The Story Telling Laboratory

The Story Telling Laboratory is a research initiative that brings together ongoing projects that study various areas related to stories: how they are constructed, how they can be explored in an interactive manner, how they are converted into text or video, how their texts are read out aloud in a convincing manner, how they can be rendered as verse…

This page describes the research that has been carried out over the past 20 years. Given the volume of accumulated results, the description has been grouped into sections under headings that seemed descriptive. Some papers may appear under more then one heading, when their content touches upon several topics.

For a description of the motivation behind the general approach of The Story Telling Laboratory and considerations on what topics get addressed when, please read the following entry On Motivation and Expected Impact

Projects that are currently under way within The Story Telling Laboratory are listed below. However, results so far on this line of research have shown that there are many open issues to explore and many different approaches will be required to explore them.

If you think you can contribute to this initiative, please get in touch with us by sending an email message to .

The TSTL acronym pays homage to Robert Louis Stevenson, as a shortened version of Tusitala, which is Samoan for “Teller of Tales”, and which he used as a pseudonym.

Narrative

The overall goal of this particular research line is the computational modelling of literary creativity in a broad sense, including both the generation of narrative (this section) and the generation of poetry (see ‘Poetry’ below). Although this has been more a result of changes of funding and personnel over the years (see the ‘Caveat’ above), there is indeed an underlying conviction that adequate modelling of literary endeavours needs to address issues both at the level of content – which is more relevant for narrative – and at the level of form – which is more relevant for poetry. As discussed in the ‘Caveat’ above, the descriptions given below are not intended as a formal theories or exhaustive accounts but rather as reports on partial expeditions of discovery into of small parts of the large territory that we hope to explore in the long run.

For completeness and ease of access, wherever specific papers contain material addressing more than one topic, they are listed below under more than one heading.

General Considerations on Narrative

Although “storytelling” is often used as a monolythic concept, we have found it useful to consider that it might be deconstructed into a number of related concepts that interact to realise its underlying complexity. Our current view of this deconstruction is still in flux, but the following subdivisions have provided us with useful partitions that have allowed significant progress to be made. We are also very much aware that any deconstruction we can consider is only partial, and that many relevant subtasks are still being left out. The following set of subtasks is proposed here only as a means for explaining the work accumulated over the years. In our experience, the different ways in which people use narrative to communicate show a significant presence of the following subtasks:

This partial deconstruction is very useful to make progress at each of these subtasks, but it also provides the means for “reconstructing” the overall tasks as a combination of the subtasks. A significant part of our research effort has gone into exploring how this might be achieved, constituting therefore an additional partition:

Additionally, it became clear that the impact that a narrative has on an audience is heavily influenced by a number of factors that are not necessarily explicit in the surface form of the narrative as the audience receives it, but which are inferred by the audience during interpretation. Examples of this are emotions – both manifested by the characters and experienced by the audience –, affinities between characters, personalities of characters, suspense… Explicit investigation of how these factors may be handled – identified, represented, built – computationally became therefore an additional partition of our research plan:

At various points in our trajectory we have made efforts to sit back, review the accumulated work, and reflect upon the insights arising from it. The following two papers constitute two different such efforts, thirteen years apart.

Representation of Narrative

A fundamental challenge for the development of any artificial intelligence solution is the choice of the right representation of the problem being considered. Ideal as it would have been to find a single representation suitable for all the various tasks related to narrative that we have considered, our experience is that specific tasks are best served by particular representations, and that the resulting set of representations is not necessarily interchangeable or even compatible. The following sections describe some of the insights arising from the specific choices that we have considered. These include representation formats, methodological approaches, or relevant phenomena that have required specific solutions.

Desiderata

When relevant insights have arisen from the accumulation of evidence from working on the same concepts but applied to different tasks and using different technologies, we have made a point of trying to make them available to the scientific community in general in the form of scientific papers. The following two papers constitute instances of this type of effort.

Corpora

Corpus-based approaches to computational linguistics have over time have shown great advantages, so they were clearly a valuable tool to consider in our various attempts at solving specific tasks related to narrative. The compilation and annotation of corpora have been important elements in our work to this point on several of these tasks, such as: identification of narrative schemas (see ‘Narrative schemas” below) and emotional connotations of text (see ‘Emotions’ below).

Controlled Natural Language (CNL)

As the number of tasks related to narrative that we considered increased (and the number of technologies, each with their associated representation format) we realised that an important challenge for the field of computational narrative would be to find a reasonable format for interchanging resources / data across different implementations. One of the possibilities we explored as a posible solution to this challenge was the use of Controlled Natural Language, a simple text that is formal enough to be parsed without problema and easily produced from any more complex representation. Such a representation would be easy enough for any computational narrative application to produce. The onus of transcribing outputs from other systems onto the internal notation of each research effort would lie on the implementation of the parser from the Controlled Natural Language onto the internal notation in each case. We relied on this idea to propose a posible format for a shared evaluation task on narrative generation, relying on such a Controlled Natural Language to provide the shared ground data, and on the parsers to evaluate the outputs.

Narrative Schemas

As we explored more and more theoretical accounts and computational implementations of various narrative tasks, it soon became clear that an important and useful concept was the idea that there are certain schemas of how information is organised in a narrative that recurr across existing stories in various formats – novels, movies, plays, TV series, folk tales… — regardless of the mode of presentation and the theory of narrative that is being considered. This concept had so far been explored mostly in accounts closer to the field of creative writing than formal narratology. We decided to explore how such accounts might be useful in computational terms, to generate knowledge resources in a first instance, and to inform generation process once those become available. This particular initiative has also been supported by the work on corpus annotation (see ‘Corpora’ above).

Ontologies

Description-logic based ontologies (prevalent in semantic web) are a useful technology for representing complex knowledge, and we have considered them as a posible solution for representing narrative. Most the work described here was the result of Federico Peinado’s PhD thesis, in which he explored a framework for developing applications of automated narration. The ontology he developed included features from Vladimir Propp’s “Morphology of the Folktale” – for the more traditional views on plot and narrative roles of characters — and from Robert McKee’s work on screenwriting – for more recent views on the progression of stories and interactivity.

Embedded Stories

Over the time we have spent attempting to model narrative, the phenomenon of embedded stories – and the associated concept of narrative levels – has emerged as a very significant feature present in most human narratives yet very rarely considered in computational accounts of narrative. For this reason, some of our most recent work attempts to shed some light on how these features might be addressed computationally.

Content Determination: Plot Building

An important task within a computational storytelling set up is the ability to construct new stories. This subtask of the whole alligns well with the stage of content determination in a classical natural language pipeline, where the content to be conveyed is either selected or constructed. The task of deciding how that content should be conveyed constitutes a separate task of discourse planning (see ‘Discourse Planning: Narrative Composition’ below). In terms of classical narratological concepts, this content determination subtask alligns reasonably well with the establishment of the fabula for the story, as opposed to the construction of a specific discourse for a given fabula (see ‘Discourse Planning: Narrative Composition’ below).

This content determination task is what is usually referred to as story generation. Historically, it was the main drive of our initial efforts in this field. During those, it became apparent very early on that a fundamental characteristic which played a very significant role in the perception of quality of a story was whether the events in it were somehow linked by a causal sequence from the start of the story towards its resolution. We therefore adopted the concept of plot defended by Forster – a chronological sequence of events linked by causality – as a driving force in our efforts. That is one of the reasons why most of the papers cited below refer to construction of plot. Another important reason is that we very early on adopted the decision to focus on the construction of conceptual descriptions of the structure of the narrative, leaving for a later stage the generation of either specific linear discourses that narrate it (see ‘Discourse Planning: Narrative Composition’ below) or text renditions of such discourses (these we have not worked much on, though some work is describe in ‘Text Generation’ below).

Our efforts to explore different alternative technologies for addressing the task of constructing plots have ranged over a number of different options:

General Efforts on Plot Generation

Propp-Based Approaches to Plot Construction

Simulation-Based Approaches to Plot Construction

Case-Based Reasoning Approaches to Plot Construction

Planner-Based Approaches to Plot Construction

Discourse Planning: Narrative Composition

Once we had identified that the challenge of automated storytelling benefited significantly from being deconstructed into a number of subtasks (see ‘General Considerations on Narrative’ above), we realised very soon that, while the stage of building a new story — story generation – was the focus of significant research efforts across the world, the stage of constructing a narrative discourse for an already established story was generally being neglected. This lead to a particular focus within our research group on this subtask over the years.

A number of specific research lines have been considered within this subtask:

Composing Narrative Discourse from Facts

Composing Narrative Discourse from Facts with Interpretability Check

Composing Discourse from Facts to Match a Plot

Building Stories with Subplots

Elements Relevant to Narrative

The impact of a given narrative on its audience is significantly affected by a number of factors that are not necessarily explicit in the surface form of the narrative as the audience receives it, but which all audiences identify without difficulty during interpretation and use to construct their opinion of it. There are many factors that fall within this category. At the present time, only a few of them have been the focus of explicit exploration within our research group. The extent of our research in each case is described below:

Emotions

Emotions are clearly central to narrative. Indeed, one of the main purposes of narrative seems to be to allow an author to induce or convey a given set of emotions to an audience, to such an extent that these emotions are themselves very difficult to represent or describe separately from the narrative form. Given their importance we have considered over the years several approaches at identifying or representing emotions computationally. These efforts attempt to take into account existing psychological theories of emotion, and tend to focus on the relation between particular emotions and words that may be used in conveying them to others. In support of this initiative we have developed corpora, ontologies and lexical resources. We have explored emotional labels, a space of emotional coordinates, and polarity/intensity values. We have developed automated taggers and automated classifiers. We are still trying to find a way of exploiting the knowledge acquired in this task to other tasks.

Annotation of Emotions

Sentiment Analysis (polarity/intensity)

Character Affinity

The evolution over time of relations between characters is another very important element in narrative. Again, narrative seems indeed to be the prefered means in our culture to convey this type of information. We have explored various possible ways of representing character affinity and taking it into account when building stories: Bayesian models, fuzzy logic, agent-based systems, BDI models.

Character personality

The personality of characters is also a fundamental ingredient of narrative. Stories that have characters with rich personalities draw the attention of the audience, and coherence between the personality of characters and their actions seems to play a fundamental role in the perception of the quality of a story.

Modelling Suspense

Suspense is a complex abstract concept very often used when discussing the quality of narrative media (movies, games, stories…) or their impact on the audience. We have looked into psychological theories of suspense and carried out empirical work on trying to identify how the various possible mechanisms of suspense operate on audience perceptions of narrative.

Computational Asssessment of Narrative

In general terms the ability to assess the quality of system outputs is known to be critical for many artificial intelligence approaches – Generate and Test, fitness functions in evolutionary solutions… In recent times the Computational Creativity community developed a strict preference for systems that included the ability to rate their own outputs. Cognitive models of the writing task assign significant important to a cyclic approach that includes reflection-based revisión of intermediate drafts. All these considerations show the importance of having adequate models of how narrative may be assessed computationally.

As in other cases considered above, there are many factors that ought to be considered in the assessment of narrative, and our research group has only manged to consider some of them to this point. The list of factors considered is no way exhaustive – there are many others that should be considered with at least equal or even higher priority – and the efforts described here are in all cases only initial approximations to the task. For what they are worth, the following factors have been explored:

Content

Structure

Novelty

Narrative as an Integrated Process of Specific Subtasks

In trying to model the construction of narrative computationally we aim neither to mirror necessarily the observed characteristics of how humans do it, nor to restrict ourselves to the processes featured in the latest fashion in artificial intelligence technologies. Instead we hope to find a happy medium that combines the best from both worlds. To this end we have come up with a deconstruction of the storytelling process into a number of subtasks that make sense in terms of how humans adresss narrative and that at the same time are suitable for computational modelling. This approach has lead to work on how these subtasks might be fruitfully combined into complex models of the storytelling task in a broader sense.
Our efforts on this research lines have addressed the following topics:

Cognitive Issues related to Narrative

Interpretation and Revision

StoryTelling Architecture

Colaborative Narrative Creation

Assisted Narrative Creation

Interactive Narrative

Text Generation

The ongoing research effort on automated storytelling is focused mainly on higher level tasks such as story construction and narrative composition, and critical ancillary aspects such as narrative representation, overall system architecture or computational modelling of aspects relevant to narrative (see sections above for details). Nevertheless, some effort has been devoted over the years to specific tasks related to the generation of text, either intended to be integrated in a storytelling system or into a poetry generation system.

Descriptive Text

Existing solutions for story generation tend to focus mostly on compiling the sequence of events that constitutes a story, with the descriptions of places and characters either altogether ommited or delegated to an accompanying graphical interface akin to those used for video games. Research on how to generate descriptive text for the conceptual representation of a given element relevant to a story was an important challenge that we hope to integrate in the future into the overall storytelling solution.

Character Description

Lexical Choice in Referring Expresions

Evolutionary Solution for Referring Expression Generation

Dialogue in Narrative

Dialogue has a fundamental role to play in narrative, to the extent that it is a very important medium for conveying narrative momentum in very prevalent genres (film and drama). Yet in prevailing approaches to computational storytelling it is generally either overlooked altogether or addressed with very basic template-based solutions that are far behing the state of the art in text generation. The recent emergence of chatbots as a promising field and the current trend of using Transformers for all natural language tasks are slowly changing this, but there is still a lot of work to do in alligning these efforts with the accumulated body of knowledge on narrative. At the risk of disappointing those that believe that these new technologies can solve all without resorting to any prior efforts, we do believe that a significant part of the complexity of narrative will still require knowledge-based solutions.

Rhetorical Figures

Rhetorical figures such as comparisons, analogies and metaphors are often considered informative indicators of the level of literary aspiration of a text: if a text has metaphors or comparisons, then one can asume that the author is making an effort to impress his audience at an aesthetic level. It would therefore seem useful to a computational program with literary aspirations to have access to such tropes. This simple truth has met with the obstacle of the inherent difficulty of modelling these rhetorical devices, but the lack of success has not been lack of attempts or perseverance. The papers listed below cover various attempts in our research group over the years to model the task of constructing rhetorical figures. The most significant difficulties in this task lie in the fact that these figures operate at the semantic and pragmatic levels of linguistics, which are at present the ones most poorly covered by computational solutions.

Poetry

Three fundamental premises must be kept in mind when reading the considerations below on computational modelling of poetry generation.

  1. Poetry generation should ideally be viewed as a refinement on prose generation, involving an additional layer of aesthetic requirements on top of the fundamental requirements on content that any literary product should satisfy.
  2. An approach to this task from an engineering point of view may consider the computational modelling of the aesthetic requirements on their own, leaving aside the requirements on content, to allow for progress to be made without the complication of the joint problem.
  3. Any computational solutions arising from this effort should not be viewed as “poets” but rather as modules that might be useful as constituent elements of an automated poet (if and only when they have been integrated with appropriate solutions for producing adequate content).

The prevalent error of labelling any program that produces poem-like artifacts as a computer poet – in which some of our earlier papers described below did indeed incurr – is conceptually incorrect and leads to extremely misleading expectations. These misleading expectations often cloud the appropriate attribution of merit to such efforts.

The poetry-aware computational models described below might at some point in the future be integrated with some of the modules on generation of narrative discourse described above. If and when that happens, they would have at least a semblance of the most elementary abilities of a human author: some means of deciding what content they should include, in what order they should present it, and some idea of how it might be received by their audience. In their present form they have none of these, so their output should be assessed exclusively in terms of the aspects they consider when building it, which are currently restricted to meter, rhyme, stanza structure and some elementary restrictions on linguistic likelihood of the sequences of words they include.

Poetry Generation

The research efforts described below have been grouped into the following categories for ease of access:

Evaluation

Multilingual

N-gram based

Case-Based Reasoning

Rule-based