While storytelling is a form of human artistic expression, we conceive of story writing and, in particular, script-writing for TV and movies as being built upon a few fairly known generative principlesa structure. Literature researchers have identified structural similarities between different stories. They claim that most stories can be attributed to a fairly small set of unique plots [3,12] about a few archetypal characters [28]. The Hollywood cliché is of the studio executive telling the script-writer: "Give me the same thing... only different!". Familiar characters and stories come with pre-existing audiences; therefore, writers can exploit pre-existing expectations [31], reduce the exposition time, introduce creative scenes, abuse the characters for comic ends (e.g., Johny English movies abuse the James Bond character) and break old social Taboos (e.g., Gay love stories).
Different authors defined different categories of story types: 7 basic plots by [3], 10 story types by [9], 20 master plots by [12], and 36 dramatic situations by [7]. The different typologies are not distinct from each other. They contain some overlapping definitions, or define similar concepts from different angle or granularity. We chose to build our research upon the 10 story types of [9] due to their detailed and clear definitions with plenty of examples. Revealing the high-level structure of movies (i.e., the story type and its main elements) is a major aspect of understanding the movie plots. Humans can identify and understand most of the elements of a story, such as the characters and their motivations, events and their consequences etc., and can categorize the story into one of the story types. However, current movie analytics technologies In: R. Campos, A. Jorge, A. Jatowt, S. Bhatia, M. Litvak (eds.): Proceedings of the Text2Story'23 Workshop, Dublin (Republic of Ireland), 2-April-2023 EMAIL: liuc@post.bgu.ac.il (C. Liu); armin@bgu.ac.il (A. Shmilovici); mlast@bgu.ac.il (M. Last) ORCID: 0000-0002-4020-6025 (C. Liu); 0000-0002-7202-6360 (A. Shmilovici); 0000-0003-0748-7918 (M. Last) are able to detect only relatively primitive story elements, such as the human actors and some low-level actions [13,22]. Researchers in the computational narrative understanding community have recently made progress in understanding narratives in text (books and movie scripts) [25,24,14,29,2,33].
We decided to focus on movies for story types, because movies present a relatively easy case: due to medium restrictions (e.g., fairly standard movie length), movies typically present relatively simple stories (e.g., one distinct main story, one or few main characters). Furthermore, most movies have a distinct inner structure [32]. To the best of our knowledge, we are the first to study story type classification for movies. Another major and common challenge faced by narrative understanding community is the collection of large-scale, reliably annotated datasets [1,19,25], especially in the movies domain.
Our first objective is to build a labeled dataset of movies to facilitate the use of supervised machine learning algorithms for the problem of story type classification. A crowd-sourcing experiment is used for constructing the dataset. The collected labels are analyzed to verify the following two hypotheses: (1) most movies adhere fairly well to a general structure, described by the screenwriting book [9] and (2) even non-experts can identify those story types after watching a movie.
Our second objective is to provide a lightweight solution for the challenging task of story type classification, with the use of relatively simple methodology and features. Although a fully automated pipeline is desired, please note that at this stage we substitute the audio-visual movie information with the dense scene-by scene textual descriptions and manually annotated features from the MovieGraphs dataset [15,27] to avoid the errors prevalent in the existing scene splitting, character identification, and other automated movie annotation tools, which often fail in common cases such as darkly illuminated scenes.
The original contributions of our paper to the domain of computational narrative understanding of movies are two-fold: a) We introduce the first benchmark dataset for the problem of story type classification that will be released to the research community; and b), We demonstrate that the story type of a movie can be automatically detected using some relatively simple movie features. Please note that in this study, we used the dense textual and feature-rich descriptions of the MovieGraphs dataset [15,27] The rest of the paper is organized as follows: Section 2 presents the definitions of 10 story types; Section 3 describes the dataset construction; Section 4 describes our feature engineering methodology for story type classification; Section 5 presents our classification experiments and their results; and Section 6 concludes with a discussion.
Below, we describe the 10 story types defined by [9]. The most important point to mention is that there are three abstract elements characterizing each story type, and those core elements can be used to distinguish between the different story types. From the story type we can infer the existence of some abstract concepts in the movie plot, such as a monster character or a detective character.
A troubled hero's only way to overcome a growing life crisis is to defeat his worst enemy -himself. Examples: Brooklyn, Inside Out. 6. Buddy Love: An inadequate hero must rise above an extremely difficult situation to be with a uniquely unlikely partner who is the only one capable of bringing him peace. Examples: E.T., Zootopia. 7. Whydunit: A devoted hero must find the truth behind an intriguing mystery before he is swallowed by the darkness he desperately seeks to expose. Examples: Captain Marvel, Bladerunner, The Silence of the Lambs. 8. The Fool Triumphant: An innocent hero whose only way to defeat the prejudices of a group is to change himself without losing what made him the group's target of contempt in the first place -his uniqueness. Example: Moneyball. 9. Institutionalized: An outsider whose only way to save his individuality is by going against the many who wish to make him like them. Examples: American Sniper. 10. Superhero: A uniquely special hero must defeat an opponent with stronger capabilities by using the same powers that disconnect him from the people he hopes to save. Examples: Iron Man, Taken.
For example, the movie Die Hard belongs to the Dude with a Problem type, where an "innocent hero" (a bullheaded policeman) visiting his wife is accidentally involved in a "sudden event" (terrorists crash an office party and take his wife as a hostage) and he has to use his brains rather than his might in a "life or death battle" to subdue the high-tech terrorists and save his wife. Sometime a story-type is also written for the antagonist character to make it more realistic. For example, from the terrorists' point of view, the movie Die Hard belongs to the Monster in the House type: in a secluded office building, their sin of greed has awakened a mighty and vengeful monster (the hidden policeman) who kills them one by one until they figure out his weakness (his hostage wife) and use her to set up a trap before a final direct confrontation. Some of the story types are typically used for producing some of the "standard" movie genres. For example, most horror movies are of the Monster in the House type; most detective movies are of the Whydunit type; and most romantic comedies are of the Buddy Love type. The same story can be told using different story types and genres. For example, a superficial story about a fictitious British secret agent who fights an evil secret organization that threatens the safety of the world can be told with different story type twists for producing a) action movies: in the James Bond movies (typically of the Superhero type) the character has superhero-like fighting skills; or b) comedy movies: the Johnny English movies are of the Fool Triumphant type. The character is innocent about his clumsiness and ridiculed by the establishment in which he operates.
Movie Selection. We limited our movie selection to the 51 movies used in the MovieGraphs dataset [15,27]. This dataset contains 7,637 manually calibrated scene boundaries. The description of each scene contains detailed manually provided information such as the characters' appearance, the characters' relations (e.g., parents), character interactions (e.g., greeting), characters' emotion etc. The reason we chose to use this dataset is that it is rather complete in terms of the available information that is crucial for narrative understanding. We had to discard 6 movies, either because we could not obtain the exact version of the movies used in the MovieGraphs dataset, or because of an irrelevant movie genre from point of view of the plot (e.g. Biographical movie), and eventually labeled 45 movies.
Story-Type Annotation. For the story type labeling, we selected 119 human annotators (out of 180 applicants, all senior undergraduate students in the Software and Information Systems Engineering Department), based on their English proficiency level and their level of interest in watching movies. During the annotation process, we ensured that: (1) Each annotator was assigned at least 3 movies (including one of the 9 movies with "gold-standard types"); and (2) Each movie was annotated by at least 5 different annotators (in practice, except for one, all movies were assigned to 6 or more annotators). The annotators were provided with the guidelines that described in detail the background concepts and definitions of story types. They were asked to choose at least one story type for a given movie out of the 10 main story types described above, while there was no limit on the maximum number of types they could choose. To evaluate the annotator's attention during the task, ten simple quizzes about the movie plot were embedded in each movie. The annotators were motivated by 2 bonus points for their course grade. The annotation was approved by the department's Ethics Committee.
Data Analytics. Our annotated dataset consists of 45 movies, with each movie labeled by 1 to 3 story type labels. There are 17 movies with only one label, 16 movies with two labels and 12 movies with three labels. On average, each annotator selected 1.5 story types per movie, and for 40 out of 45 movies, at least half of the annotators agreed on the same story type. The above agreement between annotators on so many movies confirms our first hypothesis that most movies adhere fairly well to a general structure described by the screenwriting guidebook [9]. All 10 story types assume that a movie has a single storyline. Cases of poor annotator's agreement were more common for the few movies with multiple storylines, (e.g., Crash (2004)).
To evaluate the quality of annotations, we compared our collected story type labels with the "gold standard" labels for the 9 movies for which we have "gold-standard story types" (labeled by professional scriptwriters who are proficient in the Save the Cat! Theory [4,6]). After weighting the students' annotations, the agreement between the experts' labels and weighted students' labels increased to 8 out of 9 "gold standard" movies, indicating that the weighting strategy2 correctly strengthens the annotators who understood the concepts better and weakens those who did not. The agreement between the collected labels and the gold ones confirms our second hypothesis that even non-experts can identify the story types after watching a movie and reading the guidelines.
The three most frequent story types in the dataset are Buddy Love, Dude with a Problem, and Rites of Passage. On the other hand, story types such as Superhero, Out of the Bottle, Monster in the House and the Fool Triumphant are rarely selected. The type imbalance has become more severe due to our limited selection of movies. As common in such cases of severe imbalance [9], we kept only the three most frequent story types and replaced the rest with an additional "Other" label. The infrequent labels aggregation may not be necessary in a larger, less imbalanced dataset.
Story Type Classification. In our evaluation experiments, we use two popular classification algorithms, C4.5 Decision Tree and Naive Bayes, with 5-fold cross validation. Considering the limited size of the particular dataset we use, instead of performing multi-label classification, we classify movies with respect to each story type separately, through binary classification (one-against-all strategy). This limitation can be resolved once a larger scale dataset is available.
Feature Extraction. The 10 story types contain subjective and highly abstract concepts (such as "monster" or "fool") which are, apparently, difficult to infer from a few numerical features. According to the definitions, all stories are about the hero and the hero's surroundings. Characters are the essence of a story, and the attributes of the characters in a movie determine how the story develops: their activities, their emotions, the changes they undergo, their relationships, etc. Therefore, we mainly concentrate on the character-centric features, which measure the character importance and relationships.
Basic Features: Scenes are the basic narrative units of a movie. A typical scene consists of multiple shots. We use the MovieGraphs dataset to extract the scenes and its shots information from each movie. How rapid the scenes/shots change across the entire movie might, to some extent, reflect the complexity of oe A gui rre Ennis Del Mar assie the story. Moreover, the average number of shots per scene could be a clue to the intensity of the scene events, because a larger number of shot changes imply more event details, such as dialogue exchanges during conversations, moves during a fight, or even different memories during a flashback etc.
Character Network Features: In order to capture the relationships between the characters in a given movie, we build a static character network for each movie, applying the CoCharNet algorithm [23] to the MovieGraphs dataset. The character network is a weighted graph consisting of nodes representing the characters and edges representing the co-occurrences between each character pair. Nodes are weighted by the appearance duration of the movie characters, while edges are weighted by the co-occurrence duration of the pair of characters they connect. From the story type definitions it can be inferred that some types contain latent character relationships that can be discovered from their character network. E.g., a Fool Triumphant story is expected to have a single hero (the "fool") while a Buddy Love story is expected to have two key characters, the hero and the buddy. For Example, Figure 1 (middle) presents the character network for the Brokeback Mountain movie which is a distinct Buddy Love story where two nodes (of the two main characters) are distinctly stronger Based on the character co-occurrence networks constructed for each movie, we extract the features, which reflect information about the character's social relations in the movies. The average (avg.) and standard deviation (std.) of the node degree represent how strongly "connected" the characters in the movie are and the features of edge weights are expected to show how dominant the corresponding social connections are.
Temporal Key Character Features: It is reasonable to assume that an important character appears in almost all the scenes, and therefore has an almost uniform appearance distribution. Therefore, by computing the similarity between the characters' appearance distributions and the uniform distribution using Kullback-Leibler Divergence measure (KL divergence), we can estimate the importance of each character. The smaller the KL divergence, the more important the character is.
As described above, we replaced each of the 7 least frequent story type labels by the "Other" label. Together with the three most frequent story types, we now have four story categories for the classification task. Our experiments were run on an annotated dataset of 45 movies, including 26 movies of the Buddy Love type, 19 of the Dude with a Problem type, 13 of the Rites of Passage type, and 22 assigned with the Other label. Considering the small size of the dataset, instead of applying multi-label classification algorithms, we attempted to identify each story type separately, i.e., we conducted four binary (one-against-all) classification experiments on the four categories. We evaluated the performance of the classifiers by stratified 5-fold cross validation, i.e., 36 movies for training and 9 movies for testing in each fold, and the fold splitting was done separately for each story type. Moreover, in order to further minimize the influence of data imbalance, we weighted the samples with the inverse frequency of their label. We evaluated the following classification algorithms available from the Scikit-Learn Library [26]: Logistic Regression, Naive Bayes, Support Vector Machine with linear kernel, and C4.5 Decision Tree. Table 1 presents the results of the top performing algorithms.
We performed feature selection by computing the mutual information between each feature and every class label; then, for each label, we selected the top three, four and five features with the highest information gain (no matter which set they were from) for the classification task. The features were all normalized with L1 normalization.
There were two hyper-parameters in the experiment: We limited the maximum depth for the decision tree classifier to 3, to avoid overfitting. The KL divergence threshold was set to 1.0 for determining the number of key characters.
Table 1 presents the results of our classification experiments. It shows baseline results, best results (classifier + selected features), and the features used for obtaining the best results. The baseline results were computed by labeling all movies with the positive type in each binary classification experiment, respectively, leading to the universal 1.00 recall. Although we sometimes failed to significantly improve the F1 measure, as for the Buddy Love story type and Other story types shown in Table 1, the obtained precision and accuracy were encouraging, and passed the t-test (p value less than 0.05) for all story types compared to the baseline. As we expected, the importance of some features is story type dependent. For example, the edge weights in the character networks play an important role for identifying the Buddy Love movies, because in those movies the relation (edge weight) between the two main characters (the hero and his/her buddy) is expected to be much stronger than the other features, which can somehow be reflected by the average and standard error of the edge weights. The number of characters in the third quarter of the movie is the most important feature for identifying the Dude with a Problem type. Considering the movie deconstruction, the third quarter of the movie is likely to include the "confrontation" part of story, as well as the "development", in which the protagonist (the dude) experiences his/her essential change. During this period, it is likely that either his companions or his enemies appear. Another important feature is the number of edges, which reveals how many connections the characters have, instead of how strongly they are connected to each other. In a typical Dude with a Problem movie, almost every character is connected to the protagonist, leading to a high number of edges in the character network. The most important feature of the Rites of Passage type is the average degree of the graph, which reflects on average, how connected the characters are, and may give a hint of the social network in the story. Regarding the Other class, although it is a mixture of seven less frequent story types, the four most dominant features are all character-related, showing the characters' importance within the stories. Finally, basic features (e.g. scenes per minute, average scene duration, movie duration) also played important roles in identifying the three most frequent story types.
In this paper, we introduce a new challenge of movie story type classification. Our experiments suggests that most movie stories adhere fairly well to a general structure described by the screenwriting book [9]; and that even non-experts can identify those story types after watching a movie. We provide a benchmark dataset [10] and a lightweight machine learning solution for the task of story type classification, with the use of a relatively simple methodology and features. The character-centric features are highly related to movie stories and can be used as one of the main information sources for narrative understanding. We believe that many of those features have equivalents in stories not specifically written for the cinema (e.g., character co-occurrence networks from text [19]).
The value of story type classification for automated movie understanding is in extracting the high level abstract concepts associated with each story type, e.g., an innocent hero, a sudden event, and a life or death battle for the Dude with a Problem story type. These can provide some automatic understanding of the protagonist's character traits and the motivations that drive him in facing his challenges.
The main limitation of our study is that due to a small and imbalanced movie dataset, we could identify only the three most prevalent story types while aggregating the other story types into a single "other" category. We believe that more complex feature engineering may help to distinguish between some of the related story types (e.g., Whydunit and Dude with a problem). Further progress can be achieved with larger collections of annotated movies and automated feature extraction with deep learning. With a sufficiently large dataset of annotated movies, one could attempt to identify the 20 master plots by [12], and 36 dramatic situations by [7].
A practical application of movies story type detection may be the construction of a better movie recommendation system based on the implicit assumption that most movie lovers have a strong preference towards certain story types. The viewer feedback and behavior, which are the most widely used information sources for the construction of a movie recommendation system [5,8] can also be used for story type based movie recommendation. To facilitate the user choice, a movie search engine can be geared towards retrieving movies with similar narratives, protagonists with similar character traits [28], or specific scenes (such as a life or death battle).
This research was partially supported by the Israeli Council for Higher Education (CHE) via the Data Science Research Center, Ben-Gurion University of the Negev, Israel.