2020 年 5 巻 1 号 p. 4-21
In recent years, there have been attempts to generate stories automatically by utilizing computers and various other kinds of information technology. However, it remains unclear how story sequences can be made to look natural and appear interesting. This study conducted a data description of 134 famous Japanese comic detective stories and listed 37 types of plot elements, 10 types of tricks, 9 types of criminal motives, and 10 types of relationships between victims and criminals, in order to realize a computational narratological analysis. Based on these parameters, eleven factors have been extracted through factor analysis. The structures of the plot transition networks varied according to the factors present in the detective stories. Therefore, an elaborate, human-like, and understandable creation of plot transitions can be enabled based on the detailed differences between plot transition networks.
In the field of literary studies, research on narrative structures and story patterns is called narratology. Various investigations conducted in the field of narratology have been based on humanities methodologies. For instance, Propp (1968) insisted that the function of character roles in stories can be categorized based on several patterns in specific genres, namely thirty-one story functions and seven character roles in Russian folktales. Greimas (1966) hypothesized that the structures of general stories can be divided into and symbolized by particular elements; additionally, these elements can be classified into symbols. Likewise, Genette (1972) analyzed the stylistic and semantic differences between the sequences in story narration and their chronological order.
Based on the preceding research, there have been attempts to generate stories automatically by utilizing computers and various other kinds of information technology. To create stories, it is necessary to evaluate the naturalness and creativity of the plot elements’ sequences. To facilitate this, traditional approach is to use cognitive science theory about goals and plans (Schank 1977) to generate the behaviors of each character (Meehan 1977). This type of story generation algorithm is developed based on inference of next action (e.g. assuming Markov chain). As a development of this type, the What-If Machine Project (WHIM) constructed a database of common consequences for numerous situations and produced stories by selecting logically connectable events (Granroth-Wilding, Mark, and Stephen Clark. 2013).
Another traditional approach is to use existing plot structures to create new plot sequences (e.g. Bringsjord and Ferrucci 2000). This type of system is often based on Propp’s thirty-one functions (e.g. Gervás 2014). In another case, the Kimagure AI Project made templates of general story structures to engender various easily understandable stories (Toyosawa et al. 2018). In recent years, the problem of selection from various possible plots has been examined (Amos-Binks 2016). However, there are only a few case studies with empirical data about plot sequences based on existing stories. In sum, it is not scientifically clear how story sequences can be designed to look natural and appear interesting.
In this study, the detective story genre was selected to analyze plot sequences quantitatively. Detective stories are highly patterned, include many short stories, and allow for the collection of many sample works. In addition to that, in the simple typical works of detective comics for beginners that were selected for study, complex discourse transformations such as flashbacks, or parallel narratives rarely appear. Simplicity of plots enables focused quantitative analysis from the viewpoint of fundamental plot structure. In conclusion, detective stories for beginners were deemed to be an appropriate choice for this study.
Some story genres included common elements: a typical plot sequence, a characteristic protagonist, a patterned antagonist, common backgrounds, and frequently used text styles. This study focuses on the most fundamental narratological elements: plot and character patterns. In detective stories, plot patterns include the types of tricks played by criminals, in addition to plot sequences. Moreover, relationships between characters—especially between criminals and victims, which provide the motives for crimes—are critical in detective stories. Therefore, this study analyzes plot sequences, tricks, motives, and the relationships between criminals and victims. These elements were combined to extract distinct categories in detective stories. Based on the extracted categories, characteristic patterns of plot transitions and characters were obtained.
By utilizing these quantitative data, it is possible to derive the overarching patterns to compute the probability of any plot element preceding or following another. These computational patterns could serve as the foundation for automatic plot generation by computers in the future.
In this study, the Japanese detective comic series Case Closed (also known as Detective Conan; Japanese title Meitantei Konan) was selected for analysis (Aoyama 1994–). Case Closed is a best-selling detective series and is famous as a successful multimedia property in Japan. Case Closed has been serialized in the comic magazineWeekly Shōnen Sunday since 1994, with more than one thousand episodes. The most recent comic book installment was volume 94, published in December 2017. The total number of copies printed is estimated to be more than one hundred million. This work has been aired as an ongoing TV anime series for more than twenty years, since 1996, and has been investigated in studies of influences of Japanese culture (e.g., Cooper-Chen 2012). Moreover, gender-based character analysis (Unser-Schutz 2015) and case study plot structure analysis (Hatakeyama 2003) have been done for this comic series. However, quantitative plot structure analysis has not yet been attempted.
Case Closed includes much homage to traditional detective stories like those written by Arthur Conan Doyle and Agatha Christie. The plot style of each episode follows that of traditional detective stories. Moreover, because the main reader demographic of the series is comprised of young boys, the story structure tends to be very simple and easily understood, with explicit explanations of the tricks and mysteries utilized. For this study, the comics of volumes 1–45 were analyzed as the data, with volumes covering a ten-year period (1994–2003). Since the number of episodes differed from the number of actual cases, 134 cases out of 461 episodes were targeted for analysis. On average, three to four episodes were found to correspond per case.
Although there are many methods for categorizing story plots based on particular elements, this study focused on the general functions common in detective stories in order to describe plot elements. The macro-level functions of stories were used as a basis for dividing and categorizing plot elements manually. The story functions for 134 cases were then described and synthesized on the basis of the most frequently appearing typical functions of detective stories. As a result, the study depicted 134 cases with 37 types of plot elements. Table 1 displays the 37 types of plot elements. A total of 1,123 plot elements were obtained (Murai 2018). Each case includes an average of 8.4 plot elements.
The tricks in each case were categorized into ten types, as defined below. Some cases include several kinds of tricks. Therefore, the categorization of each case’s trick is a duplicate classification (one case might be categorized in several categories). Table 2 describes the results of manual categorization.
Table 2. Type and number of appearances per trick
Criminals’ motives in each case were categorized into nine types, as defined below. Some cases include several types of motives. Therefore, as in the case of tricks, the categorization of each case’s motive is a duplicate classification. Table 3 presents the results of manual categorization.
Table 3. Type and number of appearances of criminals’ motives
The relationships between victims and criminals in each case were categorized into ten types, as outlined below. Some characters have duplicated roles in their respective stories. Therefore, some classifications are duplicated, as in the former tables. Table 4 depicts the outcomes of manual categorization.
Table 4. Type and number of appearances of relationships between victims and criminals
To categorize and group the features of the plot sequences and character roles in detective stories, all parameters have been analyzed using factor analysis based on each case. Factor analysis statistically extracts groups of given parameters by rotating parameters as vectors in multidimensional space. Those groups are called “factors.” Factor analysis makes it possible to simplify relationships of many variables by grouping. Factor analysis extracts factor loadings between each parameter and factor. Each factor loading signifies the strength of the relationship between a parameter and a factor. Fundamentally, parameters which have large factor loadings in absolute value related to some factor are regarded as included in that factor.
In total, sixty-six parameters (thirty-seven plot elements, ten types of tricks, nine types of criminal motives, and ten types of relationships between victims and criminals) in 134 cases were analyzed as sixty-six dimensions in 134 vectors for factor analysis. The Promax rotation method was used as a method of matrix rotation in multidimensional space and a parallel analysis was performed to determine the number of factors. After the factor analysis, less significant parameters (with a maximum factor loading of > 0.25) were eliminated, and a subsequent factor analysis was performed repeatedly. Finally, after performing the factor analysis five times, eleven factors were identified.
In total, forty-two parameters (nineteen plot elements, seven types of tricks, nine types of criminal motives, and seven types of relationships between victims and criminals) comprise the eleven factors. Table 5 displays the resultant factor loadings; the bold font signifies cells whose factor scores exceed 0.25. In table 5, the type (in each respective row) indicates the kinds of parameters. Type “P” signals plot elements, “T” represents tricks, “M” describes motives, and “R” signifies relationships between victims and criminals.
Considerations and the naming of the factors are described below:
Table 5. Factor loadings of plot- and character-related parameters
Factor 10: This factor includes a victim who is a “criminal of a past crime” and the motive “past crime.” Therefore, this factor is about a “past crime.” This factor also indicates that the plot “suicide,” and the tricks “concealing a crime” and “fake criminals,” are often used in this story pattern. This includes the trick of making it seem like a victim who regrets a past crime has committed suicide.
Table 6 describes the factor correlations of the factor analysis. The bold font signifies the correlation values that exceed 0.25. If correlation value between two factors is large, it signifies that those two factors have a strong influence on each other. Factor 1, “eliminating the witness,” has a weak correlation with factor 3, “request for a greedy type of crime,” and factor 6, “abduction and fighting.” These correlations may indicate that witnesses can be easily killed in connection with violent and selfish crimes. Factor 2, “vengeance for a past death case,” has a weak correlation with factor 5, “love,” and factor 7, “mysterious criminal.” The correlation between factors 2 and 5 may be related to vengeance for a dead lover. Moreover, the correlation between factors 2 and 7 may signal that the avenger would like to hide their identity. Also, factor 4, “vigilance against theft,” and factor 9, “preliminary notice of a crime,” imply the story pattern of a phantom thief who is likely to steal.
Table 6. Factor correlations of plot and character-related parameters
Based on these factor loadings, the factor scores of each case can be calculated. The factor score signifies the relationship between factors and each of the original vectors that are composed of parameters. In this paper, the original vector signifies the parameter sets from one detective story. Therefore, the factor score for each detective story signifies the strength of relationship between factors. Table 7 shows the number of stories whose factor scores exceed 1.0 in each category.
Table 7. Number of categorized stories based on factor scores
In order to investigate the sequential relationships between each pair of plot elements, the transitions between those plot elements were visualized as a direct network (fig. 1) by utilizing the software Graphviz (Ellson et al. 2003). For visualization purposes, the numbers of sequential transitions for two different plot elements were aggregated in 134 cases (stories), and frequent transitions (more than 3) were employed as edges in the network. Each plot element was a node for the network. The nodes’ font sizes represent the frequency of each plot element.
Figure 1 shows that fundamental narrative patterns can be symbolized as transitions between typical functions in detective stories. Based on frequently appearing plot elements, the common story pattern in the series Case Closed can be described as a sequence of an “introduction,” the appearance of a “cadaver,” the “investigation,” the “past case,” “reasoning,” “confession,” and “arrest.”
Although figure 1 describes the average plot transition in all cases, each typical plot transition, according to the extracted factors, can be obtained based on factor scores. For instance, figure 2 describes the plot transitions of six cases whose factor scores of factor 4, “vigilance against theft,” exceed 1.0 (in table 7) based on frequent transitions (> 2).
Figure 2 shows that the plot transitions of the story pattern “vigilance against theft” are completely different from the average detective story pattern in figure 1. Although not as extreme as in figure 2, figure 3 presents an example of a slightly different pattern of plot transition. Figure 3 portrays the plot transitions of fourteen cases, whose factor scores of factor 7, “mysterious criminal,” exceed 1.0 (in table 7) based on frequent transitions (> 2).
Figure 1. The transition network of all plot elements in the detective stories
Figure 2. The transition network of factor 4–based plot elements
Although figure 3 is like figure 1, the plot element “incursion” and “exposed identity” are characteristics of the story pattern “mysterious criminal.”
Both the average plot transition pattern (fig. 1) and the individual plot transition patterns (based on minor story types) embody the typical structures of detective stories.
In the future, based on the quantitative structure of these transitional networks, human-like plot sequence generation could be enabled by utilizing mathematical methodologies, such as the random walks in the Markov chain model.
Figure 3. The transition network of factor 7-based plot elements
This study utilized a data description of 134 famous Japanese comic detective stories and listed thirty-seven types of plot elements, ten types of tricks, nine types of criminal motives, and ten types of relationships between victims and criminals to realize a computational narratological analysis. By utilizing factor analysis, eleven factors — “eliminating the witness,” “vengeance for a past death case,” “a request for a greedy type of crime,” “vigilance against theft,” “love,” “abduction and fighting,” “a mysterious criminal,” “a crime that has not happened,” “preliminary notice of a crime,” “past crime,” and “protection”—have been extracted. These factors related to the fundamental story patterns within the detective story genre.
The transition patterns between the various plot elements were visualized as networks. The structures of the plot transition networks changed according to the factors present in the detective stories. Because this paper targeted one of the most famous Japanese detective comics, the resultant factors and networks represent that work’s characteristics. However, by analyzing other detective stories using this methodology, human-like elaborate and understandable creation of plot transitions in general detective stories would be enabled based on those detailed difference of plot transitional networks.
Although another coder needs to verify the objectivity of the coding, this computational narratological outcome could be one of the foundations of automatic story generation by AI in the future.
Just as Propp’s results are limited to Russian fairy tales, the results of this analysis of detective stories are also limited to a single story genre. However, by accumulating analytic results from several genres, the whole shape of story plots may emerge in the future. Knowledge about story structures based on quantitative analyses could enable not only automatic story generation but also more objective and precise scientific interpretation of stories.