Abstract 
The paper proposes an architecture for advanced 
NLG systems that handle narratives. Special 
attention is paid to document planning. Domain 
modelling and meta-knowledge modelling for a 
narratological structurer are exemplified. 
1 Introduction 
Natural Language Generation (NLG) systems usually focus 
on descriptive text types. Narrative structures are under-
represented in NLG. However, they are encountered in 
many naturally occurring texts, and not only in fiction. 
Often, texts are composed of segments with different 
prevailing functions: some are descriptive, some argumenta-
tive, others narrative. Therefore, knowledge about narratives 
can be useful also for NLG systems that deal with “classi-
cal” NLG tasks. For example, a system that is expected to 
describe and compare objects in a museum could include 
narrative passages about their use, history or discovery. 
The discipline concerned with the structure of narrative 
[Barthes, 1966] and narrative discourse [Genette, 1980] is 
called narratology. Narratologists distinguish between two 
main representational domains of narratives: 1. the story 
(histoire) as the totality of the narrated events, abstracted 
from their disposition in the text; and 2. the discourse 
(récit) that narrates them [Genette, 1988:13]. 
Narratology has identified a number of descriptive 
parameters. As can be seen from the explanations in Table 1 
(see Appendix), many parameters describe phenomena that 
relate discourse to story in a specific way. 
This paper envisages an NLG system with improved 
abilities to deal with narrative discourse. Section 2 summa-
rises the relation between Story Generation and NLG. In 
Section 3, an architecture for a narratologically enhanced 
NLG system is proposed. Section 4 outlines one of the de-
scriptive parameters, illustrating the kind of knowledge that 
can be acquired from narratology. Section 5 exemplifies the 
domain modelling, a prerequisite for modelling narratologi-
cal knowledge. Section 6 formalises the selected narra-
tological parameter, discusses its mapping onto discourse 
graphs used in NLG, and exemplifies its use in a system. 
Section 7 concludes the paper and mentions future work. 
2 Story Generation and NLG 
This section gives a brief overview of the relationship 
between Story Generation (SG) and NLG. Subsection 2.1 
describes the architectural gap between SG and NLG. 
Subsection 2.2 discusses the contribution of an existing 
system to fill that gap. An extended account of work in SG 
cannot be provided here. Pérez y Pérez and Sharples [2004] 
compare and evaluate three recent SG systems. For a 
discussion of earlier systems, see [Ryan, 1991:233–248]. 
2.1 The Story Generation-NLG Gap 
Story Generators aim to produce interesting, understand-
able, artistic, and creative stories [Turner, 1994:15]. In 
doing so, they focus on the story (histoire) representational 
domain. For example, the generator MINSTREL [Turner, 
1994] achieves not only thematic and consistency goals, but 
also “dramatic” and “presentational” goals solely by adding 
or transforming story events and background information. 
The story content is directly mapped onto the output text. 
SG faces the difficult problem of finding an interesting 
and logically coherent event sequence (story). Less energy 
is spent on other tasks, especially at discourse and surface 
level. In fact, most implemented SG systems use templates 
of different varieties for NLG. Document structuring and 
microplanning as NLG processing stages [Reiter, 1994; 
Reiter and Dale, 2000:60] are usually skipped in SG, so that 
SG and NLG architectures can be confronted as in Figure 1. 
 
 
 
 
 
 
 
 
 
 
 
 
Figure 1: Story Generation and Natural Language Generation 
Narratological Knowledge for Natural Language Generation 
Birte Lönneker 
Narratology Research Group Hamburg 
Institut für Germanistik II, University of Hamburg 
Von-Melle-Park 6, D-20146 Hamburg 
birte.loenneker@uni-hamburg.de 
Content det.
Doc. structuring
Document planner
Microplanner
Surface realizer
Communicative goal
Document plan
(Abstract) text specification
Surface text
Determine content (story), directly
associated with structure (discourse)
Template filling
of given genre
(Abstract) content specification
Surface narrative
Implicit goal: Write coherent narrative
Traditional SG
General NLG
2.2 STORYBOOK 
The existence of a gap between SG and NLG has also been 
discussed by Callaway and Lester [2002]. To remedy the 
situation, they implemented STORYBOOK, an “architecture 
for narrative prose generation” [2002:228]. The input to 
STORYBOOK must be a “narrative stream that reflects the 
orderly progression of events, descriptions and states” 
[Callaway and Lester, 2002:231], produced by a hypotheti-
cal Story Generator. STORYBOOK expects this input stream 
to contain so-called narrative primitives that specify scene 
changes and other aspects such as: 
• narrative person and focalization (“person, omnis-
cience” [Callaway and Lester, 2002:221]); 
• details about dialogue realization (e.g. whether an 
inquit phrase such as “he said” should precede, follow 
or interrupt direct speech, or whether it should be 
omitted altogether [Callaway and Lester, 2002:225]). 
Person and focalization are classical narratological pa-
rameters [Genette, 1980]. It is not clear from Callaway and 
Lester [2002] how many different values STORYBOOK actu-
ally implements for these parameters; their example tale is 
invariably rendered by a “third person disembodied” narra-
tor [2002:230]. — The question of whether or not a charac-
ter utterance is to be provided with an inquit phrase is 
sometimes subsumed under the narrative distance parameter 
(cf. Table 1): Direct speech framed by inquit phrases is defi-
nitely considered as imitating, but the highest level of imita-
tion is achieved by the reproduction of unframed direct 
speech [Fludernik, 2005; Genette, 1988:56]. 
In STORYBOOK, many decisions concerning discourse 
still lie with the Story Generator that is supposed to produce 
a stream with narrative primitives. STORYBOOK itself nei-
ther decides about narrative person, focalization, or distance 
for any given passage, nor is it concerned with any other 
narratological parameters listed in Table 1. Rather, 
STORYBOOK focuses on microplanning and surface phe-
nomena with special attention to features particular to nar-
rative. The architectural contribution of STORYBOOK can 
thus be seen as reducing the gap between SG and NLG to 
approximately the microplanner level, where STORYBOOK, 
as a specialized NLG system, takes over from a hypothetical 
Story Generator (see Figure 2). 
 
 
 
 
 
 
 
 
 
 
 
Figure 2: Story Generation and STORYBOOK 
3 Narratology for NLG 
Narratological aspects influence on all architectural modules 
[Reiter, 1994] or representation levels [Cahill et al., 2000] 
of NLG. The most important decisions of a narratologically 
enhanced system concern the document planner with its 
content determination and document structuring sub-
tasks (Subsections 3.1–3.2). Microplanning and surface 
realization (Subsection 3.3) are supposed to resemble classi-
cal NLG processes. As further explained below, the archi-
tecture of the envisaged system corresponds to Figure 3. 
 
 
 
 
 
 
 
 
Figure 3: Narratologically enhanced Story Generation-NLG system 
3.1 Content Determination 
In a narratologically enriched NLG system, a “narrative 
content determination module” might be implemented as a 
Story Generator. The Story Generator decides 
• what it considers to be a minimal story [Prince, 1973], 
• what it considers a good story (cf. Subsection 2.1), 
• how to select the content that represents the story: 
events, participants, and their relations. 
However, if narratological discourse parameters are to be 
applied, certain story events or participants are required. For 
example, the narrative levels parameter presupposes the 
existence of at least one NARRATING action among the story 
events of the (first) narrative (see Sections 4–6). 
This means that an entirely data-driven, pipelined process 
where content determination precedes document structuring 
[Reiter and Dale, 2000:111] will not use the full potential of 
a narratologically enhanced document planner. Rather, the 
content determination module (the Story Generator) should 
be able to satisfy constraints issued by the document 
structurer. These constraints might well take the form of 
modification requests after a first output attempt of the Story 
Generator. Therefore, a hypothesis-driven architecture in 
which the system switches back and forth between content 
determination and document structuring seems well suited 
for the task at hand (see the double arrow in Figure 3). 
3.2 Document Structuring 
In the envisaged architecture, the document structurer 
receives the narrative content representation from the Story 
Generator. In its simplest form, this input is a sequence of 
events together with their participants. The structurer shares 
a domain model with the Story Generator, where the 
semantics of events and participants are represented. 
Determine content
(story) and
part of structure
(esp. causality)
Story Generator
of given genre
Goal: Write coherent narrative
NLG
Narratological structurer
Microplanner
Surface realizer
Document plan
(Abstract) text specification
Surface text
Vary (order, person,
point of view,...);
representation
create discourse
sentation
Abstract
content
repre−
Microplanner
Surface realizer
(Abstract) text specification
Narrative stream
Surface text
person, focalization, speech reproduction
Determine content (story), decide about
of genre: fairy tale
StoryBook
Hypothetical SG
Implicit goal: Write coherent narrative
By default, the incoming event sequence is considered as 
corresponding to a discourse related by the NarrativeSe-
quence discourse relation
1
 (for discourse relations, cf. Sub-
section 6.3). The sequence might have additional internal 
structure, because most Story Generators represent conse-
quences of events and/or episodes, and therefore some kind 
of causality [e.g. Gervás et al., 2004; Pérez y Pérez and 
Sharples, 2004; Turner, 1994]. 
Combining this information with information from the 
ontology and with narratological knowledge, the narra-
tological document structurer creates variation in discourse. 
Most narrative parameters presented in Table 1 influence on 
this representational domain. Amongst others, the Narra-
tological Structurer can perform the following tasks, some 
of them in interaction with the Story Generator: 
• modify the order of events or episodes, tag the shifted 
elements as flashbacks or flashforwards, and accord-
ingly modify the discourse relations (a Cause relation 
might become an Explanation relation because of the 
inverted event order); 
• create ellipsis by suppressing events or episodes that 
the reader might as well infer; 
• add or remove narrative levels (Sections 4–6); 
• decide about point of view and focalization for the 
entire narrative or, rather, for certain episodes, and ac-
cordingly present or suppress events and relations. 
The produced document plan contains discourse relations 
and narratological information for nodes and leaves, which 
can be further used in microplanning and surface realization. 
3.3 Microplanning and Surface Realization 
The microplanner performs lexicalisation and aggregation 
and generates referring expressions. Besides the usual 
constraints, it respects narratological contraints. For exam-
ple, certain values of focalization and point of view can 
cause “unusual” referring expressions: personal pronouns 
might be used for initial reference [Harweg, 1968: 163–
166], or indefinite referring expressions for subsequent 
reference [Ushie, 1986; Wiebe and Rapaport, 1988]. The 
choice of vocabulary, tense, and syntactic complexity can as 
well depend on narratological factors. 
The surface realizer turns the microplanner output into 
natural language text. See [Callaway and Lester, 2002:224] 
for narrative-related issues in surface realization. 
4 The Narrative Levels Parameter 
The narrative levels parameter will be used to illustrate the 
kind of knowledge that can be acquired from narratology for 
integration into a narratologically enhanced NLG system. 
Narratives can consist of different levels: There are “tales 
within tales within tales” [Barth, 1984]. Genette [1988:85] 
illustrates this phenomenon by a figure similar to Figure 4, 
                                                 
1
 Similar to the relations Narration [Asher and Lascarides, 
2003:162–165] or Occasion [Hobbs, 1990:86–89]. 
 
 
 
 
 
Figure 4: Narrative levels 
in which person i produces an utterance a that contains the 
production of another utterance b by person j. Talking about 
narratives, the “persons” in Genette’s drawing correspond to 
narrative instances (narrators). Narrative instance i might 
not be identified by any referring expression throughout the 
utterances produced by i (and j): i can be “absent from the 
story” or heterodiegetic [Genette, 1980:244–245]. Narrative 
instance j, on the other hand, necessarily has to be identical 
to a character in story a. Finally, j might be absent (or not) 
from the “inner” narration b that he himself produces (see 
Table 1 on Person). 
The generation of a narration within a narration presup-
poses the existence of a NARRATING action in the “outer” 
narration: in Figure 4, a NARRATING action (where j tells b) 
occurs within outer narration a.
2
 Therefore, before narrative 
levels will be modelled in detail (Section 6), Section 5 will 
be concerned with a definition of NARRATING. 
Genette’s [1980:244] considerations suggest that narra-
tive levels is one of the simpler narratological parameters. 
For example, he claims that any narrative can be converted 
into an inner narration by adding “presentational” outer nar-
rations [Genette, 1988:95]. It should thus be possible to 
enable a generation system to create narrativel levels. 
5 NARRATING and Related Frames 
This section proposes representation structures for 
NARRATING (Subsection 5.1) and related frames, especially 
the NARRATED frame (Subsection 5.2). A frame is consid-
ered to represent structured knowledge about a given con-
cept. In terms of NLG tasks, this modelling corresponds to 
setting up part of a domain model for content determination 
[Reiter and Dale, 2000:86–89]. 
The modelling is supported by concordances from the 
German DWDS corpus.
3
 Contexts of verb forms of erzählen 
(‘tell; narrate’) were investigated, because they “evoke” the 
target frame NARRATING. As DWDS texts are of several 
types, the results reflect general (not only literary) language 
use. The analysed corpus material stems from texts pub-
lished between 1990 and 1999. The modelling is also partly 
inspired by the SUMO
4
 upper ontology [Niles and Pease, 
2001]. SUMO does not yet include the NARRATING concept. 
5.1 The NARRATING Frame 
The NARRATING frame is an indirect subframe of ACTION. 
Narratological theories [Chatman, 1978:44–45] agree with 
                                                 
2
 “Inside” vs. “outside” narrations [Barth, 1984:233] are also 
called “framed” vs. “framing” or “embedded” vs. “embedding”. 
3
 http://www.dwds.de [Query date: 24 January, 2005]. 
4
 http://ontology.teknowledge.com/ [10 March, 2005]. 
i
j
a
b
current ontologies like SUMO on the fact that an action is 
performed intentionally by a cognitive agent. The NARRAT-
ING frame thus inherits an Agent slot which I relabel 
NarrativeInstance for better orientation (cf. Table 5). 
The NARRATING frame is related to other frames sub-
sumed under the superframe COMMUNICATION_ACTION. 
That superframe represents an action in which typically at 
least two participants exchange messages. In the present 
investigation, NARRATING actions are assumed to express 
this content linguistically; therefore, NARRATING is here also 
a subframe of LINGUISTIC_COMMUNICATION (cf. Figure 5). 
In the NARRATING frame, the content is represented as 
complex filler of the Narrated slot; a more detailed 
modelling of the NARRATED frame follows in Subsection 
5.2. NARRATING is a special subframe of COMMUNICA-
TION_ACTION insofar as it usually restricts the direction of 
the information flow between the participants: One of them 
is the (main) producer, the other one the (main) recipient of 
NARRATED. The respective participant slots are the above 
mentioned NarrativeInstance, and the Addressee. Typical-
ly, the Addressee or “narratee” [Genette, 1980] is a cogni-
tive agent different from NarrativeInstance. Nevertheless, 
the current model also allows the same individual to act 
simultaneously as both NarrativeInstance and Addressee: 
cognitive agents may “tell a story to themselves”. 
Comments and questions produced by the Addressee, 
concerning the Narrated, deserve some attention. The pre-
sent conceptualization considers that even if comments and 
questions interrupt of the flow of Narrated, they do not 
necessarily mark the end of a NARRATING action (see also 
Subsection 5.2). 
Some other slots of NARRATING are inherited from the 
ACTION or EVENT frames. In the DWDS corpus material (cf. 
Table 2, Appendix), examples of fillers of Motivation, 
Consequence/Aim, ParallelEvent, Time, Place, and Man-
ner are encountered. Typically, one of the Consequences of 
NARRATING is a change of the Addressee’s mental state: the 
Narrated is added to her or his knowledge. This conse-
quence seems to belong to the lexico-semantic knowledge 
about NARRATING or, probably, COMMUNICATION_ACTIONS 
in general, and is rarely mentioned explicitly in texts. 
Table 3 gives an overview of the NARRATING frame. It 
also maps slot names on binary relation labels and formu-
lates restrictions for the fillers. However, as a representation 
of a conceptual entity, it does not specify linguistic restric-
tions such as “addressee-oriented, non-message-oriented” 
which are considered to hold language-specifically for indi-
vidual verbs [Bateman et al., 1994]. For the use in a specific 
NLG system, the frame and its slots thus need to be mapped 
onto the lexicon used by that system during lexicalisation. 
5.2 The NARRATED Frame 
A text can refer to different aspects of what is narrated in a 
NARRATING action. Hence, instead of introducing further 
slots within the NARRATING frame, the filler of Narrated 
will itself be represented as a frame (cf. Table 4, Appendix). 
One way of referring to the NARRATED_LINGUISTIC_MES-
SAGE, abbreviated as NARRATED, is to mention the Text- 
 
 
 
 
 
 
 
 
 
 
 
Figure 5: Frame hierarchy containing NARRATING 
NARRATING 
Slot Relation Filler Restriction 
NarrativeInstance isCausedBy CognitiveAgent (> 0) 
Narrated produces Narrated (1) 
Addressee isAddressedTo CognitiveAgent (>= 0)
Motivation isMotivatedBy AbstractEntity (>= 0) 
Consequence/Aim causes Event (> 0) 
ParallelEvent overlapsWith Event (>= 0) 
Time happensAtTime TimePosition (> 0) 
Place locatedAt PhysicalEntity (> 0) 
Manner is Attribute (>= 0) 
Table 3: The NARRATING frame 
Type of the narrated material. Typically, this is a literary 
genre, or one of its parts. Further, the NARRATED uses a 
MessageSupport, which might be written (e.g. book, let-
ter), spoken (e.g. speech, song) or thought (e.g. dream). 
Thought seems to lend itself to somewhat problematic sub-
divisions; see e.g. [Genette, 1980:231] on “retrospections”. 
The abstract content of NARRATED is represented linguis-
tically through a LinguisticSurfaceForm. As can be seen 
from the DWDS corpus, the LinguisticSurfaceForm oc-
curs as a new sentence or clause, and can be more or less 
“direct”. Research on the reproduction of speech is an area 
of both narratology and linguistics. Usually, the forms direct 
speech (D in Table 4), free indirect speech and indirect 
speech (I in Table 4) are distinguished in increasing order of 
narrative distance (see Table 1) [Fludernik, 2005]. 
Comments and questions produced by the Addressee of a 
NARRATING action, concerning the Narrated, should also be 
part of the LinguisticSurfaceForm: Metaphorically speak-
ing, they might be seen as discursive catalysts or parasites of 
that surface form. 
Finally, the Topic slot of the NARRATED frame is typical-
ly filled by sentence constituents like objects or preposi-
tional phrases. Since these constituents can themselves be 
complex, the topic might be quite detailed, taking the form 
of a summary. Still, it does not provide access to the 
“original” linguistic surface form. With respect to narrative, 
this category is also referred to as narrative report of 
discourse or content summary [Fludernik, 2005]. 
Figure 6 illustrates the knowledge modelled by 
NARRATED and related frames. 
Happening
Event (Process)
Action
... ...Experiencing...
CommunicationAction ...
...LinguisticCommunication
DirectingCommittingNarratingReporting ...
produces(LinguisticMessage)
produces(NarratedLinguisticMessage)
 
 
 
 
 
 
 
 
 
 
 
Figure 6: The NARRATED_LINGUISTIC_MESSAGE frame 
At a meta-level, the NARRATED frame can be mapped 
onto representation structures of the envisaged system that 
combines SG and NLG (cf. Figure 3 above). The TextType 
is implicitly given in the goal of the Story Generator; also 
the Topic or a topic restriction is often given at this stage. 
The Support is, depending on the stage of processing, some 
digital entity, written (printed), or possibly speech. The Ab-
stractContent is the “abstract” output of a Story Generator 
(the product of content determination). The LinguisticSur-
faceForm corresponds to the final NLG output based on 
this abstract content. Using knowledge and reasoning proce-
dures, a system might additionally or alternatively summa-
rise AbstractContent into a phrase that represents Topic. 
6 Representing Narrative Levels 
This section is concerned with describing a formalism that 
should enable a discourse planner to handle narrative levels 
(see Section 4). Via the formalism, the system can be told 
• how many levels a generated narrative should have, 
• which narrative instance should be the “direct” 
mediator to the reader, and 
• how outer and inner narrations, or parts thereof, 
should follow each other in discourse. 
The basic ingredients for the formalism (Subsection 6.1) 
are applied to a description of example constellations (Sub-
section 6.2) and mapped onto discourse relations (Subsec-
tion 6.3). Subsection 6.4 briefly exemplifies the cooperation 
between Story Generator and Narratological Structurer. 
6.1 Basic Ingredients: Classes and Relations 
Three classes and two relations need to be modelled to 
represent possible structures of global narratives with 
respect to their levels. 
NARRATION 
An instance of NARRATED is a NARRATION if its Linguistic-
SurfaceForm is of the subtype DirectLinguisticSurface-
Form (and is, as such, explicitly present in the output text). 
This narratological constraint correctly excludes from 
opening a new narrative level such NARRATING actions 
which are merely mentioned, or whose NARRATED is 
reproduced indirectly. 
NARRATIVE_INSTANCE 
A NARRATIVE_INSTANCE is a cognitive agent that occupies 
the agent role (NarrativeInstance slot) of a NARRATING 
action. Neither the instance itself nor the action needs to be 
referred to directly in the surface form of any NARRATION. 
For example, if i isA NARRATIVE_INSTANCE, the variable 
i is not necessarily “resolved” against a character name. 
toldBy(NARRATION, NARRATIVE_INSTANCE) 
The relation toldBy(a,i) holds if there is a NARRATING action 
x, a NARRATION a and a NARRATIVE_INSTANCE i such that 
both isCausedBy(x,i) and produces(x,a) are true (Table 3). 
toldIn(NARRATION, NARRATION) 
ToldIn(b,a) or “b is told in a” holds if there is a NARRATION 
b, a NARRATION a, and a NARRATING action x such that: 
• a is not b; 
• a is an event sequence such that one of its events is x; 
• produces(x,b). 
DISCOURSE_SEQUENCE 
A DISCOURSE_SEQUENCE is an ordered list of variables 
standing for different NARRATIONS, as they sequentially oc-
cur within the discourse of a global narrative. In a DIS-
COURSE_SEQUENCE, the same variable may appear more 
than once: for example, an outer narration can frame an 
inner narration at the start and the end of a discourse. 
Variables for separated parts of the same narration receive 
numbered subscripts (see a
1
 and a
2
 in Figure 7 below). 
6.2 Example Constellations 
Type 0: Zero-Instantiation 
Zero instantiation of the narrative levels parameter occurs if 
there is no change in level. There is only one narration, told 
by a narrative instance. An example of this type is the fable 
The Fox and the Crow by Aesop, because neither the fox 
nor the crow produce any narrations. The instantiation of the 
classes and relations corresponds to the scenario in Table 5. 
I am not aware of any Story Generator (cf. Section 2) that 
exceeds the zero instantiation. This would mean that con-
temporary Story Generators ignore narrative levels. 
Type 1: Outer and one Inner Narration 
A typical pattern of Type 1 is as follows: An outer narration 
contains a NARRATING action whose agent is a narrative 
instance different from that of the outer narration. The NAR-
RATED of this action is the inner narration. After the NAR-
RATING action, the rest of the outer narration is told (Table 6). 
An example is Heart of Darkness (1902) by Joseph 
Conrad. A homodiegetic narrative instance i (= “I”) tells 
outer narration a, while j (= “Marlow”) tells inner b: 
 
 
 
Figure 7: Narrative levels and discourse sequence 
uses MessageSupport
isOfTextType TextType
hasForm LinguisticSurfaceForm
isAbout
Topic
Relation Filler Restriction
LinguisticSurfaceForm
MsgSupport
IndirectLSFFreeIndirectLSFDirectLSF
Written Thought
Spoken
NarratedLinguisticMessage
represents(AbstractContent)
i
j
a
b
ba
1
a
2
Classes and Relations Instantiation 
Narration a 
NarrativeInstance i 
Attribution of NarrativeInstance toldBy(a,i) 
Attribution of Narration -- 
Sequence discourse_sequence=[a] 
Table 5: Zero-Instantiation of Narrative Levels 
Classes and Relations Instantiation 
Narration a,b 
NarrativeInstance i,j 
Attribution of NarrativeInstance toldBy(a,i), toldBy(b,j) 
Attribution of Narration toldIn(b,a) 
Sequence discourse_sequence=[a
1
,b,a
2
] 
Table 6: One Inner Narration; Framing Outer Narration 
[a
1
 toldBy i] – there was nothing else to do till the end of 
the flood; [...] we knew we were fated [...] to hear about one 
of Marlow's inconclusive experiences. 
“[b toldBy j] I don't want to bother you much with what 
happened to me personally,” [comment by i] he began, 
showing in this remark the weakness of many tellers of tales 
[...]; “[b toldBy j] yet to understand the effect of it on me 
you ought to know how I got out there, [...], how I went up 
that river to the place where I first met the poor chap. [...] 
“[b toldBy j] I had then, as you remember, just returned 
to London [...].”
5
 
The “implicit inquit” phrase he began [...] is a comment 
by i concerning the Manner of NARRATING. It is not mod-
elled in the discourse_sequence (see also Subsections 5.1–
5.2 on comments). — The closing part of outer narrative a in 
Heart of Darkness is very short (cf. Figure 8, Appendix). 
Type 1 has some variants. For instance, an outer narration 
might not be resumed at the end of the discourse (Table 7, 
Appendix), as in Abbé Prévost’s Manon Lescaut (1731). In 
contrast, an outer narration (here: b) might manifest itself in 
discourse only after the NARRATED (here: a) produced in the 
NARRATING action it contains (Table 8, Appendix). Philip 
Roth’s novel Portnoy's Complaint (1969) shows this pattern. 
Yet another variant occurs when the narrative instances of 
outer and inner narration are identical (Table 9, Appendix). 
An example is Balzac’s Sarrasine (1831), where a homo-
diegetic narrator (i) tells a first narration (a) to an “un-
known” addressee (the “reader”). In the course of a, a NAR-
RATING action is performed by i himself, but with a different 
addressee, namely a character in a (cf. Figure 9, Appendix). 
Other Types: More Than Two Narrations 
More than two NARRATIONS can easily be combined 
extending the patterns presented above. 
Common “outward” extensions of levels include explicit 
introductory or closing NARRATING actions, often produced 
on written support by (fictional) editors, translators or 
“discoverers” of the inner narration. An example is the 
                                                 
5
 http://www.online-literature.com/conrad/heart_of_darkness/ 
[10 March, 2005]. 
novel Der goldne Spiegel (‘The golden mirror’, 1772) by 
Christoph Martin Wieland, for which a structure with three 
opening narrations “leading to” the innermost narration d, 
has been identified
6
 (cf. Table 10, Appendix). 
At least equally common are “inward” extensions of the 
levels. While several NARRATING actions within one and the 
same outer narration are considered to produce parallel 
inner narrations at the same level, further levels are added 
when a narration is produced within an inner narration. In 
The 1001 Nights, Scheherazade entertains King Shahryar 
with parallel inner narrations. In these narrations, the char-
acters sometimes tell one another stories, which are inner 
narrations at the next level. Table 11 (Appendix) shows a 
simplified version of The 1001 Nights, in which Sche-
herazade (j) tells only three stories (b,c,d). The third level is 
added by narration e, told by a character (k) of narration d. 
6.3 Discourse Graphs 
The basic assumption underlying discourse graphs is that 
discourse relations (e.g. [Mann and Thompson, 1988; 
Hobbs, 1990:83–111; Asher and Lascarides, 2003]) exist 
between discourse segments, where a segment is either a 
proposition or a graph consisting of already related proposi-
tions. Most theories assume that a limited set of discourse 
relations covers all semantic links that might exist between 
discourse segments. 
Based on the status of segments they relate, multinuclear 
and nucleus-satellite discourse relations can be distin-
guished. The definition of nuclearity by Mann and Thomp-
son [1988:266] is centered around the notions of compre-
hensibility (a text is less understandable if a nucleus is 
deleted), substitutability (satellites can more easily be 
replaced) and writer's purpose (nuclei are more essential). 
Similarly, Hobbs [1990:104] subdivides discourse relations 
into coordinating and subordinating. 
There are correspondences to narratological theory (e.g. 
[Barthes, 1966; Pavel, 1985; Chatman, 1978]). Chatman 
subdivides events, as constituting elements of story, into 
kernels and satellites. Kernels cannot be deleted without 
disturbing the logic of the plot. Satellites contribute first of 
all aesthetically, so that their deletion is logically possible. 
A difference lies in the size of the analysed segments. 
Linguistically inspired theories include propositions 
(clauses) as basic level, but descriptions using these theories 
often stop at the level of one or several paragraphs (an ex-
ception is [Longacre, 2003]). Narratological analyses usu-
ally concentrate on larger segments, represented by para-
phrases of their propositions [Chatman, 1978:54]. 
Discourse Graphs for Type 1 Narrative Levels 
To a certain degree, the narratological description of narra-
tive levels constrains which relations might hold between 
narrations of a “levelled” global narrative. Still, it does not 
completely disambiguate the global discourse structure. 
In fact, in a typical Type 1 narrative (Table 6), the inner 
narration b might be related in two competing ways to nar-
                                                 
6
 By Jörg Schönert in FGN Forum Narratologie 
(http://www.narrport.uni-hamburg.de/) on 21 December, 2004. 
ration a. The first analysis relates a and b based on a local, 
lexical consideration: a NARRATION is regarded as an argu-
ment of the verb phrase that evokes NARRATING. This is 
illustrated by Example (1). If NARRATED TextType filler 
(e.g. “a story”) is present in a sentence, the English verb tell 
allows an implicit LinguisticSurfaceRealization (1a). This 
– initially underspecified – argument of tell might then be 
realized in a subsequent sentence (1b), so that an Elabora-
tion relation holds between (1a) and (1b). In nucleus-satel-
lite terms, (1a) is a nucleus elaborated by (1b). 
(1) a. [a toldBy i] John told a story to Mary. b. “[b 
toldBy j] I was seventeen years old, and [...].” 
The proposed analysis is in line with Asher and 
Lascarides [2003:285], who identify an Elaboration relation 
between the propositions (2a) and (2b) of Example (2). 
(2) a. John made a promise to Mary. b. He would 
phone her. 
The possibility that the resulting discourse graph (Figure 
10) indeed holds for a global narrative increases if the 
NarrativeSequence representing narration b is relatively 
short. This also confirms that in these cases, the DirectLin-
guisticSurfaceRepresentation of the content of b is an 
aesthetical choice rather than a “plot-logical” necessity. 
The presented analysis based on local features does not 
always seem globally appropriate as well. Outer narrations 
often set the background for a (longer) inner narration at the 
start of the discourse_sequence, and draw conclusions, 
interpret or evaluate it at the end. Figure 11 shows the 
corresponding graph, in which the segment representing 
narration b is nucleus of two relations.
7
 This corresponds to 
a “schema” identified by Mann and Thompson [1988:247]. 
Danlos [2004:130] presents similar double-nuclear struc-
tures (“factorized” nuclei) at the sentence level. 
Both presented graphs are plausible. The final choice for 
a given global narrative depends not on narratological dis-
course description alone, but on story contents as well. 
6.4 Processing Examples 
This subsection briefly sketches two examples of how the 
Story Generator and the Narratological Structurer (cf. Fig-
ure 3) could cooperate during the creation of typical Type 1 
narrative levels (cf. Subsection 6.2). 
Adding an Outer Narration 
The Story Generator creates Story A. Respecting the condi-
tions for a Type 1 narrative, the Narratological Structurer 
requests an additional Story B, with some restrictions: 
• B is shorter than A; 
• B contains a NARRATING event. 
The Story Generator generates Story B. This might be a 
minimal story consisting of the events MEETING (of char-
acters c
1
 and c
2
), NARRATING (of a story by c
1
), REACTION 
(of c
2
). Upon receipt of B, the Narratological Structurer fills 
                                                 
7
 At a global level, the Background relation can be compared 
only in a very abstract way to the definitions given by Asher and 
Lascarides [2003:165–168] or Mann and Thompson [1988:273]. 
 
 
 
 
 
 
 
 
Figure 10: Discourse Graph 1 for Type 1, Framing Outer Narration 
 
 
 
 
 
 
 
 
Figure 11: Discourse Graph 2 for Type 1, Framing Outer Narration 
Story A into the Narrated slot of the NARRATING event in 
B, tags it as direct speech for the subsequent processing 
modules, and arranges all contents into a discourse tree 
corresponding to Figure 11. Story A has become an inner 
narration and a nucleus, Story B is the outer narration. 
Adding an Inner Narration 
The Story Generator creates Story A which contains a 
NARRATING event. The Narratological Structurer requests an 
additional Story B, with the following restrictions: 
• B is shorter than A; 
• the contents of B illustrate a part of A, for example an 
event present in A, or a character trait of the Narra-
tiveInstance of the NARRATING event. 
The Story Generator generates Story B. For instance, the 
illustration of A might be achieved by similarity: if Story A 
contains a CRIME event (e.g. KIDNAPPING), Story B can con-
tain a CRIME event as well (e.g. ROBBERY). The Narratologi-
cal Structurer now fills the new Story B into the Narrated 
slot of the NARRATING event in A, tags it as direct speech, 
and arranges all contents into a discourse tree, this time 
corresponding to Figure 10. Story A has become an outer 
narration, Story B is the inner narration and a satellite. 
7 Conclusions and Future Work 
The proposed system architecture combines Story Genera-
tion, narratological structuring and traditional NLG compo-
nents into an advanced NLG system for the production of 
narratives. Starting with the narrative levels parameter, do-
main modelling and meta-knowledge modelling for the 
Narratological Structurer were exemplified. Future work 
will include comments by narrators, some of which can be 
compared to “global level” discourse markers. 
Further narratological parameters will be modelled, and 
constraints will be formulated for their accumulation. A 
cognitive approach might work out their aesthetical effects. 
a
1
NarrativeSequence
a
2
NarrativeSequence
Elaboration
[31]John told:[1−30] [33][32] [34]
"Outer Narration" a
NarrativeSequence
NarrativeSequence
"Inner Narration" b
Satellite
Nucleus
[35−40]
a
1
NarrativeSequence
[1−3]
a
2
NarrativeSequence
[13−15]
Background
Interpretation/
Evaluation/...
NarrativeSequence
"Outer Narration" a
NarrativeSequence
"Inner Narration" b
SatelliteSatellite Nucleus Nucleus
[4−12]
For example, certain receptional states or strategies could be 
identified as “prototypically” activated, enabled, or blocked 
by a given combination of narratological parameter instan-
tiations. Finally, content-related studies will lead to concrete 
proposals for the interaction between a Narratological Struc-
turer and one of the existing Story Generators. 
Acknowledgments 
This research is supported by DFG (German Research 
Foundation) grant ME 1546/2-1. Thanks go to all members 
of the Narratology Research Group Hamburg, especially to 
Rolf Krause, Jan Christoph Meister and Stefanie Thiedig, as 
well as to three anonymous reviewers for their comments. 

References 
[Asher and Lascarides, 2003] Nicholas Asher and Alex Las-
carides. Logics of Conversation. Cambridge University 
Press, Cambridge and New York, 2003. 
[Barth, 1984] John Barth. The Friday Book, chapter Tales 
Within Tales Within Tales, pages 213–252. John 
Hopkins University Press, Baltimore and London, 1984. 
[Barthes, 1966] Roland Barthes. Introduction à l'analyse 
structurale du récit. Communications, 8:1–27, 1966. 
[Bateman et al., 1994] John Bateman, Bernardo Magnini 
and Fabio Rinaldi. The Generalized {Italian, German, 
English} Upper Model. In Proceedings of the ECAI-94 
Workshop on Implemented Ontologies, pages 35–45, 
Amsterdam, August 1994. 
[Cahill et al., 2000] L. Cahill, C. Doran, R. Evans, R. Kib-
ble, C. Mellish, D. Paiva, M. Reape, D. Scott, and N. 
Tipper. Enabling Resource Sharing in Language Gen-
eration: An Abstract Reference Architecture. Technical 
Paper, MITRE, 2000. 
[Callaway and Lester, 2002] Charles B. Callaway and James 
C. Lester. Narrative Prose Generation. Artificial Intelli-
gence, 139(2):213–252, August 2002. 
[Chatman, 1978] Seymour B. Chatman. Story and Dis-
course. Narrative Structure in Fiction and Film. Cornell 
University Press, Ithaca, NY, 1978. 
[Danlos, 2004] Laurence Danlos. Discourse Dependency 
Structures as constrained DAGs. In Proceedings of the 
5th SIGDIAL Workshop on Discourse and Dialogue, 
pages 127–135, Cambridge, MA, April 2004. 
[Fludernik, 2005] Monika Fludernik. Speech Representa-
tion. In David Herman, Manfred Jahn and Marie-Laure 
Ryan, editors, Routledge Encyclopedia of Narrative 
Theory, pages 558–563. Routledge, London, 2005. 
[Genette, 1980] Gérard Genette. Narrative Discourse. An Es-
say in Method. Cornell University Press, Ithaca, NY, 1980. 
[Genette, 1988] Gérard Genette. Narrative Discourse Revis-
ited. Cornell University Press, Ithaca, NY, 1988. 
[Gervás et al., 2004] Pablo Gervás, Belén Díaz-Agudo, Fe-
derico Peinado, and Raquel Hervás. Story Plot Genera-
tion based on CBR. In Applications and Innovations in 
Intelligent Systems XII (at AI-2004), pages 33–46, 
Cambridge, UK, December 2004. 
[Harweg, 1968] Roland Harweg. Pronomina und Text-
konstitution. Fink, München, 1968. 
[Hobbs, 1990] Jerry R. Hobbs. Literature and Cognition. 
CSLI, Stanford, CA, 1990. 
[Longacre, 2003] Robert E. Longacre. Holistic Textlinguis-
tics. Electronic working paper, SIL International, 2003. 
[Mann and Thompson, 1988] William C. Mann and Sandra 
A. Thompson. Rhetorical Structure Theory. Toward a 
Functional Theory of Text Organisation. Text, 8(3):243–
281, 1988. 
[Niles and Pease, 2001] Ian Niles and Adam Pease. Towards 
a Standard Upper Ontology. In Proceedings of the 2nd 
International Conference on Formal Ontology in Informa-
tion Systems, pages 2–9, Ogunquit, ME, October 2001. 
[Pavel, 1985] Thomas G. Pavel. The Poetics of Plot. The 
Case of English Renaissance Drama. Manchester Uni-
versity Press, Manchester, 1985. 
[Pérez y Pérez and Sharples, 2004] Rafael Pérez y Pérez and 
Mike Sharples. Three computer-based models of story-
telling: BRUTUS, MINSTREL and MEXICA. Knowl-
edge-Based Systems, 17:15–29, 2004. 
[Prince, 1973] Gerald Prince. A Grammar of Stories. An 
Introduction. Mouton, The Hague and Paris, 1973. 
[Reiter, 1994] Ehud Reiter. Has a Consensus NL Generation 
Architecture Appeared, and is it Psycholinguistically Plau-
sible? In Proceedings of the 7th International Workshop 
on Natural Language Generation, pages 163–170, Ken-
nebunkport, ME, June 1994. 
[Reiter and Dale, 2000] Ehud Reiter and Robert Dale. Build-
ing Natural Language Generation Systems. Cambridge 
University Press, Cambridge, 2000.  
[Rumelhart, 1975] David Rumelhart. Notes on a Schema for 
Stories. In Daniel G. Bobrow, editor, Representation and 
Understanding. Studies in Cognitive Science, pages 211–
236. Academic Press, New York et al., 1975. 
[Ryan, 1991] Marie-Laure Ryan. Possible Worlds, Artificial 
Intelligence, and Narrative Theory. Indiana University 
Press, Bloomington and Indianapolis, 1991. 
[Turner, 1994] Scott R. Turner. The Creative Process: A 
Computer Model of Storytelling. Lawrence Erlbaum, 
Hillsdale, 1994. 
[Ushie, 1988] Yukiko Ushie. ‘Corepresentation’. A Textual 
Function of the Indefinite Expression. Text, 6(4):427–
446, 1986. 
[Wiebe and Rapaport, 1988] Janyce M. Wiebe and William 
J. Rapaport. A Computational Theory of Perspective and 
Reference in Narrative. In Proceedings of the 26th 
Annual Meeting of the Association for Computational 
Linguistics, pages 131–138, Buffalo, NY, June 1988. 
