Author: Nick Montfort

This is the fourth in a series of posts about my interactive fiction system, Curveship.

I was recently asked to elaborate on the difference between nn (the research system I developed during my dissertation work at the University of Pennsylvania) and Curveship.

The most important difference is that nn is a research system that I used for making some advances related to computer science, computational linguistics, and narratology. The system was developed to prove certain points; it was used only by yours truly to implement narrative and text generation ideas and to run demos.

It worked pretty well for this. Using nn, I developed a new representation for the order of actions, one that can be used flexibly in text generation. I put together a string-with-slots representation used in templates, one that is much less cumbersome to use than a general abstract syntax representation. And I worked out how a focalizer world (now called a “concept”) can model an actor’s theory about the “real” or “actual” world that is central to an interactive fiction.

That’s not to say that you’d want to pick up this system, though, and start writing games in it. While nn in 2007 had all the basic components of an IF system (a parser, the ability to save an restore games, the ability to model a world), they were not all ready for prime time, because my development work was focused on trying to answer particular research questions.

So, I took the nn code and began working on a new project, Curveship. I wanted to release a system that was developed with other IF authors, other AI researchers, and other teachers of narrative theory and electronic literature in mind. Curveship is getting close to being this system – within a few months, I hope, I will get it to that point, with the help of a group of people, the Curveship crew, who are testing and commenting on pre-release versions.

Concretely, there are three main sorts of differences. First, bugs in nn that my limited use of the system didn’t expose have been fixed – and more will be. Second, there are improved capabilities that didn’t relate to my research goals, but which are important to IF authors. These range from better parsing and clarification (disambiguation) to better ways of formatting text to better capabilities for simulating physical aspects of the world. Finally, I have impeoved the interface, which includes changing the names of some components of the system. What used to be a “plan for narrating” is now “spin.” What I called “existents” (a perfectly nice but abstruse term from narratology) are now, more simply, called “items.” To avoid confusion, I even changed the name of the “Narrator” module, because “narrator” is a narrative function and a parameter that can be set within the system. This module is now called the “Teller.” There have been other name changes and a great deal of refactoring, some writing of docstrings, and other improvements to make the Curveship code more understandable and manipulable to those who want to get into it. The way string-with-slots templates are written has changed, too, as has the way items and actions are created.

The discussion of reply structures, focalizer worlds, and similar aspects of the system in previous publications (such as my dissertation) remains largely applicable to Curveship. But because Curveship is being developed with an eye toward a new group of author/programmers, researchers, and teachers, there have been many changes beyond the parts of nn that implement those core research contributions. Even some of publications and talks about Curveship from recent months have names for components that are now different and mention the old style string-with-slots formalism. That’s one of the reasons I’ve started in on this series of posts.

This is the third in a series of posts about my interactive fiction system, Curveship.

Before I start descending into detail, I’ll explain why I think Curveship is a big deal.

Curveship does the usual work of an interactive fiction system when it comes to simulating a world: There are discrete rooms that make up the fiction’s locations, actors can inhabit and wander around these rooms, and things can sit in them, be taken and carried off or otherwise moved around. Items can change state, so that a lamp, for instance, can be turned on and off. Items can go into or onto other items, if they allow it. None of this is surprising; plenty of interactive fiction development systems already do all of it very well.

Hypertexts don’t do that. Hypertexts, including Web hypertexts but also pieces in HyperCard, Storyspace, Twine, and similar systems, present a structured document in which some phrases are hooked up to other pages. They might have guard fields or other pieces of code, but at their most essential, they are really strange surfaces of text, with phrases connected in an unusual topology.

Nothing’s wrong with hypertext. If you want to create an unusual language surface, to link pages together into a curious structure, it’s clearly the form to use. The play and connection of language can be intriguing and relevant. You can make a compelling literary work in this form, as several people have done.

But interactive fiction systems allow something else. You can put a simulated world underneath the text that people see. This lets you create different types of literary systems, which can be nice: Several people have created interactive fiction systems that are compelling literary pieces and that couldn’t have been realized as hypertexts.

In an important way, Curveship is to interactive fiction what interactive fiction is to hypertext. Just as interactive fiction added a new and significant model – a model of a simulated world – Curveship adds a model of the telling, of the narrating, of the way that the simulated world is represented in language. Curveship adds spin, as in the spinning of a tale and as in Spin Alley.

Curveship can leave actions out of the narrative, can reorder actions in the narrative to create flashbacks and flashforwards and retrograde narration and narration by category, can narrate at greater or lesser length, and can represent several actions in one statement or repeat the narration of a single action several times. The system can tell about what is happening in the simulated world from the perspective of any specified actor, based on that actor’s knowledge and perceptions, or it can narrate from a global perspective. It can tell about what has happened in the past just as easily as it can represent a current action.

The system is able to do this because it models every action that happens in the world, just as existing IF systems model every item. First-order representations of actions differentiate Curveship from other IF systems. And on that of that, a complete text generation system has been built, allowing for grammatical changes that follow from high-level narrative parameters – the spin.

Consider The Odyssey. Why is it interesting? Why has it shaped world culture and remained memorable for thousands of years? Yes, lots of interesting stuff happens in it. But the way it is told is also tremendously compelling – Odysseus listening to the song of his own exploits, shielding his eyes and weeping, being identified as the great hero, and then going on to tell the next part of the story in his own voice. That’s as fascinating a part of The Odyssey as is the encounter with the Cyclops or the slaying of the suitors. And the telling surely becomes more important in contemporary literature. Is Lolita really more interesting because of what happens or because of how it is told? What about Ulysses? To the Lighthouse?

Curveship’s modeling of the telling adds to the significant power that IF already provides by way of a model of a simulated world. Just as IF adds a simulated world to hypertext, Curveship adds control over the telling (that is, it adds spin) to the simulated world.

Those who want to create strange surfaces of text will want to work in hypertext. Those who want to have a simulated world will probably use an existing IF system. For those who are interested in varying the telling, as many great authors have done in oral, written, typed, and word-processed literature, it will be best to use a simulated world and spin: Curveship.

This is the second in a series of posts about my interactive fiction system, Curveship.

In writing about Curveship in any detail, I’ll have to use terms such as action, event, and order, which sound ordinary but are used in a special way in the system. Furthermore, I’ll have to use terms such as focalization and narratee, which do not sound ordinary, but have a meaning within narratology (a.k.a. narrative theory) and are important to the way Curveship works. I’m going to define a few of these terms – some I’ll save for later. Rather than sort them alphabetically, I’ll group them by how they figure in the system.

Worlds of Curveship

Actual World. Following Marie-Laure Ryan’s terminology, I consider the main, simulated interactive fiction world to be the “IF actual world,” that is, what is “actual” from the standpoint of the player characters and other actors. I’ve bothered to give this a special name (rather than just calling it the world) because it isn’t the only ball in play in Curveship.

Concept. A concept, formerly called a “focalizer world,” is a particular actor’s theory of the actual world, based on knowledge (put into a fiction file to begin with) and perceptions (gained as the actor moves around and looks at things). Concepts are almost never complete, and they may be wrong. They allow the telling of actions and the description of items to be focalized, that is, based on what a particular actor knows and sees.

What Exists in the World

Item. Basically, any object that has a representation. I could have called them “objects,” but that term means something different in object-oriented programming, even though items do end up being represented as objects. I could have called them “existents,” the term Seymour Chatman and other narratologists use, but that caused gnashing of teeth among some of the people I consulted, including one researcher whose first language wasn’t English. Items must be in one of three categories.

Actor. An item that can take action on its own, due to either code that an IF author has written or a script the author has dropped in. Any item can react, but only an actor can initiate action. Actors are also the only category of item that have concepts.

Cosmos. The special actor that is the root of the item tree. Responsible for earthquakes, power outages, and any occurrence where the author doesn’t want to model the cause as its own actor. The cosmos can also change the plan for narrating, which allows for a connection between the simulated world and the way the telling of this world is done.

Room. A discrete location which has exits leading to other rooms. Rooms are all on the first level of the item tree, directly below cosmos.

Thing. Generally, an inert item. Any item that isn’t a room and doesn’t need to act or have a concept is a thing. Things can react when something is done to them or done in the same room, so that pressing a switch on a lamp can cause the lamp to react by increasing its glow.

Item Tree. A structure representing all of the items in the actual world (or in a concept) and the relationship between them. It is a graph with a designated node (cosmos), directed links from parent to child, and no loops – that is, no item is its own descendant. Hence, a tree.

Parent-Child Relationship. The relationship between an item and an item a level under it, and connected by a link, in the tree. When an actor walks into a room, he or she becomes the child of the room. If the room moves (perhaps it’s an elevator) the actor will move with it. If the lights in the room become brighter, the actor and everything else in the room will be better illuminated. An apple placed in a sack similarly becomes the child of the sack and, for instance, is stolen if the sack is stolen.

Link. The edges (as they are called in graph theory) of the item tree, each one connecting an item to another item and indicating a parent-child relationship. Links are labeled, with the label indicating more about the relationship. If a person is holding a sack with an apple inside and wearing a cloak, the sack and cloak are the children of the person and the apple is the child of the sack. Furthermore, the link between the sack and the person is labeled “of” (indicating a possession) and the link between the cloak and the person is “on” (indicating something being worn). The link between the apple and the sack is “in” (indicating containment). If you have ever tried to use Inform 6 to model a desk that you can put things on (on top of) and in (in a drawer), something which is not very straightforward to do, you will see an advantage to the way Curveship works.

What Happens in the World

Action. A specific but fairly high-level action taken by an actor. An intended command (such as “get lamp”) usually corresponds to a single action. It also usually doesn’t make sense for an action to partly succeed; it either works or doesn’t, although the first components of it may succeed and later ones may fail. The set of actions is open. Authors can easily make up new ones. Actions are what the system narrates, because they are meaningful and high-level.

Event. A lower-level component of action. There are only five, representing different things to be done in the world. Any action is made of one or more of these with appropriate arguments. The five primitive events are sense, impel, misc, modify, and configure. Only the last two change the state of the actual world, modify by changing a feature of an item and configure by changing the item tree. Events are good for simulating the world – unlocking and opening doors, moving actors from room to room, and so on. But because they are so low-level, they are not very good as a basis for narration. We usually do not want to hear that an actor exerted force on a switch, then turned the lamp on, and then increased the glow of the lamp. But it’s nice to be modeling this under the hood, because if the lamp’s battery is dead, an actor might be able to do the first two things but not the third. If the battery is then replaced, the lamp should glow brightly right away, since the switch is in the on position.

Files of Curveship

Fiction File. What others call the “story file” or “game.” It can be a game, of course. Essentially, though, it defines a fictional (and simulated) world.

Spin File. A very short file with a few parameters and perhaps a bit of code, all of which could be included in a fiction file. These parameters determine how the story is told at a fairly high level. Released games probably won’t use spin files, but sets of these may be used in research and teaching dealing with narrative, because they allow different methods of telling to be applied to the same fiction file (or a set of fiction files).

These are the building blocks of the simulation done in Curveship. I should note that Curveship is designed to mainly simulate the physical world, not emotional states or conversations between actors. And, of course, it models the way actions are represented and items are described based on concepts of that world – it narrates in different ways based on parameters and code provided by IF authors. I’ll discuss how it does that in a future post.

This is the first of a series of posts about my interactive fiction system, Curveship.

Curveship is an interactive fiction development system that provides a model of a physical world, and its workings, as do existing state-of-the-art IF systems (such as Inform 6, Inform 7, TADS 2, and TADS 3). It will not have as many libraries, and will have no multimedia features, when it is released, but it will provide another significant capability: it will allow IF authors to write programs that manipulate the telling of the story (the way actions are represented and items are described) as easily as the state of the IF world can now be changed. While existing IF systems allow for the simulation of a character who can move around and change the state of the world, Curveship provides for control over the narrator, who can tell as if present at the events or as if looking back on them, who can tell events out of order, creating flashbacks or narrating what happens by category, and who can focalize any character, not just a fixed PC or a hard-coded series of them, to tell the story from the perspective of that character’s knowledge and perceptions.

Curveship will have a closed round of testing and revision soon, and will be publicly released soon thereafter. This may happen as early as the end of summer, but if that timing makes little sense for the project or the IF community (IF programmers may be too busy finishing their IF Comp games at that point) the release could happen later.

I have had a very short page about the system up for a while at Curveship.com. As it explains in slightly more detail, the system’s starting point was my too-cleverly-named research system nn. I have been working for a while to get a usable programming interface set, to clean up the code, and to add some capabilities that are necessary for a usable IF system but didn’t have much research value to me, at least initially. Finally, the name of the system comes from a poem by Hart Crane. “Curveship is being fashioned to model the essential qualities of variation—the curve of a story through its telling—just as friendship and authorship represent the essence of being a friend and author.”

I’ve recently given a talk about the system to a group of roboticists and computational linguists at a workshop, and more recently (yesterday) got to discuss the practicalities and details of the system a bit with the Boston IF group; those folks provided some good suggestions at the levels I’m most concerned about now. I’m now going to begin writing, as often as I can, about the system, what principles it is based on, and how it works at high and low levels.

Curveship is a Python framework; to write a fiction file you simply write some Python code, instantiating classes for the most part, writing your own subclasses if you want items to do something unusual. There are a few fairly radical differences between Curveship and existing IF systems, but the two that will probably be most visible are the modeling of actions and the string-with-slots representation that is used for text generation. In Curveship, items do not change their own state. Changes of state, and the movement of items from one place or container to another, are accomplished only by actions. And, when actions are narrated or objects described, this is done based on a formalism that is a bit harder to write than a simple string would be. These are two of the the main aspects of Curveship that, although they demand more work from the IF author, allow for powerful, general ways of narrating.

Ideally, as I continue these posts, I’d like to write in response to your particular interests and curiosities. I’ll be glad to answer questions in comments if there’s a reasonably sort answer; if not, I’ll try to answer them in the context of a longer post.