When interpreting a narrative in which someone tells a story, !rst-person accounts of events that happened to the narrator earlier are relatively easy to process. But cases in which the narrator is telling something that she witnessed - but was not personally involved in - require more elaborate pragmatic inferences based on the assumption that the narrator was present as a witness of the events being told. Appropriate handling of these cases requires means of correct interpretation of where events experienced and events witnessed - by the narrator - are happening in terms of locations in the storyworld. The present paper considers a simple model of how information on changes in location across events may be inferred from narrative discourse, and how this information can be exploited to ensure appropriate temporal placement - with respect to the events of the main or frame story - of the events for the embedded stories that were witnessed by a narrator at some earlier point in the discourse.
When we consider a story as a stand alone product we tend to disregard the context in which it is
produced: the story is usually an act of communication between a narrator and an audience [
Three topics are reviewed to inform the work reported: basic challenges of reading narrative that involves embedded stories, how to place embedded stories with respect to the frame story and how to infer changes in location from narrative discourse.
Psychologists studying what takes place in the mind of a reader processing !ction have identi!ed that a
fundamental part of the process is the construction of a physical model of the !ctional world in which
the story takes place [
Todorov [6] and Tenev [7] both identify modality as a fundamental ingredient in the construction
of narrative, with con"ict in stories often arising from contrast between actual and desired situations
or between opposing obligations/desires. Ryan [8] elaborates on this idea, describing narrative as a
sequence of moves that transition between states of the storyworld that include models of characters
views on wishes, obligations or beliefs. In both cases, the ability to model this kind of evolution on the
modal views of characters becomes a fundamental aspect for narrative interpretation. As the views of
characters on wishes, obligations or beliefs are often presented as embedded stories [
Gervás [9] proposed an algorithm for the interpretation of embedded stories that relied on on a stackbased mechanism for handling the changing contexts of interpretation. Rather than work on text, this algorithm operated on a simpli!ed format for conceptual description of narrative discourse designed to capture the information relevant for testing hypotheses on the operations and data structures required for adequate processing of embedded stories. The discourse for a story is therefore represented as a list of updates to the story, where an update is a conjunction of statements – each a predicate with arguments – that jointly describe the event. The sequence of updates to the story describes the sequence of events or facts for the story. Speci!c statements are used to indicate the start of an embedded story and the fact of its telling with the frame story, with the sequence of statements for the embedded story appearing bracketed between them in the sequence for the discourse of the frame story. The algorithm allowed the correct separation between narrative levels and the identi!cation of the spans of events in the discourse corresponding to embedded stories.
An extension of this algorithm [
• start with empty story interpretation for frame story, empty stack for initial narrative level, and empty table of embedded stories • on start of an embedded story (start_story <story-name>): – push to stack interpretation of frame story so far – create new empty story interpretation for embedded story • process updates for embedded story onto story interpretation for embedded story • on end of embedded story (tell_story statement <narrator> <narratee> <story-name>): – store accumulated interpretation for embedded story in table for embedded sub-stories
indexed by name of sub-story (<story-name>) – pop from stack interpretation for frame story acting as context, establish it as context
for rest of frame story – add special tell_story <narrator> <narratee> <story-name> statement to interpretation of frame story to encode how telling of embedded story fits into frame story
The algorithm for the interpretation of embedded stories described in can be summarised in Table 1. This results in a story interpretation for the frame story – with telling predicates to indicate where embedded stories are told – and a table of (story interpretations of) embedded stories indexed by name. The stack should be empty at the end. Each narrative level is represented by a di#erent story interpretation: the frame story in the main interpretation, those for embedded stories in the representations stored in the table. The recursive nature of the embedded stories is captured by the presence in the corresponding narrative level of telling predicates that refer in each case to the index of the corresponding embedded story in the table.
The representation to this point partitions the input narrative discourse into its constituent substories, but it does not capture the relations that may connect together the di#erent sub-stories. Some of the sub-stories may simply not be connected at all to the frame story (unrelated stories). Some of the sub-stories may be connected to the frame story by referring to events that have happened in the same storyworld as the frame story but which had not been mentioned before (preceding stories). Some of the sub-stories may be connected to the frame story by referring to the events in the storyworld for the frame story that have already been mentioned in the preceding discourse (anaphoric stories), or by presenting alternative versions of some events in the storyworld that have already been told (con!icting stories). When sub-stories refer to events in the same storyworld, an important relation between them that needs to be established is the relative chronology. Anaphoric stories often appear in a story when events already told in the frame story are retold by a witness to someone who was not present. Preceding stories are the main tool used by authors to introduce "ashbacks.
The algorithm proposed for this in [9] operates by constructing a branching partially ordered graph that compiles the events and sub-stories in the discourse according to estimates of those relations determined by basic heuristics. The procedure proposed for building this branching partially ordered graph is shown in Table 2. The procedure for establishing whether an embedded story partially matches a given span of discourse relies on alligning updates in both spans whenever the corresponding conjunctions of statements have matching statements. Statements are considered to match if they share the same predicate name and at least the value for one argument (in the same argument position of the predicate). Whenever more than one match is found, the !nal match is selected based on length of matching span and average value of percentage of shared arguments across all the matched statements. • insert events from frame story into graph a!er preceding event in frame story • on reaching embedded story: – if embedded story involves characters not present in frame story, mark as unrelated
story and store separately – otherwise: search preceding spans of frame story for matches: ∗ if match is found, mark embedded story as conflicting story and insert into graph before start of matching span and marked as conflicting view on events in the span ∗ otherwise, mark embedded story as preceding story and insert into graph before
start of frame story (it refers to a time before that point)
A related e#ort to model features involved in the interpretation of narrative discourse considers how a physical model of the world may be built from a discourse [10]. This approach relies on a two-step process: !rst the explicit information available in the discourse on storyworld locations and location changes is compiled into an intermediate data structure, then a set of heuristics is applied to expand this data structure into a fuller representation of the locations in the storyworld. The heuristics in question rely on basic assumptions about continued existence of locations in the storyworld if they are mentioned at any point in the discourse and continued presence of characters at a given location unless they have been explicitly described as moving away.
Table 3 shows an example of the process of interpretation of the information on locations and character presence as extracted by the algorithm. An example of story is given at the top, showing both a text summary and the discourse representation employed here. This story corresponds to an excerpt from tale 155, as analysed by Propp in the appendix of this book [11]. The two steps of world model construction are shown below. The gaps in continuity that were apparent in the dynamic model have been !lled in, and the representation now re"ects that movement of characters across the locations in the world: brother to dragon’s lair, princess away when she is released.
The stories we are interested in involve situations in which one of the characters tells stories about events he has witnessed, but in which he was not involved. Although the narrator may mention explicitly in his story what he was doing at the time, this does not necessarily happen. When there are no such mentions, embedded stories of this type cannot be associated to the frame story by relying on the simple heuristic of identifying shared characters across frame and embedded story. Table 4 shows examples of such a story, presented as text in column (a) and as the full transcription of the conceptual representation of the discourse used as input in column (b).1
When this story is processed with the heuristics applied in [9], the two embedded stories are identi!ed as preluding stories (because they take place at locations mentioned in the frame story, they are considered related) but the relative chronology of each one of them with respect to the frame story and the relative chronology between them cannot be identi!ed. An example of this imperfect 1Although I would have preferred to use examples of real-world narrative, the example used here has been engineered to include the maximum set of problematic features in the minimum space that would allow for the resulting output structures to !t in the available space. As mentioned later in Sections 4 and 5, further work should consider the application of the procedures described here to real-world narratives. reconstruction of the fabula is shown in column (a) of Table 5. This presents a schematic representation of the fabula reconstructed for John’s story by the original algorithm.
The solution proposed in this paper is to consider the models of the storyworld constructed for each of the stories (one for the frame story and one for each of the embedded stories) to inform the process of determining the relative placement in time of the embedded stories with respect to the frame story.
To achieve this, we rely on the assumption that when the protagonist of a story such as this one tells other characters about events that have happened at locations that he is visited during the story, he is most probably describing events that he was witnessed while he was present at those locations.
With respect to the algorithm described in Section 2.2, the following modi!cations are required to achieve this: • if an embedded story is found to be related to the frame story (not an unrelated story), but no match is found with the frame story (not a con"icting story) • the world model for the embedded story is checked for the set of locations at which it takes place • the narrator of the embedded story is identi!ed • the world model for the frame story is checked for time points at which the narrator of the embedded story was present at the locations in which the embedded story takes place • the embedded story is assumed to have taken place between the start and the end of the period during which the narrator was at those locations • the embedded story is placed in the graph for the fabula at the estimated time point An example for the graph obtained for the fabula by applying this procedure is shown in in column (b) of Table 5. This shows a schematic representation of the fabula reconstructed for John’s story by the
John woke up at home and had breakfast. Then he went to school. He had a maths class in the maths classroom. He had lunch.
He had an English class in the English classroom. He went to the soccer field to have soccer practice. Then he returned home. When he got home, he told his mother how Peter had released his pet rat in Maths class, the pet rat scared the teacher and Peter got punished. Then he told his mother how Mike had an accident during soccer practice, and how the coach helped Mike, who had to go to hospital. extended version of the algorithm, enhanced with heuristics for identifying parallelisms between embedded stories and spans of the frame story based on location information inferred from the storyworld models for frame and embedded stories.
The contrasting representations – (a) and (b) – presented in Table 5 for the input story shown in Table 4
illustrate how the proposed extension to the narrative interpretation algorithm allows for more accurate
representation of the interpretation that intuitively occurs to human readers. This is not surprising
as the proposed extension enriches the previous version of the algorithm with a process of explicit
construction of a model of the storyworld, known to be a fundamental element in human interpretation
of narrative [
An important insight arising from the work presented is that such a model of the physical aspects of the storyworld has an importance in the interpretation process beyond the simple acquisition of information about the world: without such a physical model of the storyworld inferences on relative placement of embedded stories may be incorrect or underdetermined.
The algorithm in its present form has been implemented as a Java program that takes as input text !les with the encoded version of discourse and produces text !les that represent the structures shown in the tables in the paper. Some e#ort has gone into transcribing these structures as readable LaTeX tables but their topology and information content has been respected.
The algorithm as described should be considered as a baseline approximation, intended mainly to establish the relative importance of addressing the problem in models of narrative interpretation. The current version is susceptible to incorrect assumptions as to which moments of a character’s past correspond to the point in which a given embedded story is witnessed. More re!ned heuristics to inform such decisions may be considered as further work.
An additional issue arises from the fact that the process of constructing the physical model of the storyworld happens in two stages, one of recording explicitly provided information on location and one of !lling in missing information about the world based on heuristic-driven inference. Because the Main plot line: at_location home wakes_up john has john breakfast sets_out john to_location school at_location maths_class has john maths_class has john lunch at_location english_classroom has john english_class sets_out john to_location soccer_field has john soccer_practice sets_out john to_location home tell_story john mum peter_incident tell_story john mum mike_incident (a) interpretation by original algorithm
Main plot line (continued I): has john lunch at_location english_classroom has john english_class sets_out john to_location soccer_field
Main story + side story John’s soccer practice Mike’s incident has john soccer_practice start_story mike_incident at_location soccer_field had_accident mike decide_to_help coach mike sets_out mike to_location hospital Main plot line (continued II):
sets_out john to_location home tell_story john mum peter_incident tell_story john mum mike_incident (b) interpretation by revised algorithm second step must always be defeasible – in the sense that if information provided later in the discourse contradicts any of the inferences made, those inferences need to be revised – this implies that if the placement of any embedded story has been based on such defeasible inference, it will itself be defeasible. The procedure described here may need to be reconsidered in these terms in further work.
The introduction of this type of defeasibility into the computational process opens up possibilities for the author wishing to exploit it for particular e#ects. An author may know the type of inferences that a reader may be making in processing each possible instantiation of the discourse, in terms of default assumptions as to which character is where, and therefore also about which character is aware of which events in the storyworld. By exploiting this type of information, such an author may craft particular discourses that, without being openly misleading, may result in the reader making assumptions that will obfuscate the parts of the plot that the author wants kept in the dark.
The work described in this paper is limited to a conceptual description of discourse restricted to a small subset of the features that may be expressed in natural language. In particular, it focuses on representation of events and location information of where they take place. More elaborate solutions to the problem of relative temporal placement of the events in embedded stories with respect to the frame story will be required once additional contextual cues – such as temporal expressions or discourse markers – are considered.
In more pragmatic terms, future work should consider identi!cation of appropriate machine learning solutions to replace the hand-coded heuristics currently powering the implementation. However, in order to train such machine learning solutions a corpus of examples annotated with all the relevant data would be required. If the compilation of such a corpus were to be carried out entirely by hand, it would involve a very signi!cant e#ort of knowledge engineering in itself. To simplify this process, we proposed a three stage procedure for bootstrapping such a corpus. An initial stage would rely on state of the art technologies for natural language processing to construct a corpus of examples of real world narratives transcribed into the conceptual representation used in this paper. We assume that solutions based on Large Language Model would not !nd the task onerous. The existing implementation of the algorithm would then be used to construct examples of outputs that capture the complex structure of the narratives in terms of embedded stories and relative temporal placement of the corresponding fabulae. These outputs may then be revised by human experts to ensure correctness. This step would allow repair of any errors resulting from the use of limited heuristics. This would result in two parallel corpora of the initial texts paired with the corresponding representation for their underlying narrative structure. Based on such the corresponding parallel corpus, machine learning solutions for the task may be trained.
The present paper argues the importance of including a construction of the physical model of the storyworld in any process of narrative interpretation, and how such a model can help inform the correct temporal placement of embedded stories with respect to the fabula for the frame story in cases where the narrator of the embedded story is merely a witness of the events in the story. The algorithm proposed to address this task is a simple baseline intended to underline the importance of the task within the broader context of models of the interpretation of narrative. Additional research is needed in terms of both more re!ned heuristics for the task and further experiments with a wider range of examples of input stories.
The model as described has implications for automated storytelling, AI-based narrative understanding and discourse analysis. Automated story tellers may need to consider the use of embedded stories as part of the toolkit available to describe their storyworlds. E#orts at narrative understanding need to ensure that their solutions can handle embedded stories. Approaches to discourse analysis must include a description of discourses involving more than one narrative level.
In terms of further work, there is a clear need for additional empirical evaluation and computational experiments over a broad range of real-world narrative datasets. The functionalities outlined in this paper should be tested over di#erent narrative styles or storytelling conventions. The pragmatic considerations presented at the end of Section 4 may provide a possible way forward in this sense. Once the approach is taken beyond simple explanatory examples as the ones presented in this paper, issues will arise with more complicated narratives, for instance, having situations in which multiple embedded stories con"ict or overlap in terms of location and chronology. The extensions based on machine learning contemplated in Section 4 may fare much better than hand-crafted heuristics in dealing with such complex cases.
This paper has been partially supported by project CANTOR: Automated Composition of Personal Narratives as an aid for Occupational Therapy based on Reminescence, Grant. No. PID2019-108927RBI00 (Spanish Ministry of Science and Innovation), project DARK NITE: Dialogue Agents Relying on Knowledge-Neural hybrids for Interactive Training Environments, Grant No. PID2023-146308OB-I00 (Spanish Ministry of Science and Innovation) and project ADARVE (Análisis de Datos de Realidad Virtual para Emergencias Radiológicas) funded by the Spanish Consejo de Seguridad Nuclear (CSN). [5] D. Herman, M. Jahn, M. Ryan, Routledge Encyclopedia of Narrative Theory, Taylor & Francis, 2010. [6] T. Todorov, Poétique de la prose, Média Di#usion, 2014. [7] D. Tenev, Point of view and the modalities of narrative, Prace Filologiczne. Literaturoznawstwo [PFLIT] 1 (2018) 27–42. [8] M.-L. Ryan, The modal structure of narrative universes, Poetics Today 6 (1985) 717–755. [9] P. Gervás, A model of interpretation of embedded stories, in: Text2Story: 4th International Workshop on Narrative Extraction from Texts, CEUR Workshop Proceedings, CEUR Workshop Proceedings, Lucca, Tuscany, 2021. [10] P. Gervás, Basics of narrative interpretation: Physical model and character-speci!c views of the storyworld, in: 8th International Workshop on Computational Models of Narrative (CMN’25), Geneva, Switzerland, 2025. [11] V. I. Propp, Morphology of the Folktale, volume 9, University of Texas Press, 1968.