How do you pronounce biopic, synod, and Breughel? - and why? Do our cake and archaic sound the same? Where does the stress go in stalagmite? What's odd about the word epergne? As a finale, the author writes a letter to his 16-year-old self.
SUMMARY This book presents core concepts and insights from humanities narratology, such as time, plot, and the narrator, for automatic narrative understanding and generation in natural language processing and artificial intelligence, including interactive entertainment. It also describes how computational developments may also contribute to narratological theory and digital humanities. The book is mainly a primer on narratology and a review of computational developments for computational linguists and game developers, but may also be of interest to cognitive scientists and narratologists interested in applications.
Chapter 1, Narratological Background, provides an overview of the main narratological concepts addressed and establishes terminology. Mani defines ‘narrative’ with a pre-theoretic notion of ‘story,’ to include parts of blogs, emails, news, and technical and literary works, among others, that involve storytelling with causal coherence. The relevant modes of presentation (or media) also include games, as game developers with an interest in interactive narrative are among the primary intended audience for the book. The author also defines ‘Narrative Structure’ to encompass global dimensions of phenomena related to narrative representation, including narrator and audience variables. In this way, his view of narrative structure goes beyond discourse formalisms such as Discourse Representation Theory (Kamp, 1984) and Rhetorical Structure Theory (Mann & Thompson, 1988), which focus on relations of coreference, time, and/or communicative roles at a more local level. One of the most important distinctions that Mani makes -- relevant throughout the book -- is between ‘fabula’ as the underlying (story) content and ‘discourse’ as its expression or form. In a section on narrator characteristics, he introduces Genette’s (1980) narratological concepts of a homodiegetic narrator (who participates in the story) and a heterodiegetic narrator (who doesn’t), as well as the narrative distance continuum (Genette, 1988), which is represented in Mani’s annotation scheme, NarrativeML, with six values, ranging from narrated speech to immediate speech. Mani also introduces different cases of narrator perspective: non-focalized (omniscient), internally-focalized, and externally-focalized (Genette, 1980), with room for ‘other’ possibilities in NarrativeML. He also discusses embedded stories, as in “The Thousand and One Nights,” in which readers often have to revise their beliefs about the factual status of characters or events. Subordinated discourse with propositional attitudes also helps determine the factual status of events. With regard to narrative time, Mani introduces Genette’s (1980) narrative time relations -- subsequent, simultaneous, and prior -- indicated by the dominant tense in the narrative. He also mentions Chatman’s (1980) distinction between story time (in the fabula) and discourse time -- discourse processing time in Chatman’s original formulation, but operationalized as textual length for Mani’s NarrativeML. Mani uses the ratio between the two to represent narrative pace, which indicates pacing techniques such as stretching and speed-up. Seven types of narrative ordering from Genette’s (1980) account -- chronological (‘Chronicle’) and others -- are also represented in NarrativeML, indicating the relationship between discourse order and the actual order of events in the fabula. The issue of audience response, or reader affect, is also mentioned. After a brief discussion of the ontology for a fabula or story world, Mani also introduces accessibility relations, used for representing narrative embeddings and subordinated discourse with propositional attitudes while avoiding the representation of multiple worlds (multiple story worlds and the actual world). Changes in accessibility relations can represent the audience’s belief revision, e.g., in a plot twist. The chapter ends with a description of NarrativeML, an annotation scheme that incorporates many of the narratological constructs. Such a scheme can be used for human annotation of corpora, and the annotated corpora can then be used for supervised learning of systems for automatic narrative understanding and generation. The possibility of importing other annotations from the PropBank (Palmer, Gildea & Kingsbury, 2005) and other schemes is also discussed.
In Chapter 2, Characters as Intentional Agents, Mani presents the ‘planning perspective’ on narrative, in which the task of narrative understanding is to recognize the characters’ goals and plans, and that of narrative generation is to synthesize a series of plans. Since there have been more developments in AI on intentionality than in humanities narratology, Mani focuses more on the AI side in this chapter while suggesting how these developments may provide insight for narratology. Interpreting actions and events in terms of goals and plans presumes world knowledge. Preconditions and consequences for actions and stereotypical structures of event sequences have been represented in narrative understanding systems, such as those based on Schank and Abelson’s (1977) scripts, but Mani points out that these systems were too domain-specific and focused too much on events rather than characters. He also mentions the use of case-based reasoning in both understanding and generation, which allows generalization from stored narrative fragments beyond simple retrieval. In particular, he points out some general limitations in story planning systems, such as domain specificity, and lack of natural variation in narrative order, distance, etc. For purposes of ‘lightweight’ annotation with NarrativeML, Mani adopts Pavel’s (1982) Move-grammar, which provides a coarse-grained analysis of goal structure in the fabula -- often corresponding to long stretches of text. Pavel’s analysis decomposes an action into a problem and a solution. In the section on interactive narrative, which often involves incremental, non-monotonic plan revision after new events, Mani discusses Mateas and Stern’s (2005) FAÇADE to illustrate dynamic re-planning in reaction to audience feedback, but points out the problem of a large number of branching possibilities having to be spelled out. He suggests that a good user model for reader affect or aesthetic preference may help constrain possible paths in interactive narrative generation, proposing a Boolean model of the reader’s attitude toward the agent of an event -- positive (sympathy), negative (antipathy), or neutral. Mani also mentions the possibility of using sentiment analysis for a model of narrator attitudes as well, though not represented in his NarrativeML. He also discusses the problem of balance between authorial and audience control in interactive narrative. The representations of intentionality discussed in this chapter are closely tied to Chapter 4, Plot.
In Chapter 3, Time, Mani focuses on narrative understanding for story time. After a quick review of rule-based systems making use of tense and aspectual marking or causal knowledge, he introduces Allen’s (1984) interval calculus with seven basic temporal relations, which is adopted in NarrativeML and can be extended with logics for branching time for underspecified relations. Mani then discusses TIMEX2 (Ferro et al., 2005) and TimeML (Pustejovsky et al., 2005) annotation schemes for tagging duration, end time, and relations among events and times, in which subscript indices indicate the order of mention, allowing inferences about narrative ordering in relation to the actual order of events in the story world. Subordinating links, or SLINKs, for relations such as remembering require branching time models. Discourse time, measured in number of words, can be compared to story time in the fabula to measure narrative pace, or tempo. The author also mentions an interesting possible extension with estimates of minimum and maximum duration of events to represent commonsense intuitions (Pan, Mulkar-Mehta & Hobbs, 2011). Issues in human annotation with temporal links among times and events, or TLINKs, are discussed. Successful automatic tagging in these temporal aspects of narrative can facilitate narratological investigations. In a brief section on automatic narrative generation, Mani describes two particular systems, Callaway’s (2000) STORYBOOK and Montfort’s (2011) CURVESHIP. Back to narrative understanding, Mani reviews recent success in automatic tagging and resolution of temporal expressions, as well as recent developments in automatic tagging of events, factuality, coarse-grained duration, and TLINKs. For temporal relation classifier systems based on local pair-wise decisions, the problem of global inconsistency may arise. Solutions combining a ranking method with Integer Linear Programming or Markov Logic Networks are discussed. Habituals and scene-setting descriptions also pose a challenge to narrative time representation.
In Chapter 4, Plot, Mani provides background on important concepts, including abstract event summaries and Aristotelian mythos, based on a view of plot as a compact structural unit with emphasis on event sequences. Other narratological concepts regarding plot, such as a turning point, a narrative arc with stages, the heroic quest, and Propp’s (1968) narrative functions, are discussed, along with applications to interactive narrative. Overly fine-grained distinctions are often impractical for reliable annotation, but the use of more global structural representations based on story grammars (Rumelhart, 1977) and Macrostructures (van Dijk, 1980) in automatic story generation is discussed next. In this section, Mani repeats the distinction between story time at the level of fabula and discourse time as text length, pointing out that generation systems often fail to decouple the two and claiming that Pavel’s Move-grammar adopted into Mani’s NarrativeML may improve upon that aspect. Mental states and intentionality in plot are then emphasized, with descriptions of previous accounts involving affect states (+, -, and a neutral M; Lehnert, 1981), which capture motivations and intentional actualizations. Despite the coarse-grained differentiation of emotional states, classifying the (emotional) polarity of verbs based on their arguments may improve representation of causality (Goyal, Riloff & Daumé III, 2010). Recurring patterns of transitions are considered plot units, and applications including Elson’s (2012) Story Intention Graphs for DramaBank are discussed. Acknowledging the inherent difficulty of inferring intentional states that are not directly expressed, Mani mentions event summarization approaches, such as Chambers’ (2011) Narrative Event Chain, combined with salience filtering or event-based causal reasoning algorithms as a potentially more feasible alternative. After a quick overview comparing the different plot models discussed in the chapter, Mani suggests some applications for narratology, such as intelligent searching for stories with similar plots.
In the final chapter, Summary and Future Directions, the author summarizes each previous chapter, and illustrates the representation of the major aspects of a narrative in NarrativeML with an example (pp. 96-99), mentioning the problem that long literary genres with substructural units such as scenes or episodes would pose a challenge to annotation efforts. In his concluding remarks, Mani speculates that developments in narratological theories of character psychology may continue the tradition of narratological insights for computational applications, and also hints at the opposite direction of inspiration, namely, automatic narrative computing systems inspiring developments in digital humanities with enhanced search, analysis, translation, clustering, and recommendation, among others.
EVALUATION This book offers an easy-to-read introduction to the core issues in narrative representation, both traditional narratological insights and more recent computational developments. Mani defines the terminology and explains his choice of terms carefully, and suggests promising interdisciplinary contributions between humanities narratology and computational narratology throughout, achieving his main goals. He concludes each chapter with a clear illustration of how the aspects of narrative discussed in the chapter are represented in NarrativeML.
The book is primarily intended for computer scientists working on narrative processing and generation and for narrative theorists interested in applications, and its direct relevance to other fields of cognitive science, including formal semantics, seems more limited than the back cover suggests. For example, Mani himself points out that he is more concerned with an entire story at a more global level, compared to the more local focus of Discourse Representation Theory or Rhetorical Structure Theory. In addition, although some interesting experimental findings (e.g., Gerrig & Bernardo, 1994) and human annotation studies (e.g., Pan, Mulkar-Mehta & Hobbs, 2011) are mentioned, discussion of relevant findings in cognitive psychology (e.g., Graesser, Singer & Trabasso, 1994, on causality and intentionality; Zwaan & Radvansky, 1998, on the situation model of discourse) is otherwise lacking. Comparing NarrativeML representation to Zwaan and Radvansky’s (1998) situation model reveals that the core dimensions of narrative align well between the two models, but that in NarrativeML spatial relations are not represented to a level of granularity that the situation model would predict or to the same level of granularity as temporal relations in NarrativeML. Incorporating Mani’s recent work (Mani & Pustejovsky, 2012) on spatial representation may be useful for narrative representation as well. Another possible addition to NarrativeML is to represent character prominence, which has been found to be important for narrative production (e.g., Sanford, Moar & Garrod, 1988).
Narrative computing is an exciting field, very much burgeoning, and it will be interesting to see how Mani’s proposals, e.g., the use of Pavel’s (1982) Move-grammar for representing character goals, stand the test of time.
REFERENCES Allen, J. (1984). Towards a general theory of action and time. Artificial Intelligence, 23(2), 123-154.
Callaway, C. (2000). Narrative prose generation. Doctoral dissertation, North Carolina State University.
Chatman, S. (1980). Story and discourse: Narrative structure in fiction and film. Ithaca: Cornell University Press.
Ferro, L., Gerber, L., Mani, I., Sundheim, B. & Wilson, G. (2005). TIDES 2005 standard for the annotation of temporal expressions. http://timex2.mitre.org/annotation_guidelines/timex2_annotation_guidelines.html
Genette, G. (1980). Narrative discourse (J. Lewin, Trans.). Ithaca: Cornell University Press.
Genette, G. (1988). Narrative discourse revisited (J. Lewin, Trans.). Ithaca: Cornell University Press.
Gerrig, R. & Bernardo, D. (1994). Readers as problem-solvers in the experience of suspense. Poetics, 22, 459-472.
Goyal, A., Riloff, E. & Daumé III, H. (2010). Automatically producing plot unit representations for narrative text. In Proceedings of EMNLP 2010 (pp. 77-86). Cambridge, MA.
Graesser, A. C., Singer, M. & Trabasso, T. (1994) Constructing inferences during narrative text comprehension. Psychological Review, 101(3), 371-395.
Kamp, H. (1984). A theory of truth and semantic representation. In J.A.G. Groenendijk, T.M.V. Janssen & M.B.J. Stockhof (Eds.), Truth, Interpretation, and Information (pp. 277-322). Dordrecht: Foris.
Lehnert, W. G. (1981). Plot Units: A narrative summarization strategy. In W.G. Lehnert & M.H. Ringle (Eds.), Strategies for Natural Language Processing. Hillsdale, NJ: Lawrence Erlbaum.
Mani, I. & Pustejovsky, J. (2012). Interpreting motion: Grounded representations for spatial language. New York: Oxford University Press.
Mann, W.C. & Thompson, S.A. (1988). Rhetorical structure theory: Towards a functional theory of text organization. Text, 8(3), 243-281.
Mateas, M. & Stern, A. (2005). Structuring content in the Façade interactive drama architecture. In Proceedings of the AIIDE 2005 (pp. 93-98). Menlo Park, CA: AAAI Press.
Montfort, N. (2011). Curveship’s automatic narrative variation. In Proceedings of the 6th International Conference on the Foundations of Digital Games (pp. 211-218).
Palmer, M., Gildea, D. & Kingsbury, P. (2005). The Proposition Bank: A corpus annotated with semantic roles. Computational Linguistics, 31(1), 71-105.
Pan, F., Mulkar-Mehta, R. & Hobbs, J.R. (2011). Annotating and learning event durations in text. Computational Linguistics, 37(4), 727-752.
Pavel, T. (1982). Plot-structure and style: Remarks on an unstable relationship. Canadian Review of Comparative Literature, 9(1), 27-45.
Propp, V. (1968). Morphology of the folktale (L. Scott, Trans.). Austin: University of Texas Press.
Pustejovsky, J., Ingria, B., Saurí, R., Castano, J., Littman, J. Gaizauskas, R., Setzer, A., Katz, G. & Mani, I. (2005). The specification language TimeML. In I. Mani & J. Pustejovsky (Eds)., The language of time: A reader (pp. 545-558). New York: Oxford University Press.
Sanford, A.J., Moar, K. & Garrod, S.C. (1988). Proper names as controllers of discourse focus. Language and Speech, 31(1), 43-56.
Schank, R.C. & Abelson, R.P. (1977). Scripts, plans, goals, and understanding: An inquiry into human knowledge structures. Hillsdale, NJ: Lawrence Erlbaum.
Zwaan, R.A. & Radvansky, G.A. (1998). Situation models in language comprehension and memory. Psychological Bulletin, 123(2), 162-185.
ABOUT THE REVIEWER:
Choonkyu Lee is a postdoctoral researcher at the Utrecht Institute of Linguistics OTS. His research interests include time in narrative discourse and commonsense knowledge in semantics/pragmatics, with interdisciplinary approaches involving cognitive psychology, theoretical linguistics, and computational linguistics.