An Introduction to AI Story Generation

An Introduction to AI Story Generation

. 27 min read

Table of Contents


Automated story generation is the use of an intelligent system to produce a fictional story from a minimal set of inputs. This is a problem that has long been explored by AI researchers, since it strikes at some fundamental research questions in artificial intelligence.

To tell a story, an intelligent system has to have a lot of knowledge, both about how to tell a story and about how the world works. These concepts need to be grounded to be able to tell coherent stories. Story generation is therefore an excellent way to know if an intelligent system truly understands something. To understand a concept, one must be able to put that concept into practice — telling a story in which a concept is used correctly is one way of doing that.

For example, if an AI system tells a story about going to a restaurant, as simple as that sounds, we discover very quickly what the system doesn’t understand when it messes up basic details. Besides understanding concepts, storytelling also requires an understanding of the listener or reader, known as a theory of mind -- a model of the listener to reason about what needs to be said or what can be left out and still convey a comprehensible story.

In addition to these fundamental AI research problems, automated story generation is also worth studying for the applications it may enable. Aside from the grand challenge of an AI system that can write a book that people would want to read, storytelling appears in many places in society:

  • Human-AI coordination: there are times when it is easier to communicate via narrative. For example, communicating via vignettes helps with coordination because it sets expectations against which to gauge the appropriateness of behavior. Humans often find it easier to explain via vignettes, and are often able to more easily process complex procedural information via vignettes.
  • Human-AI rapport: Telling and listening to stories is also a way that humans build rapport.
  • Explainable AI: Explanations can help humans understand what an AI system does. For sequential decision making tasks (e.g. robotics) this might entail a temporal component to the explanation resembling a story.
  • Computer games: many computer games feature stories or plots, which can be generated or customized. Going beyond linear plots, interactive stories are those in which the user assumes the role of a character in a story and is able to change the story with their actions. To be able to respond to novel user actions requires the ability to adapt or re-write the plot.
  • Training and education: inquiry-based learning puts learners in the role of experts and scenarios can be generated to meet pedagogical needs (similar to interactive stories above).

The remainder of this article will present a primer on the field of research that I think my students need to know to get started on research on automated story generation, and that anyone interested in the topic of automated story generation may find it informative. A caveat: since I have been actively researching automated story generation for nearly two decades, this primer will be somewhat biased toward work from my research group and collaborators.

What is Automated Story Generation?

This is an article from Popular Mechanics in 1931.

Let’s define some terms:.

  • Narrative: The recounting of a sequence of events that have a continuant subject and constitute a whole (Prince, 1987). An event describes some change in the state of the world. A “continuant subject” means there is some relationship between the events—it is about something and not a random list of unrelated events. What “relates” events is not entirely clear but I’ll get to that later.
  • Story: A narrative that tells a story has certain properties that one comes to expect. All stories are narratives, but not all narratives are stories. Unfortunately I cannot point to a specific set of criteria that makes people regard a narrative as a story. One strong contender, however, is a structuring of events in order to have a particular effect on an audience.
  • Plot: A plot is the outline of main incidents in a narrative.

We might distinguish between automated story generation and automated plot generation. Recently, some have started distinguishing story generation from plot generation as whether the output of the system reads as an outline of main events versus having natural language that describes aspects of the story that are not strictly events, like descriptions, dialogue, and other elaborations. The distinction seems to be: does it read like a high-level outline of events, or does it look like something someone might find in a book. I’m not certain I would make this distinction, but I see the appeal for those that are more interested in surface form versus structure (two equally important and valid perspectives).

The fictional criteria distinguishes automated story generation from other storytelling technologies such as news writing, where the events are those that really happened in the real world. That is, news relies on the real world as the “generator” and then creates natural language prose. News generation is an important problem, but one that, in my opinion, should be separated out from automated story generation.

Minimal input is a criteria that I add to automated story generation to distinguish from story retelling. Story retelling is a problem where most or all of the story/plot is given and the automated system is producing an output that tracks the input closely. For example, one might give a story retelling system a trace of facts about a story in some abbreviated or structured form and the system might generate natural language prose that conveys those facts in order. In this case the “story” was already known but the surface form of the telling is variable. As to what a “minimal” set of input is is subject to debate. Should it be a single prompt for the start of the story? Should it be the start and a goal? Should it be a small number of plot points that get filled in? What about given domain knowledge? For machine learning systems, should we consider the corpus that is trained on as an input? Many researchers and practitioners will explicitly or implicitly assume different input parameters.

Narratology and Narrative Psychology

Before diving into technology, let’s look at some of the things we can learn from narratology and narrative psychology. Narratology is a humanistic field that concerns itself with the study of narratives. Narrative psychology is a branch of psychology that looks at what happens in the human mind when reading stories.

There are a lot of branches of narratology. As a technologist I find the most value in structural narratology, which has provided frameworks for thinking about the structure of narratives. I have found some of their theories and frameworks to be operationalizable (which is not to say that other branches of narratology are not useful, just that I haven’t been able to directly map their contributions to system engineering).

Structural narratology, drawing heavily from Bal (1998), analyze narratives at two levels:

  • Fabula: The fabula of a narrative is an enumeration of all the events that occur in the story world between the time the story begins and the time the story ends. The events in the fabula are temporally sequenced in the order that they occur, which is not necessarily the same order in which they are told. Most notably, the events in the fabula might not all exist in the final telling of the narrative; some events might need to be inferred from what is actually told. For example: “John departs his house. Three hours later John arrives at the White House. John mutters about the traffic jam.” The fabula clearly contains the events “John departs house” and “John arrives at the White House” and “John mutters”. We might infer that John also drove a car and was stuck in a traffic jam — an event that was not explicitly mentioned and furthermore would have happened between “depart” and “arrive” instead of afterward when the first clue is given.
  • Sjuzhet: The sjuzet of a narrative is a subset of the fabula that is presented via narration to the audience. It is not required to be told in chronological order, allowing for achronological tellings such as flash forward, flashback, ellipses (gaps in time), interleaving, achrony (randomization), etc.

Some further distinguish a third layer, text or media, which is the surface form that the reader/audience directly interfaces with. The text/media is the specific words or images from which the fabula is inferred.

While in reality human writers cognitively operate at all levels simultaneously, the distinction between fabula, sjuzhet, and text can be useful for operationalizing story generation systems. For example, some systems use a tripartite pipeline [citations needed] in which (1) an explicit fabula is generated as an exhaustive list of events, (2) a sjuzhet is generated by selecting a subset of events and possibly reordering them, and (3) the generation of natural language sentences that describe each event in the sjuzhet in greater and more readable detail.

One line of narrative psychology looks at the mental models that readers create when reads a story. The psychology of narrative comprehension point to the importance of causal relations (Trabasso 1982) when human readers model the relationships between story events. In narrative, a causal relation between two events is an indication that the temporally latter event is made possible in part by the presence of the temporally former event (narrative causality is not the same as causality as used in statistics or Bayesian inference, where causality between random variables means that the value of one random variable affects the distribution of values of the other random variable). Narrative causality can be due to direct cause — a wheel falling off a car causes it to crash — or through enablement —the could not have crashed if the ignition hadn’t been turned on.

Trabasso (1982) determined that human narrative understanding involved modeling of the causal relations between events in stories. A causal relation exists between two events e1 and e2 when e1 occurs before e2 and e2 could not occur if e1 had not occurred. Although causal relations implies that e1 causes e2, it is more accurate to say that the presence of a causal relation captures enablement: e1 enables e2.

Graesser et al. (1991) hypothesize that readers also model the goal hierarchies of story world characters, recognizing that characters emulate real humans in having goals and that actions also enable goals and intentions. Some events are goal events and some events are sub-goal events that are necessary steps in achieving a final goal event. For example, the event “John shoots Fred” might be a goal event preceded by a number of sub-goal events such as “John buys a gun” and “John loads the gun”. These sub-goal events precede the goal event and are also related to the goal event as part of a hierarchy. Some events also directly cause a character to form an intention, which manifests itself as a goal event later in the story (or at least the attempt to achieve a goal event). For example “Fred threatens to kidnap John’s children” might be an event that causes John to intend to shoot Fred. This captures the notion of direct cause between two events as mediated by the mental state of a character.

Non-Learning Story Generation Approaches

Let’s get into technologies. This cannot be exhaustive, so I have attempted to create some broad classes and give some examples of each. This section looks at non-machine-learning based approaches. Non-learning systems dominated much of the history of automated story generation. They can produce good plots, though the emphasis on natural language output has been reduced. The key defining feature of these techniques — for the most part — is the reliance on knowledge bases containing hand-coded knowledge structures.

Story Grammars

Computational grammars were designed to decide whether an input sequence would be accepted by a machine. Grammars can be reversed to make generative systems. The earliest known story generator (Grimes 1960) used a hand-crafted grammar. The details are largely lost to history.

The earliest known story generated by a grammar-based story generation system (1960).

In 1975, David Rumelhart (1975) published a grammar for story understanding. It was followed by a proposed story grammar by Thorndyke (1977).

The Rumelhart story grammar.

Black and Wilensky (1979) evaluate the grammars of Rumelhart and Thorndyke and come to the conclusion that they are not fruitful for story understanding. Rumelhart (1980) responds that Black and Wilensky misunderstood. Mandler and Johnson (1980) suggest that Black and Wilensky are throwing the baby out with the bathwater. Wilensky (1982) revisits story grammars and doubles-down on his critique. Wilensky (1983) then goes on to propose an alternative to grammars called “story points”, which resemble schemata for plot points (see next section) but aren’t generative. Rumelhart goes on to work on neural networks and invents the back-propagation algorithm.

Story Planners

Story planners start with the premise that the story generation process is a goal-driven process and apply some form of symbolic planner to the problem of generating a fabula. The plan is the story.

The system that is most commonly acknowledged as the first “intelligent” story generator is Tale Spin (Meehan 1977). Tale Spin originated out of work on story understanding, flipping things around to do story generation instead. It builds off two core ideas:

  • First is conceptual dependency theory, a theory of human natural language understanding that (for the purpose of conciseness) boils everything down into a small number of primitives: ATRANS (transfer an abstract relationship such as possession), PTRANS (transfer physical location), PROPEL (apply a force such as a push), GRASP (grasping an object), MOVE (movement of a body part).
  • Second is scripts, which are procedural formulae (made up of conceptual dependencies) for achieving goals. Given a goal for a character, Tale Spin would generate a story by picking a script. This might result in further goals by the main character or other characters and further script selection.
A story generated by the Tale Spin system.

One of the more notable contributions of Tale Spin were the “mis-spun tales”, which demonstrated the tight connections between the knowledge engineering of the scripts and the quality of the stories, as illustrated by example runs that failed in interesting ways.

Mis-spun tales generated by the Tale Spin system.

The Universe system followed with a more modern understanding of planning. We would now call Universe a greedy hierarchical task planner. Universe had a character generator (Lebowitz 1984) and a story planner (Lebowitz 1985). Universe, which generates soap operas, uses a library of schemas that indicate how a scene plays out. The schema can reference sub-schemas. Starting with a high-level scene goal (e.g., “churn two lovers”), the system would pick a schema that described how to make two characters unhappy. The schema would contain a number of sub-goals. The system would construct the story by iteratively decomposing sub-goals.

A plot fragment schema from the Universe system.
A story generated by the Universe system.

In the mid-90s symbolic planning became much more formalized and grounded in predicate logic. Formally, planners find a sequence of actions that transforms an initial state (given as a list of fact propositions) to a state in which a goal state holds (given as a list of propositions that must be made true). Action schema consisted of an action, operands, a precondition and an effect. A precondition is a set of propositions that must be true for an action to execute. An effect is a description of how the world changes in terms of propositions that become true and those that become false.

An action schema for a POCL planner.

There are many ways of doing planning. A Partial-Order Causal-Link (POCL) planner chains from the goal backward to the initial state by seeking actions that make goal conditions true, and then seeking actions that make the preconditions of those actions true. This process continues until everything grounds out in the propositions listed in the initial state. An argument can be made that a least-commitment plan resembles the mental model of a reader because each chain from a precondition of some action to an effect of another action resembles enablement relations.

A story plan generated by a POCL planner from Riedl and Young (2010).

One of the problems with symbolic planners is that enablement isn’t the only consideration for stories. If everything is driven by enablement of preconditions it ends up looking like all characters are collaborating on a shared goal — the given goal state. My Fabulist system (Riedl 2010) used a newly invented type of POCL planner that built plan structures that created the appearance of character goal hierarchies and character intentions. Fabulist uses the tripartite framework: it first generates a plan data structure as the fabula, sub-selects actions to be part of the sjuzhet, then uses templates to render actions into natural language.

A story plan generated by Fabulist. The orange bubbles show actions that are part of goal hierarchies.
A story generated by Fabulist corresponding to the above plan data structure.

Stephen Ware (2014) further modified POCL planners to ensure the appearance of character conflict (CPOCL). Planners like to avoid conflict because conflict breaks one or more of the plans of characters.

An example CPOCL plan with character conflict and un-executed actions.
A story generated by CPOCL.

Case Based Reasoning

Case based reasoning is a theory of intelligence based on the idea that most reasoning is not done from first principles but instead adapts memories of solutions to related problems to new contexts. When a problem is encountered, the agent retrieves a solution to an older related problem, applies the old solution to the new problem, adapts the old solution to better fit the needs of the current problem, and then stores the new solution.

There is reason to believe that humans tell stories by adapting existing stories to new contexts. Case Based Reasoning approaches to story generation typically assume the agent has access to a library of existing stories and uses the above retrieve-apply-adapt-store process to transform an old story (or several old stories) into a new story.

The Minstrel system (Turner 1993) was arguably the most famous case-based story generation system. It used a lot of special rules for adaptation.

A story generated by the Minstrel system.

ProtoPropp (Gervas et al. 2005) later used more general adaptation techniques. ProtoPropp used Vladimir Propp’s analyses of Russian Folktales to construct a case library.

Case library for the ProtoPropp system.

The Mexica system (Perez y Perez and Sharples 2001) operationalizes a theory of creative writing. It alternates between two processes. The first process (engagement) operates a bit like a case-based reasoner. It compares the generated story so far to existing stories in its library in order to produce some interesting continuations. These continuations aren’t guaranteed to link up with existing story elements so a second process (reflection) uses something that looks a lot like a backward chaining planner to revise the new snippets and add new actions to link everything up.

A story generated by the Mexica system. Regular text was generated during the engagement phase. Text in italics was generated during the reflection phase.

Some of my work attempts to further reconcile case-based reasoning and POCL planning (Riedl 2008) as well as did some closer exploration of the adaptation phase of case-based planning for stories (Li and Riedl 2010). Related to case-based reasoning is story generation via analogical reasoning (Riedl and Leon 2009).

The SayAnything system (Swanson and Gordon 2012) is an example of textual case-based reasoning. Textual case-based reasoning treats a large text corpus as an unstructured case base. Most case bases are structured for easier retrieval and adaptation, so retrieving and adapting is much more challenging with unstructured text. The SayAnything system mines stories from blogs and retrieves snippets from the blog stories in response to user-written segments of story. The system takes turns between human and computer adding to the story.

Character-Based Simulation

The above approaches can be thought of as author-centric (Riedl 2004) — the story generators assume the role of a singular author responsible for plotting out all the actions and events of all the characters.

There is another way to generate stories: simulate characters. In a character-centric simulation, each character is an autonomous agent with its own independent reason. They respond to their environment and to other characters based on their own beliefs, desires, and intentions. One example of this is the system by Cavazza, Charles, and Mead(2001), which used a modern hierarchical task network planner. Each agent had a goal and formed its own plan. Those plans may come into conflict, requiring one agent to replan (which can be done very quickly in HTN planning). This system was technically an interactive storytelling system because a user could inject new facts or beliefs into the story world, causing agents to replan in response.

The hierarchical task network for character agents in a story simulation.

Virtually any agent technology capable of responding to changes in the environment can be applied to character-centric simulation based approaches to story generation. One of the limitations of the simulation approach is that there can be few guarantees about what will happen when the agents start executing. Theoretically if the agents were models of actors instead of characters, then they would have some ability to make decisions based on storytelling principles (Louchart and Aylett 2007) instead of being fully operating from the perspective of the character (the difference between a model of James Bond vs. building a model of Sean Connery). There has been some work in trying to model improv theatre actors (Riedl, 2010; Magerko and O’Neill 2012).

Machine Learning Story Generation Approaches

In this section we explore machine learning approaches that do not use neural networks.

One of the challenges of the non-learning approaches is that a vast majority of non-learning systems either require schemas encoded into a symbolic format (e.g., hierarchical plot schemas, symbolic scripts, action schema with preconditions and effects, cases stored in symbolic formats, etc.). Machine learning can be used for knowledge acquisition.

The Scheherazade system (Li et al. 2012, 2013) maintains a memory of plot graphs which are partially-ordered graphs that represent the most likely ordering of events on a particular topic. For example, a plot graph for going to a restaurant might have key events such as ordering food, eating, paying, leaving, etc. The arcs are temporal order constraints, e.g., paying occurs before leaving. Plot graphs resemble scripts in spirit except (a) they are partially ordered instead of fully ordered, and (b) events are arbitrary strings. When Scheherazade is asked to tell a story about a topic but doesn’t have a corresponding plot graph, it crowdsources example stories and then learns a plot graph that takes the most often mentioned events and learns the most likely ordering constraints. The system doesn’t have a prescribed dictionary of what construes an “event” — an event is a cluster of semantically related strings. Once a plot graph is learned, the generation process involves selecting event nodes such that no temporal constraints are violated. Since events are clusters of semantically related sentences from the crowdsourced corpus, the final story can be produced by selecting a linear sequence of nodes and then selecting a sentence from each node.

A plot graph learned by Scheherazade for going on a date to a movie theatre.
A story generated by Scheherazade for the plot graph above.

Neural Story Generation Approaches

The past few years have seen a steady improvement in the capabilities of neural networks for text. The literature on neural network based story generation techniques is growing rapidly, which requires me to focus on just a few of the systems and works that I found notable at the time of writing.

Neural Language Models

A language model learns the probability of a token (or sequence of tokens) based on a history of previously occurring tokens. The model is trained on a particular corpus of text. Text can be generated by sampling from the language model. Starting with a given prompt, the language model will provide one or more tokens that continue the text. The prompt plus the continuation can be input into the language model to get the next continuation, and so on. Training a language model on a corpus of stories means the language model will attempt to emulate what it has learned from a corpus. Thus sampling from a language model trained on a story corpus tends to produce text that looks like a story.

Recurrent neural networks such as LSTMs and GRUs were initially explored for text generation (Roemelle and Gordon 2017, Khalifa et al 2017). Initially, if the story training corpus was too diverse, RNN-based language models would have difficulty producing sequences that didn’t appear too random. This is due to sparsity — stories are written to be unique from each other and any sentence or event in a story might be unique or nearly unique. This makes learning a language model on stories difficult.

Martin et al. (2018) addressed this by trying to reduce the sparsity of story corpora by converting stories from natural language to event tuples e=<s, v, o, p, m> where s is a subject, v is a verb, o is a direct object of the verb, and, optionally, p is a preposition and m is an additional noun that accompanies the preposition. For example, “Sally repeatedly shot her cheating boyfriend in the bedroom” would be represented as <Sally, shot, boyfriend, in, bedroom>. These terms were further generalized by replacing them with semantic word classes from WordNet or semantic verb frame classes from VerbNet, for example <person-1, murder-42.1, male.2, in, area.5>.

In theory this reduced the sparsity and it was shown that generators could learn to make better event sequences than when working with raw text. This created the problem that event tuples were hard for humans to read so they introduced a second neural network to convert event tuples back to full sentences. “Eventification” of sentences is a lossy process so this neural network was tasked with restoring information in a contextual way. This is a challenging problem tantamount to re-telling a story abstraction with natural language. An ensemble was found to work best (Ammanabrolu 2020).

The generation loop for Martin et al. (2018).

A single language model doesn’t distinguish strongly between characters, which are just tokens. The work by Clark, Ji, and Smith (2018) learn to represent different entities so that stories can be generated about different characters.

Controllable Neural Story Generation

One of the main limitations of neural language models is that they generate tokens based on a sequence of previous tokens. Since they are backward-looking instead of forward-looking, there is no guarantee that the neural network will generate a text that is coherent or drives to a particular point or goal. Furthermore, as the story gets longer, the more of the earlier context is forgotten (either because it falls outside of a window of allowable history or because neural attention mechanisms prefer recency). This makes neural language model based story generation systems “fancy babblers” — the stories tend to have a stream-of-consciousness feel to them. Large-scale pre-trained transformers such as GPT-2, GPT-3, BART, and others have helped with some of the “fancy babbling” issues by allowing for larger context windows, but the problem is not completely resolved. As language models themselves they cannot address the problem of forward-looking to ensure they are building toward something in the future, except by accident.

Tambwekar et al. (2019) use reinforcement learning to fine-tune a sequence-to-sequence language model to generate story continuations that move toward a given goal. Reinforcement learning, generally, is a technique that can be used to solve sequential decision-making problems. However the latent space of a language model is too large for true trial-and-error learning. The reinforcement learner is used as a non-differentiable loss function. But how does the reinforcement learner know whether the continuation is getting closer to a particular goal? The system extracts patterns of verbs from the story corpus, clusters them according to how far each verb is from the goal, and then rewards the language model when it produces a continuation with a verb in the next cluster closer to the goal.

Fine-tuning the event2event neural network from the Martin et al. (2018) framework.
Fine-tuning reward is calculated by analyzing how close verbs are to each other in the corpus.

While typically language models are given a prompt that is the first line of the story, neural language models can also be conditioned on plot points that need to be present in the story. This increases the probability that certain events will occur as the story generation progresses. The hierarchical fusion model (Fan et al. 2018) takes a one-sentence description of the story content and produces a paragraph. PlotMachines (Rashkin et al. 2020) conditions a generator on a set of concept phrases given by a user. However, instead of using this set of concept phrases as an outline to elaborate on, the system can decide for itself what order to introduce the concepts. Instead of being given a high-level plot, a system can first try to generate a high-level plot and then elaborate. The plan-and-write technique (Yao et al. 2019) also uses a two-level approach. Instead of a single high-level sentence, the plan-and-write system learns to generate a sequence of keywords. Each of these keywords is then used to condition a language model so that it produces content about that keyword. If the keywords present a coherent progression then it stands that the lower-level elaborated content will as well. The technique by Fan, Lewis, and Dauphin (2019) has one model trained on entities and relations extracted from stories and a second natural language generator conditioned on the outputs of the first model.

Stories generated by the hierarchical fusion model.
Stories generated by the plan-and-write system.
Illustration of inputs and outputs of the PlotMachines system.

Another way of controlling the direction of a story is to provide key points in the story and in-filling. Wang, Durrett, and Erk (2020) attempt to interpolate between a given beginning and a given ending of the story. However, the language model they use doesn’t know how to attend to the end prompt. Instead they generate a number of candidates starting with the beginning prompt and use a re-ranker to judge the overall coherence (beginning, middle, and end), taking the best. Ippolito et al. (2019) propose an in-filling approach where they use the given beginning and ending to a story and generate keywords that might be found in the middle of the story. The keywords are then used to condition a language model that generates the middle of the story.

Neuro-Symbolic Generation

One of the issues with neural language models is that the hidden state of the neural network (whether a recurrent neural network or a transformer) only represents what is needed to make likely word choices based on a prior context history of word tokens. The “state” of the neural network is unlikely to be the same as the mental model that a reader is constructing about the world, focusing on characters, objects, places, goals, and causes. The shift from symbolic systems to neural language models shifted the focus from the modeling of the reader to the modeling of the corpus. This makes sense because data in the form of story corpora is readily available but data in the form of the mental models readers form is not readily available.

Assuming the theories about how reader mental models can be represented symbolically are correct, can we build neurosymbolic systems that take the advantages of neural language models and combine them with the advantages of symbolic models? Neural language models gave us a certain robustness to a very large space of inputs and outputs by operating in language instead of limited symbols spaces. But neural language model based story generation also resulted in a step backward from the perspective of story coherence. Symbolic systems on the other hand excelled at coherence through logical and graphical constraints but at the expense of limited symbol spaces.

Lara Martin (Dissertation, 2021) proposes to take a neural language model such as GPT-2 and constrain it with reasoning about causality. GPT-2, as a language model, can generate story continuation probabilistically based on a history of prior word tokens. Martin’s system parses the generated continuation sentence and uses VerbNet to infer what would be known to readers about the preconditions and effects of the sentence. If the preconditions of the continuation are not supported by the effects of prior sentences in the story, the continuation is rejected. If the preconditions are supported, then the effects of the sentence are used to update a set of logic-like prepositions that describe the world. These prepositions, which are drawn from VerbNet exactly because they are based on what readers can infer from reading a sentence, can be thought of as a simple reader model. By tracking the hypothetical reader’s model this system is less likely to make transitions from one event to another that do not make sense to the reader.

The neurosymbolic architecture by Martin. The World Engine maintains a set of propositions about the story world.

The CAST system (Peng et al. 2021) works in a similar to the above neuro-symbolic generation system. CAST uses the COMET commonsense inference model as the world engine. It infers the needs and wants of the characters in the story and attempts to match against the wants and needs of the characters from previous events. Character goals and intents are another way beyond preconditions and effects that readers track the coherence of stories.

The CAST pipeline.

Other Neural Approaches

Directly sampling continuations from a language model is not the only plausible way of using neural networks to generate stories. One might imagine search-like algorithms that use neural networks as resources for making decisions.

A lot of storytelling involves commonsense knowledge. Commonsense knowledge, or more specifically shared knowledge, is knowledge that one can assume is held by most people. To create a story that will make sense to a human reader, it might be advantageous to leverage commonsense knowledge. Neural networks such as COMET and GLUCOSE are trained to take input sentences and make inferences about what humans might also infer. The C2PO system (Ammanabrolu et al. 2021) takes a starting event and ending event and attempts to fill in the middle of the story. Instead of using a language model, it uses COMET to infer what most humans might believe comes next after the start and what might come before the end. This process is repeated, creating a directed acyclic graph of plausible successors and predecessors until a complete path is found from start to end.

The graph built by C2PO. 1) the initial event. 2) the final event. 3) the event found that bridges the forward and backward sub-graphs.

Example stories generated by C2PO.

The earlier-mentioned work on narrative psychology established the linkage between story coherence and question-answering. Another approach to story generation is thus to treat a language model as a resource to be queried in order to answer questions about a story world — the answers to these carefully chosen questions become the content of the story itself. Castricato et al. (3rd Workshop on Narrative understanding) explore this. They generate a story backwards, starting with a sentence about the end of a story. They generate some plausible questions about how the story might have gotten to that ending. They train an explanation generation language model to answer the question and then make the resulting explanation the previous segment of the story. They then repeat this process using the last fragment as the source of questions. Theoretically, this results in more coherent stories because each segment added to the story explains what comes next.

Stories generated via question-answering. The bold text is given input.


The field of automated story generation has gone through many phase shifts, perhaps none more significant than the phase shift from non-learning story generation systems to machine learning based story generation systems (neural networks in particular).

Symbolic story generation systems were capable of generating reasonably long and coherent stories. These systems derived much of their power from well-formed knowledge bases. But these knowledge bases had to be structured by hand, which limited what the systems could generate. When we shifted to neural networks, we gained the power of neural networks to acquire and make use of knowledge from corpora. Suddenly, we could build story generation systems that could generate a larger space of stories about a greater range of topics. But we also set aside a lot of what was known about the psychology of readers and the ability to reason over rich knowledge structures to achieve story coherence. Even increasing the size of neural language models has only delayed the inevitability of coherence collapse in stories generated by neural networks.

A primer such as this one makes it easier to remember the paths that were trodden previously in case we find opportunities to avoid throwing the baby out with the bath water. This is not to say that machine learning or neural network based approaches should not be pursued. If there was a step backward it was because doing so gave us a powerful new tool with the potential to take us further ahead. The exciting thing about working on automated story generation is that we genuinely don’t know the best path forward. There is a lot of room for new ideas.

This is an updated version of a piece originally released on Medium.

Further Reading

Amal Alabdulkarim, Siyan Li, and Xiangyu Peng Automatic Story Generation: Challenges and Attempts Proceedings of the 3rd Workshop on Narrative Understanding (2021).

Pablo Gervas. Computational Approaches to Storytelling and Creativity. AI Magazine. Vol. 30 No. 3: Fall 2009.

Mark Riedl, Vadim Bulitko. Interactive Narrative: An Intelligent Systems Approach. AI Magazine. Vol. 34 No. 1: Spring 2013.

Author Bio
Mark Riedl is an Associate Professor in the College of Computing and Associate Director of the Machine Learning Center at the Georgia Institute of Technology.  Dr. Riedl's research is on human-centered artificial intelligence and machine learning, specifically focusing on ethics, safety, explainability, storytelling, and computer games. Dr. Riedl earned a PhD degree in 2004 from North Carolina State University. From 2004 to 2007, Dr. Riedl was a Research Scientist at the University of Southern California Institute for Creative Technologies.  Dr.  Riedl joined the Georgia Tech College of Computing in 2007.

Contributing authors (alphabetical): Amal Alabdulkarim, Louis Castricato, Siyan Li, and Xiangyu Peng.

For attribution in academic contexts or books, please cite this work as

Mark Riedl, "An Introduction to AI Story Generation", The Gradient, 2021.

BibTeX citation:

 author = {Riedl, Mark},
 title = {An Introduction to AI Story Generation},
 journal = {The Gradient},
 year = {2021},
 howpublished = {\url{ } },