Ordering Events in Interactive Fiction Narratives

An interactive fiction system has been developed with a rich representation of simulated locations, actors, and things as well as events. This paper discusses one particular type of narrative variation that the system can generate: variation in order. To determine how to concisely specify a possibly non-chronological order for narrating events, a formalization of Genette’s categories of order and his concept of the time of narrating is developed. An ordered tree representation for reply structures is introduced that uses Richenbach’s concepts of speech time, reference time, and event time to determine grammatical tense. Varying the Narrative Discourse in Interactive Fiction Interactive fiction (typically abbreviated IF) is a venerable form of computer amusement. Some, including this author, believe it holds further literary and gaming promise (Montfort 2003). It is certainly rich as a platform for researching narrative text generation. There is a good deal of hand-crafted prose in existing IF, but a simulated world forms the basis for the textual exchange between user and program, providing a foundation for the generation of narratives. There is also an established form of IF interface that allows meaningful, ontological interaction. The standard IF world model is simple enough to be worked upon by a single author without commercial backing yet complex enough to provide compelling experiences. Interactive fiction has objects and characters which are positioned in simulated space; simulated incidents involving these can happen. Current IF systems have not provided any facilities for arranging the way these incidents will be told, however. This paper deals with one particular capability of an IF system called nn. This system was developed to address the level of the telling, between underlying events and textual output. Its architecture has been described elsewhere (Montfort 2006, 2007). A detailed and systematic treatment of how the telling of a narrative can be considered apart from the existents and events represented in it is found in Gérard Genette’s Narrative Discourse: An Essay in Method (published in French in Figures III in 1972; English translation, 1980). In this discussion, Genette covers variations in temporal relationships or tense: how events can be narrated in a sequence that differs from their chronological sequence (order), how the telling can move morCe rapidly or more slowly and narrate events for shorter or longer periods of time (duration, later called speed), and how events can be narrated once each, one for several occurrences, or several times for each occurrence (frequency). Genette continues to consider the narrative analog of what is called in grammar mood, which includes the perspective from which a narrative is told (focalization). As part of the final category, voice, the time of narrating and its function in narrative is discussed. While the basic concepts have been extended, revised, and further discussed in various ways, the general outline of relationships was groundbreaking and has proven essential to decades of development in narratology. The concepts introduced in Narrative Discourse and developed by other scholars pertain to the form and function of narrative and have developed from formalist and structuralist ideas, so they are formal in this sense. To implement different types of narrative variation in an interactive fiction system, however, additional refinement and formalization of narratological concepts is needed: In general, most Humanities models of narrative contain formalizations only at very abstract levels, if at all. By formalizations, we mean here a representation in some logic language (e.g., predicate calculus) or other structured representation, including tables, graphs, etc. Indeed, most works dealing with narrative and not going back directly to the structuralist tradition are composed in “plain prose”. Especially, there seems to be a tendency to apply formal notions to the abstract histoire level only. Phenomena at discours level that apply to the structure of discourse (e.g., discourse relations) are sometimes formalized in linguistics and are usually described in words only — sometimes accompanied by tables — by literary scholars (Genette, 1980). Where models are based on the discours (text) layer of a narrative or include it, genuine Humanities models usually lack formality, though their descriptions might offer a variety of authentic examples. (Gervás et al. 2006) The Narrator module of nn is a text generation system that follows a standard three-stage pipelined architecture (figure 1). The focus of this paper is on the first stage, the Reply Planner, where content selection and ordering are done. To achieve the sort of formalization necessary for implementation in a computer system, the inputs to the Reply Planner, the internal operations of the Reply Planner, and the output from the Reply Planner to the Microplanner are defined in detail alongside the concepts of narrative discourse that these operations are based on. The lowerlevel work of the Microplanner, and the lowest-level work of the Realizer, while important, are not covered this paper. The input to the Reply Planner consists of a focalizer world, a set of indices to actions indicating what in that focalizer world has transpired in the most recent turn, and a plan for narrating. The focalizer world is a representation of the IF actual world from the standpoint of a particular actor. It includes existents, along with the capability to roll back to a point in the past and see what existents were like at that point. It also includes events with causal connections between them and temporal information about each. The Reply Planner uses this input to build an ordered tree called a reply structure (RS), with proposed expressions (PEs) as leaves. The PEs indicate how the narration of an event, the description of something in the content plane, or the creation of some non-diegetic text is to be done. For instance, a standard transformation to produce the sort of narration often used in existing IF would result in chronologically-ordered PEs being placed in an RS of depth 1; each PE would be marked with the default speed (.5). For every PE that is in the output RS, some text will be generated — all content selection is done in the Reply Planner, and nothing selected at that point may be elided at a later stage. The details of what text is generated from PEs are handled by the Microplanner and the Realizer. The Microplanner will output a longer abstract paragraph or sentence representation (to whatever extent this is possible) when the speed is slower and less when it is faster. In the last stage, the Realizer, the abstract representations provided by the Microplanner are converted into strings of English for formatting and output.