A Computational Mechanism of the “Distance” in Narrative: A Trial in

A Computational Mechanism of the “Distance” in Narrative:
A Trial in the Expansion of Literary Theory
Takashi Ogata ([email protected])
Department of Medicine and Engineering
University of Yamanashi, Takeda Kofu 4-3-11 Japan
Sayaka Yamakage ([email protected])
AXIS Software, Japan
1. Introduction: Narrative Generation Process and the Aspect of Discourse
A narrative generation process consists of three structural aspects of story or narrative content, discourse, and
representation or narration. These narratological aspects are executed through a kind of communicative power between
narrator and receiver. In a portion of our research program, at first, we analyze narrative texts and literary texts based on
narratology and literary theories to understand the knowledge structure and its characteristics on each aspect of relatively
different narrative processes. Next, we implement computer simulation systems using computational techniques and
cognitive theories expanded by narratology and literary theories. At Last, we consider about a more general system of
narrative and literary knowledge and simultaneously plan a whole narrative generation system integrating organically it.
We call such framework of research expanded literary theory (Ogata 2002). First of all, we would like to explain two
important narrative aspects and our approaches to them as a premise of the discussion.
Story, first category, means a temporal sequence of events to be narrated and a kind of narrative potential structure.
As a starting point, we applied a famous and classical theory by V.Propp, a Russian folklorist and literary theorist, in
order to use as the mechanism of an experimental story generation system (Ogata 1991). On the other hand, narrative
discourse is an aspect how to narrate a story in various ways, and is directly reflected to the structure of surface text. We
have done the expansion and precision (and their integration) of each technique in the discourse aspect through
computational modeling.
Describing concretely, Genette’s discourse theory (1972) is composed of following three big categories: “time” or
“tense”, “mood”, and “voice”. Following explanation is based on the summary of Prince (1987). Time or tense is the set
of temporal relations - speed, order, distance, etc. – between story (narrative contents) and discourse. Order is one of its
subcategories, and shows the order in which events occur and the order in which they are recounted. Duration means the
relation between story’s time (narrative world’s time) and discourse time (described time). Frequency is a category that
means the set of relations between the number of times an event occurs and the number of times it is described. Mood,
next major category, is the set of modalities, distance and point of view or focalization regulating narrative information.
Distance, first subcategory of two major factors in mood, means whether a narrator shows (the aspect of mimesis) or tells
(the aspect of diegesis) narrative information. Namely, showing is taken to institute less distance than telling. Perspective
or point of view or focalization is the perceptual position in terms of which the narrated situations and events are rendered.
It is divided into three types: zero focalization, internal focalization and external focalization. Voice, last major category,
treats the temporal and spatial relations between a narrator’s narrative action and a narrated text.
Until now we have implemented some simulation systems by Common Lisp based on some theories of Propp and
Genette (Ogata 1991, Hosaka & Ogata 2002, Mukouyama & Ogata 2002). In this paper, we would like to show a
computational approach to the concept of “distance”, and discuss future directions for more precise modeling.
In artificial intelligence and cognitive science, some researches about narrative generation system used techniques
and theories such as story grammar, scripts and goal-plan was attempted, and their knowledge representation technologies
are currently progressing. But in previous researches in this area, comprehensive and systematic exploration about the
knowledge of literary themes like story, discourse, interaction or communication between narrator and narratee, and so on
has not done. Therefore, in a research program which we call expanded literary theory, we are thinking the acquiring and
modeling of more deep and essential literary knowledge and techniques using technologies and concepts of artificial
intelligence and cognitive science.
144
2. Mood and Distance as one of the Subcategories
There are “perspective” and “distance” as lower classes in the narrative mood that processes “the degree of the presence in
the information of a story”. The perspective means the category for determining “the quality of the presence in a story ”,
and the distance means “the quantity of the presence in a story ”. The concept of distance is originated with Plato. Plato
called the state that narrates in the name of poet (author) itself a story “diegesis” (pure narrative discourse), and the state
that a story is narrated without the intervention of any poets or authors “mimesis” (imitation). Genette defined these
relations as the problem of distance between a narrator and a story or the information to be narrated. From the viewpoint,
the state of diegesis means longer distance, and the one of mimesis means short er distance. In the state of longer distance,
we can find various characteristics or narrative techniques, for example like the compression of information in a story, the
emphasis of a narrator’ existence or the exposure of a narrator itself, the insertion into discourse of the narration by a
narrator itself, the usage of indirect speech in language expression, and so on. On the other hand, in the state of short er
distance too, we can find many discourse methods like descriptive and dramatic narration, excessively detailed description
of events and objects, the usage of internal monologue, the usage of direct speech in language expression, and so on.
Genette also describes the law of “the degree of a narrator’s existence in a story + (descriptive quantity about the) story =
equal”.
Because Genette himself did not explicitly shown any concrete narrative techniques for operating the concept of
distance, we analyzed a novel (Miyabe 1993, pp.138-140) in order to make a hypothesis about the operational narrative
techniques that embodies the conceptual idea of distance. In this analysis, we regarded the original text as a point of
criteria, and made two types of text s with longer and shorter distance than each original text. An example of the result is
shown in table 1 (original language is Japanese, and this example is translation by us). Through this analysis, we
considered on the problems and characteristics in the operation of distance.
Table 1: An example of the analysis
Original texts
Toru says and turns in the package.
It is wrapped in light brown water
repellent sheet, is banded with the hemp
string, is tied ribbon which is shaped
halting flowe r knot.
Longer distance
Toru turns in the package.
Shorter distance
The ribbon which is shaped halting flower knot
on light brown water repellent sheet, … , This is
very strange gift. Whoever sent me this?
From the viewpoint of natural style in narrative, we think that in the case of lon ger distance, the compression of narrative
information based on various concrete ways by the intervention of a narrator itself is an important type of techniques, and
in the case of shorter distance, the internal monologue of characters is an essential technique. In first experimental
simulation system for considering the operation of distance, we mainly use these kinds of techniques.
In addition to above consideration, we would like to point out an important problem about narrator. Namely, if we
try to generate an extremely long narrative in distance, the narration exceeds the range of the story ’s content or the story
world. In such case, we thought that the narration about the world of the narrator itself that does not have any relations
with the story should be generated. If narrative discourse is the simply transformational methodology based on a story, the
narration with the extreme long distance will become the narration in which the story does not exist. For solving such
problem, through the discussion with a literary theorist (Aoyagi 2003), we set “discourse world” that defines the world of
the narrator and the receiver themselves relating with a narrative discourse, and creates the concept of “narrator’s events”.
3. A Computational Mechanism of Distance
We explain the overview of a simple simulation system of distance in this section.
Input data into the system are a sequence of events, and for each event, a user inputs the value of distance from 0 to
10. The system interprets that if the value is nearer 0, the degree of mimesis is higher, and if it is nearer 10, the degree of
diegesis is higher. Based on these data, the system processes next four operations: (1) Renewing an events sequence using
a story tree, (2) removing event slots, (3) transforming to narrator’s events, and (4) determining original narration by the
narrator.
(1) Renewing an events sequence using a story tree: All leaves and nodes in the tree correspond to events, and higher
level’s nodes mean more abstract events, and lower level’s ones mean more concrete events. A sequence of events as
input data correspond to node(s) or leaf (leaves) in the story tree that the system preliminary holds. However, child nodes
under a same parent are all included in the sequence of events or is not all included in the sequence.
When the values of distance in events with a same parent are sufficiently high, the system removes all child events
from the sequence of events, moves up the hierarchy in the tree by one step, and substitutes the child events into the place
where the parent event existed. Figure.1 and figure.2 show this processing. If the part of nodes surrounded by a net in
145
figure 1 is an original sequence of events, renewed sequence of events is the part of nodes surrounded by a net in figure 2.
Conversely, if the values of distance of a parent event in a sequence of events us sufficiently low, the system removes the
parent event from the sequence of events and replaces the same place by the child events. Going up in a hierarchy means
more abstract processing, namely longer distance, and going down means more concrete processing, namely shorter
distance.
The generation of a new event sequence
The movement of the class
n
n
If the value of distance with regard to the event that colored
red is large, class of tree is transferred upward.
The system generates the new event sequence from the
inputted event sequence by value of distance.
Longer distance
Story tree
Story tree
summary
Class is moved
by value of distance
summary
Class is moved
by value of distance
event
explanation
event
explanation
description
description
input
(
event frame)
Shorter distance
The inputted value of distance is large
Figure 1: Operation on a story tree (1)
Figure 2: Operation on a story tree (2)
(2) Removing event slots: The system removes slot or slots in each event in the sequence of events after above operation.
The number of slot in an event to be removed is determined by the rate based on the value of given distance. For example,
if the value of distance is “4”, forty percent slots are removed. The selection of the slot(s) to be removed is (are) decided
at random.
(3) Transforming to narrator’s events: The system transforms the value of distance that each event has to the percentage,
and interprets it as the probability for generating of narrator’s events.
(4) Determining original narration by the narrator: After the generation of narrator’s events was determined, the system
uses a knowledge frames network for determining narrative objects that does not have any relations with the story itself.
The knowledge frames network is a knowledge structure that has events and various concepts linked to them like
characters, places, objects, human relations, etc. Using this, the system decides the number of links to search according to
the value of distance, and selects the slots in the knowledge frame that it finally reached as the object of narrator’s event.
Figure.3 shows the overview of this process. At the same time, the system generates the narrator’s mental actions like
opinion and imagination using specially prepared slot ’s values in the event frame.
闘う0
リドル
語り
イ
手イ
ベント
ベント
(event
(actor リドル)
(object ヴォルデモート
卿)
(goal ハリー))
(人
(名前 ト
ム・
マールヴォロ・
リドル)
(性別 男)
(出身地 日記帳)
(外観 聡明でハンサム))
物
イベント
ハリー
イベント
:
:
語り
語りの対象
の対象
奪う
ハリー・
ジェームズ・
ポッター)
男)
(ジェームズ リリー))
ホグワーツ)
(ロン ハーマイオニー))
勝利する2
所属
(人
(名前
(性別
(両親
(所属
(友人
:
闘う1
明かす
話す
現れる
呼ぶ
出現する 闘う2
治る
刺す
人
Knowledge frame network
怪我をする
Figure 3: Determining original narration by the narrator
勝利する
Figure 4: A story tree used in the experiment
4. Consideration about the System
The system was implemented by Allegro Common Lisp on Windows. As the material of first experiment, we used a
simple and short text based on Harry Potter (Rowling 2000, Terashima 2002). The program is basically written by
Japanese, but we show English partially for explanation. Figure 4 shows a story tree used in the simulation. The part of
146
消える
net is a sequence of events corresponding to input data. Table 2 shows input events and values of slots corresponding to
each event that the system preliminary holds. From Table 3 to table 5 are respectively results of simulation based on
different values of distance.
Table 2: Input events and corresponding slots and their values
明かす[reveal]
Event
Slots
actor リドル[Riddle] / object ヴォルデモート[Voldemort] / goal ハリー[Harry]
現れる[appear]
actor フォークス Fawkes / item 組み分け帽子[sorting hat] / goal ハリー[Harry]
呼ぶ[call]
actor リドル[Riddle] / object バジリスク[basilisk] / goal 秘密の部屋[chamber of secret ]
出現する[appear]
actor 組み分け帽子[sorting hat] / item 剣[sword] / goal ハリー[Harry]
闘う 2[battle 2]
actor ハリー[Harry] / object バジリスク[basilisk] / item 剣[sword]
勝利する 2[defeat 2]
actor ハリー[Harry] / object リドル[Riddle] / item (毒牙[sting] 日記帳[diary])
Table.3: Result (1): the case that events moves up a hierarchy in operation (1)
Event
Renewed event(s)
Value of
distance
7
明かす
[reveal]
明かす
[reveal]
現れる [appear]
闘う 1 [battle 1]
9
勝利する 2 [defeat
2]
4
呼ぶ [call]
出 現 す る
[appear]
闘 う 2 [battle
2]
勝利する 2
[defeat 2]
Result
(narration-event
(type 空想[imagination])
(object ((用途[use] (書く[write] 残す[remember]))
(名前[name] 日記帳[diary]))))
(narration-event
(type 意見[opinion])
(object ((友人[friend] (ハリー[Harry ] ハーマイオニー[Hermione] ))
(姉妹[sister] ジニー[Ginny] )
(兄弟[brother]
(ビル[Bill] チャーリー[Charlie] パーシー[Percy] フレッド[Fred] ジョージ[George]))
(所属 ホグワーツ[Hogwarts] ) (両親[parents] (アーサー[Arth ur] モリー[Molly]))
(外観[appearance] (そばかすだらけの顔 [freckled face]
背が高い[tall] 痩せている[thin]))
(性別 男[male]) (名前[name] ロラルド・ウィーズリー[Weasley]))))
(narration-event
(type 感想[impression])
(object ((外観[appearance ] 聡明でハンサム[cool and handsome])
(名前[name] トム・マールヴォロ・リドル[Tom Marvolo Riddle]))))
Table 4: Result (2): the case that events moves down a hierarchy in operation (1)
Event
Renewed event(s)
Value of
distance
Result
明かす
[reveal]
奪う [rob]
話す [speak]
2
4
現れる[appear]
現れる [appear]
4
呼ぶ[call]
呼ぶ [call]
5
出 現 す る
[appear]
闘う 2[battle 2]
出現する [appear]
3
(event (actor リドル[Riddle]) (object ハリー[Harry]))
(event
(actor リドル[Riddle])
(object
“TOM MARVOLO RIDDLE” = ”I AM LOAD VOLDEMORT ”)
(goal ハリー[Harry]))
(narration-event
(type 感想[impression]) (object ((種族[tribe ] 不死鳥[phoenix]))))
(narration-event
(type 予想[anticipation])
(object ((性別[sex] 男[male])
(所属[belonging] ホグワーツ創始者[founder of Hogwarts]))))
(event (actor 組み分け帽子[sorting hat]) (goal ハリー[Harry]))
闘う 2 [battle 2]
1
(event (actor ハリー[Harry]) (object バジリスク[Basilisk]))
勝利する 2
[defeat 2]
治る[recover]
刺す[stab]
0
7
消える [disappear]
6
(event (actor フォークス[Fawkes]) (object ハリー[Harry]))
(narration-event
(type 空想[imagination])
(object ((名前[name] アルバス・ダンブルドア[Albus Dumbledore])
(所属[belonging] ホグワーツ[Hogwarts] )
(外観[appearance]
(背が高い[tall] 淡いブルーの眼[pale blue eyes])))))
(narration-event
(type 願望[wish])
147
(object ((出身地[hometown] 日記帳[diary])
(名前[name] トム・マールヴォロ・リドル[Tom Marvolo Riddle])
(外観[appearance] 聡明でハンサム[cool and handsome]))))
Table 5: Result (3): the case that moving on a hierarchy in operation (1) did not occur
Event
Value of distance
Result
明かす[reveal]
9
現れる [appear]
8
呼ぶ[call]
6
(narration-event
(type 意見[opinion])
(object ((所属[belonging]) (両親[parents] (ルシウス・マルフォイ・ナルシッサ[Lucius Malfoy] ))
(名前[name] ドラコ・マルフォイ[Draco Malfoy] ) (性別[sex] 男[male])
(外観[appearance] (尖った顎[sharp jaw]
プラチナ・ブロンドの髪[platinum blond hair] 灰色の目[gray eyes])))))
(narration-event
(type 批判[criticism])
(object ((所属[belonging] ホグワーツ[Hogwarts]) (性別[sex] 男[male])
(名前[name] アルバス・ダンブルドア[Albus Dumbledore])
(外観[appearance] (背が高い[tall] 淡いブルーの眼[pale blue eyes])))))
(distanced-event (object バジリスク[Basilisk]))
出 現 す る
[appear]
闘う 2[defeat 2]
3
(distanced-event (goal ハリー[Harry]) (actor 組み分け帽子[sorting hat]))
7
勝 利 す る 2
[defeat 2]
6
(narration-event
(type 空想[imagination])
(object
((名前[name] ホグワーツ魔法魔術学校[Hogwarts school of witchcraft and wizardry])
(メンバー [member] (ハリー[Harry] ロン[Ron] ハーマイオニー[Hermione] フレッド[Fred]
ジョージ[George] パーシー[Percy])))))
(distanced-event (object リドル[Riddle]))
In table 3 and table 4, the number of events is changed by operation (1). For example, in table 3, events of “appear”, “call”,
“battle2” are abstractly described in an event of ”battle 1”. Conversely, in table 4, “reveal” is given detailed description
by “rob” and “speak”,and “defeat 2” is described by three events of “recover”, “stab” and ”disappear”. These show
examples in the operation of distance through the abstract arrangement and the concrete one of input story but in more
macro level.
Next, we think about the operation of distance on each event. Table 6 is the pair of input events and output example,
same as above result (3) of table 5. However, we translate original conceptual expression to natural language.
Table 6: Natural language expression of result (3)
Value of
Narrator’s events
distance
9
?
8
?
6
×
3
×
7
?
6
×
Input (natural language expression)
Output (natural language expression)
Riddle reveals to Harry that his identity is
Voldemort.
Draco Malfoy who has Lucius as father has sharp jaw which
same his father and platinum-blond hair. He is in the
Hogwarts which same Harry. But I' m afraid it is mistake that
they set same school. (opinion)
Fowkes appears in the front of Harry.
What did Dumbledore who is principal of Hogwarts school of
witchcraft and wizardry consider and sent Fowkes to him? I'm
afraid he should send more helpful thing to him. (criticism)
Riddle calls Basilisk into chamber of secret. Basilisk which is called "king of snake" as another name
appears.
Sorting Hat appears in the front of Harry.
Sorting Hat is hat to decide freshman's house. Harry is struck
to dump because he sees it which is ragged and full of patches.
Harry fights with Basilisk.
If Hermione and Ron were here now, Harry would be saved.
(imagination)
Harry defeats Riddle.
Riddle lost and left jeremiad and disappeared.
In table 6, first line shows the output by long distance. This information has the description of “Draco Malfoy” which
does not be related directly to the event in the story and narrator’s opinion about him. This is an example that narrator’s
description occurred according to the operation of narrator’s events. Conversely, fourth event shows an example of short
distance. In this case, first input information is only slots of “sorting hat ” and “Harry ”, but the description about “sorting
hat” and the explanation of Harry ’s action are generated. Because the frame of Harry is linked to other frames, the narrator
can create different expressions using this knowledge network.
148
5. Conclusion: Toward the Integration of Aspects of Narrative Discourse and Narrative
Generation
To get the evaluation on this research from the side of literary theory and narratology, we asked comments and opinions to
a literary researcher (Aoyagi 2002-2003). At first, the idea that we modeled the processing of distance as the move on the
hierarchy of a story tree was evaluated at the point that we defined as an operational processing of conceptual discussion
in literary theory. Next, we got a comment that the insertion of narration by a narrator itself, namely the idea of “narrator’s
events” is a new idea, because any examples for trying to model the concept of distance from the viewpoint of information
or system did not exist until now, and our main goal was the hypothetical proposing of an idea for expanding narrative
discourse theory to embody the comparatively conceptual discussion in literary theories and narratology by more
concrete techniques like narrator’s events, the operation of story tree and so on.
However, as this research itself is very simple attempt, there are many problems: (1) The abstraction of more
concrete and detailed techniques for operating distance. Of course, because distance itself is not the concept directly
corresponding to any concrete techniques and is a kind of strategic knowledge for discourse, we should find rhetorical
techniques contributing for the realization of distance based on texts analysis. (2) We should consider the processing of
each event in the case of short distance. For example, we can think the generation of subjective information in each
character itself like opinion or internal monologue to realize more detailed narration. As more directly technical
possibility, there is the adding of new slots into frames. (3) The generation of discourse that has medium distance. For
example, a kind of synthesis of narrator’s events and the operation of slots will also necessary. We can mix a narrator’s
event generated in the distance “
5”and a normal event removed fifty percent slots inside it.
At last, originally, there are many overlaps in lower techniques in narrative discourse theory. For example, more
detailed narrative techniques like description and explanation are used in different categories in narrative discourse theory
like distance discussed in this paper and duration. Especially, the problem of distance is very abstract theoretical category,
and probably a kind of strategic category for controlling the use of various techniques.
Other many categories, for example duration and temporal order, are comparatively smoothly classified into lower
level’s techniques like Genette himself showed the hierarchical classifications. But, in the case of distance as a kind of
ideal discourse strategy, the analysis and modeling from the viewpoint of mutual interaction with other various discourse
categories become important. But, from the style of traditional literary study based on the conceptual discussion about
reading and interpretation of literary text s, it will be very difficult to get to the penetration about interactions among
discourse techniques and a whole model as an organic integration of various discourse techniques because of the lack of
means for experiment and logical accumulation. A new style of literary study by computer modeling and simulation
described in this paper is effective for the consideration of more detailed discourse classification. But in the very abstract
strategy like distance, it works as a very useful method for the generation of a hypothesis what techniques are used for
realizing such abstract strategy too.
References
Aoyagi, E. (2002-2003). Private e-mails (24/11/2002, 7/12/2002, and 23/2/2003).
Genette, G. (1972). Discours du recit, essai de methode, Figures III, Paris:Seuil. (Translated by Hanawa, H. and Izumi, R.
『物語のディスクール』, Tokyo:Suisei-sha, 1985.)
Hosaka, Y. and Ogata, T. (2002). Story Generation Based on Narratology and Representation with CG Animation,
Proceedings of 17 th Congress of the International Association of Empirical Aesthetics, 561-564.
Miyabe, M. (1993). Warera ga Rinjin no Hanzai, Tokyo:Bungei-shunnjuu-sha.
Mukouyama, K., Shinohara, K., Kanai, A. and Ogata, T. (2002). Rhetorical Analysis and Automatic Editing of the Film,
Proceedings of 17 th Congress of the International Association of Empirical Aesthetics, 571-574.
Ogata, T. and Terano, T. (1991). Explanation-Based Narrative Generation Using Semiotic Theory, Proceedings of
National Language Processing Pacific Rim Symposium 91, 321-328.
Ogata, T. (2002). Expanded literary theory: Cognitive/computational expansion of literary theories and narratology,
Proceedings of the 17th Congress of the International Association of Empirical Aesthetics , 163-166.
Prince, G. (1987). A Dictionary of Narratology, The University of Nabraska Press.
Rowling, J.K. (1998). Harry Potter and the Chamber of Secrets , Bloomsbury Publishing Plc.. (Japanese translation by
Matsuoka, Harry Potter to Himitsu no Heya, Tokyo:Seisan-sha, 2000.)
Terashima, K. (2002). Harry Potter Daijiten (The Encyclopedia of Harry Potter ), Tokyo:Toyo-kan.
149