-trained Transformers or GPTs.
This raises an interesting and crucial point about the history of all these endeavours to create machines which, in various ways, represent or mimic human language use. As already noted Narrative Generators stands at a juncture between what might now be termed ‘traditional’ AI where the aim is to understand how human cognition works by building dynamic models of cognitive processes. As noted by the authors this might be thought of as the representational approach, building testable models of how minds work, albeit in restricted domains such as storytelling. To build such models relies on science where theory building and hypothesis testing are paramount. Chapters 1 to 8 in this book take us through that approach and various projects that are described here are based on postulations about the elemental components from which stories are made.
The book is well structured and progresses from describing relatively simple and accessible models using templates much like online forms that are such a familiar part of everyday online life. Phrasal substitutions are made from a simple database stored in a spreadsheets containing lists of character names, descriptions of actions and locations, and can be developed to be surprisingly subtle even if in their simplest form they are limited, ultimately, by a lack of flexibility. Chapters 2-8 takes us through increasingly sophisticated approaches that retain something of this template model but develop it to allow for greater variation and subtlety. Certain kinds of reasoning, computationally expressed, can be introduced into the manner in which characters, events, and outcomes can be ‘substituted’ to increasingly sophisticated levels such as authorial intentions and thematic frameworks.
But today we have entered the age of what might be termed neural AI. The most prominent, and for many perhaps the more disturbing, example is the power of the ‘neural net’ built from artificial ‘neurons’. The difference is profound. Its most visible incarnation is in the form of GPTs. Whereas in Chapters 1-8 we can trace much of the intellectual reasoning and computational engineering behind the design and implementation of story making machines, the arrival of the artificial neuron outlined in Chapters 9 and subsequently in chapters 10 and 11 its role in the architecture of artefacts such as GPT-4 offers little by way of any structural analysis of how stories are generated. All we have is a digital analogue for what is presumed to be the biological basis of brains and hence of thought – the neuron. And here too is a disjunction between the attempts to model intellectual processes described in other sections of the book for neural nets rely almost entirely on sophisticated statistical mathematics. Astonishing they may be but we are guessing, wondering yet somewhat baffled, as to how they achieve their linguistic capability. They are yet another example of Arthur C Clarke’s famous dictum that “any sufficiently advanced technology is indistinguishable from magic.”
There is history here of course and the idea of an artificial neuron, as the authors point out, reaches back into the origin stories of Artificial Intelligence, probably arising before the emergence of cognitive theories about how minds work which simply enabled more tractable approaches to computational modelling. The only ‘model’ that neural nets provide is an electronic analogue for the brain in so far as we agree that (a) the brain comprises neurons and (b) that these neurons are the building blocks of minds (Minsky’s famous meat machine) but we do not have a model of how a GPT can tell us a story, only that it can. In other words while GPTs offer a performance of story making, as yet they do not provide a model of how stories are made. Perhaps, ironically, a new science is required if we are to understand this – an AI of AI!
This may cast us back to one of the early debates in the history of the cognitive sciences and emergent AI when, in 1959, Noam Chomsky critiqued the Skinnerian approach to understanding language generation. In simple terms, Skinner theorised that all learning, and language learning in particular, arises from the acquisition of examples that are ‘reinforced’, i.e. made ‘correct’, by various social means (parental engagement, pedagogical practice etc.). Chomsky on the other hand argued that such a mode of learning cannot explain why grammatically correct but nonsensical sentences are nevertheless possible such as “green ideas sleep furiously” (or, from an earlier age, Lewis Caroll’s Jabberwocky poem). For Chomsky there had to be an underlying, conceptual ‘grammar machine’ on which all reinforcement relies and by which it is calibrated, a ‘machine’ that could in principle be described and explained.
This book might not appeal to an audience for whom language and storytelling are regarded as fundamentally human activities and an expression of our non-mechanical spirit. However, it is timely and relevant in today’s cultural landscape where the machine generation of linguistic content has reached new heights of sophistication often indistinguishable from that produced by human writers. Restricting itself to the understanding of how fictional, imaginative texts are produced allows for a focus on the type of cultural object in which we are immersed on a daily basis.
There is however an audience ranging from computer scientists engaged in this type of investigative work to, importantly, teachers at all levels of school and university education for whom the role of computation in understanding how machines can be made to do these things does not deny the ineffable qualities of human creativity. Indeed, a rather striking, if understated, observation throughout this book is how often practical human intervention is required to edit the output of the various systems described including GPTs.
All too often, and all too obviously, the stories that they produce cry out for an editorial hand, for a human to come along and improve it. This weakness is a strength because the limitations of our computationally generated narrative models are difficult to hide. In contrast, tools such as GPT that rely on mathematically driven models of neural architecture may present a greater risk to human creativity because for the illusory finesse with which they produce refined grammatical texts may all too easily obscure underlying inaccuracies of fact or distorted moral and cultural values.