Current discourse re昀氀ects a growing skepticism towards contemporary popular culture, speci昀椀cally the realm of cinema, with an emerging consensus that its creative capacity is on a waning trajectory. This study introduces a novel approach which employs natural language processing techniques and embedding methods to measure semantic novelty of cultural items' descriptions. We apply this methodology to cinema, analyzing plot summaries of over 19,000 movies from the United-States spanning more than a century. Our measure's robustness is validated through a series of tests, including a 昀椀t with a genrebased novelty score, a manual inspection of 昀椀lms identi昀椀ed as highly innovative, and correlations with award recognitions. The application of our Innovation Score reveals a compelling pattern: an increase in the rate of cinematic innovation throughout the 20th century, followed by a stabilization in the rate of innovation in the 21st, despite an ever-growing production of 昀椀lms. Contrary to the o昀琀en-voiced lament that cinema is losing its innovative edge, our study suggests that the level of innovativeness in cinema is not in decline.
The 昀椀lm industry has been the subject of numerous debates regarding its perceived decline in innovation. Critics and audiences alike have voiced concerns about the increasing prevalence of sequels, remakes, and franchise 昀椀lms, arguing that these trends would re昀氀ect a lack of originality and creativity. These apprehensions have been ampli昀椀ed by the publicized sentiments of esteemed 昀椀lmmakers such as Francis Ford Coppola, who likened popular 昀椀ctions to “prototypes made over and over and over again,” Ken Loach, who compared them to “commodities like hamburgers,” or Alejandro Iñárritu, who dismissed them as “basic and simple.”. The apprehensions over the diminishing creative vigor in cinema are not con昀椀ned to academic or elite circles; they resonate deeply with the broader public. A quick glance at social media, 椀昀lm forums, and audience reviews reveals a torrent of sentiments expressing disappointment with the perceived stagnation of storytelling, reliance on formulaic plotlines, and the increasing tendency to prioritize pro昀椀t over artistic innovation. But is it the case? Is innovation in cinema on decline?
The challenge in de昀椀ning innovation, and in determining whether a product is innovative,
lies in its subjective nature. We de昀椀ne innovation as what is novel for humanity at large, as
opposed to novelty, which is what is contextually novel for an individu1a8l,1[
In this paper, we develop and apply a computational measure of innovation to movies based on summaries, which are standardized and rather homogenous in both IMDb and Wikipedia. This new measure is straightforwardly applicable, not to individuals, but to cultural products and aims at measuring theirobjective level of innovativeness. It could in principle be applied in di昀erent cultural domains, in di昀erent periods and countries, on di昀erent human productions such as scienti昀椀c papers, patented technologies, or literary novels—or any other products with textual descriptive metadata.
The Sentence-BERT (SBERT) algorithm 1[
SBERT excels at computing semantic proximity between words, phrases, or even entire paragraphs. It learns the contextual relationships between words (i.e., word embeddi9n,g1;0[]) and can calculate the semantic similarity between new and existing text based on shared context. This allows SBERT to accurately determine which words, sentences, or paragraphs are most similar to each other, even if they are not exact matches. For instance, in the context of cinema, SBERT could compute the semantic distance between the plot summaries oStfar Wars IV (1977) and Star Trek: The Motion Picture (1979), recognizing their relatedness and scoring them accordingly.
In our measure of cultural innovation, we utilize SBERT to encode descriptions of cultural products into 昀椀xed-length vectors that encapsulate the semantic meaning of the descriptive text. These vectors are then used to compute the pairwise cosine similarity between them, serving as a measure of their semantic similarity. The advantage of using SBERT over traditional bag-of-words models is its ability to capture the meaning of the text, rather than just the frequency of words, which is particularly useful when dealing with short texts and texts coming from di昀erent periods or di昀erent contexts—where words di昀er although meaning is similar.
To compute our measure of cultural innovation for each individual cultural product, we 椀昀rst encode their description (here, movie plots) into high-dimensional vectors using SBERT. We then compute the cosine similarity between each vector and all previous vectors. That is, we compute the similarity between each product and all previous products. By reversing this score, we transform the resulting similarity values into distance scores. Therefore, the 椀昀nal score, for each product, is computed asthe average of its distance scores from all previous products. Building on this methodology, it is important to note that our measure inherently captures the increasing di昀케culty of innovation as more products are released within the same domain. As the number of preceding products grows, the space for unique, unexplored ideas naturally shrinks, making it increasingly challenging to create something truly innovative.
In the domain of movies, here, the innovation score for a given movie thus quanti昀椀es how di昀erent its summary is from all previous movies in the dataset. This approach allows us to capture the level of novelty of a movie plot compared to the plot of all movies having been previously released.
Our measure of innovation is conceptually similar to other methods that compute innovation
by assessing semantic distances between individual cultural products and their predecessors
within the same domain. For example, in the realm of French theater, Ca昀椀ero and Gabay1][
have employed a similar approach, as have Kelly and colleagu6e]si[n the analysis of
technological patents (see 1[
We can further formalize this measure. Given a set of cultural products, each with a description , the Innovation Score (IS) for th e-th product is calculated as: = 1
∑(1 − − 1 <
⋅ || || ⋅ || || ) where = SBERT( ) is the high-dimensional vector representation of the descriptio n obtained using the Siamese-BERT (SBERT) algorithm⋅, denotes the dot product, and|| || denotes the Euclidean norm of vector . IS essentially measures how distinct or innovative a cultural product’s description is compared to the descriptions of all the previous products in the set. A higher IS would indicate that the product’s description is more unique and innovative within the given set. Conversely, a lower IS would suggest that the product’s description shares more similarities with the descriptions of previously seen products.
Our study utilizes a comprehensive dataset compiled from IMDb and Wikipedia, encompassing metadata for 19,254 movies produced in the United States. The IMDb data, obtained directly
We turned to IMDb plot summaries, which di昀er markedly from their Wikipedia counterparts in length and standardization. Despite these di昀erences, our Innovation Score remained consistent, demonstrating a signi昀椀cant positive correlation between the scores derived from both sources ( = .28 , < .001 ). This result underscores the adaptability of our measure, capable of capturing innovation irrespective of the descriptive metadata available.
To further assess the robustness of our Innovation Score, we altered the random seed used in the calculation process. This analysis consistently revealed a strong positive correlation ( = 0.89 , < .001 ) between the Innovation Scores obtained using two di昀erent random seeds. This substantial correlation reinforces the reliability and stability of our Innovation Scores, demonstrating its consistency even when varying the initial randomization.
The notion of cultural forgetting would suggest that a movie can appear innovative even if its narrative resembles older 昀椀lms, if it is at least di昀erent fromrecent ones. To examine this, we calculated three new Innovation Scores, progressively considering a narrower temporal window: we computed the average distance of each movie from movies released 10 years before, 5 years before, and just 1 year before. Strikingly, our analysis revealed that the level of innovation remained nearly identical across these di昀erent timescales (for all correlat io>ns.9,8 , < .001 ). This suggests that a movie’s innovativeness is a consistent trait, irrespective of the speci昀椀c timeframe under scrutiny.
Upon manual inspection, we found that our algorithm indeed identi昀椀ed movies widely acclaimed as innovative, such as2001: A Space Odyssey, Pulp Fiction, and Interstellar, as highly innovative (see Figure 2.A.). However, it is important to note that these are cherry-picked examples, and we could have chosen others that would not have aligned with our intuition. For instance, whileInception is also widely considered innovative, it received a relatively low Innovation Score in our analysis. This qualitative examination serves as an initial validation to check that some scores 昀椀t our intuitions.
To ensure the robustness of our measure beyond qualitative inspection, we compare our Innovation Score with a Novelty Score derived from a method proposed by Luan and K8im], [ which gauges the uniqueness of a movie’s genre combination relative to preceding 昀椀lms. This Novelty Score is genre-based: it rewards 昀椀lms introducing rare combinations of genres. We used this Novelty Score as a benchmark to evaluate the e昀ectiveness of our Innovation Score in capturing a movie’s deviation from genre conventions. As anticipated, we found a signi昀椀cant positive correlation between our Innovation Score and the genre-based Novelty Sco=re.0(4 , < .001 ), bolstering the external validity of our measure (see Figure 2.B.).
We conducted multiple two-sample t-tests comparing the aggregated Innovation Scores of movies within a speci昀椀c genre to those outside of it (with Bonferroni correction for multiple testing). This test allowed us to discern whether there was a signi昀椀cant di昀erence in the mean level of innovation between movies belonging to a genre and all the other ones (see Figure 2.C.). Genres that encompass formulaic narrative plots, such as Film Noir, Mystery, Crime, Sport, and Thriller, exhibit lower average Innovation Scores. This phenomenon arguably occurs because, to belong to these genres, a given movie needs to adhere to speci昀椀c narrative conventions. In contrast, genres like Adventure, Action, and History, characterized by non-speci昀椀c themes, allow for greater innovation because they accommodate a broader spectrum of narratives that can deviate from traditional storytelling structures. While it may seem counterintuitive at 昀椀rst, the high average Innovation Scores in War 昀椀lms can be attributed to their ability to draw from various historical events and kinds of warfare. Science Fiction’s high average Innovation Scores can be attributed to its futuristic focus and the inherent audience expectation for novelty in Science Fiction movies.
Awarded movies have higher average Innovation Scores. Building on the common intuition that award juries, who are cinema experts, tend to reward innovation in cinema, we sought to investigate the relationship between our Innovation Scores and movie awards. We created a binary variable indicating whether a movie had won at least one award, based on mentions in Wikipedia pages, and conducted a two-sample t-test to compare the innovation levels of awardwinning movies and those without awards. This analysis was extended to speci昀椀c awards like the Academy Awards, Sundance Awards, and the Palme d’Or (with Bonferroni correction for multiple testing). Our 昀椀ndings revealed that, in general, award-winning movies tend to have higher Innovation Scores (see Figure 2.D.). Notably, movies recognized at the Sundance Film Festival, known for its focus on innovative independent 昀椀lms, were associated with higher innovation levels. However, this correlation was not observed for other speci昀椀c awards, such as the Independent Spirit Awards. This discrepancy could underlie the idea that our Innovation Score captures a speci昀椀c type of narrative innovation, which may not align with the criteria used by all award bodies. Nevertheless, the overall trend supports the validity of our Innovation Score as a measure of innovation in cinema.
In our exploration of the temporal dynamics of innovation in cinema, we aggregated the Innovation Scores by year and conducted a series of regression analyses. We 昀椀tted three models: a linear model, a quadratic model, and a logarithmic model, each with the mean Innovation Score as the dependent variable and the year as the independent variable.
The linear model, which assumes a constant rate of change in innovation over time, accounted for approximately 18% of the variance in the data. The logarithmic model, which posits a decelerating rate of innovation, performed similarly to the linear model, explaining approximately 18% of the variance. However, the quadratic model, which allows for a changing rate of innovation, performed signi昀椀cantly better, explaining about 28% of the variance. To further compare these models, we calculated the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), both of which balance the goodness-of-昀椀t of a model with its complexity. Lower values of AIC and BIC indicate a better model. The quadratic model outperformed the other two models on both criteria, further supporting its superiority (Figure 3.A.).
The quadratic model suggests a non-linear relationship between time and innovation in cinema. The positive linear term in the model indicates an overall increase in innovation over the years, while the negative quadratic term (lower in magnitude) suggests a slowing down of this increase (Figure 3.B.). Speci昀椀cally, by observing the plot of the 昀椀tted quadratic model, we can infer that innovation in cinema experienced a surge throughout the 20th century. However, this rate of increase appears to have decelerated and reached a plateau in recent years, indicating a stabilization of innovation levels in the cinematic landscape.
In addition to our regression analyses, we wanted to explore to what extent the production of 昀椀lms over the course of history deviates from what it would have been if the 昀椀lms had appeared randomly over time. We used a Monte Carlo simulation approach. This simulation involved generating 1000 datasets, each with movies randomly shu昀툀ed across di昀erent years while keeping the number of movies per year constant. The results of this Monte Carlo simulation are striking. They reveal a decrease in the average Innovation Score during the initial years, followed by a relatively constant, lower level of innovation per year. The reason is straightforward: the more 昀椀lms already exist, the more di昀케cult it is to innovate on a purely random basis. This pattern contrasts sharply with the actual data, which showed an increase in innovation throughout the last century.
This suggests that if movies were randomly distributed in time, without consideration of what came before, they would produce movies that share similarities with previous eras purely by chance, leading to a 昀氀atter and lower average Innovation Score. In contrast, in the real dataset, they seem to actively strive to innovate and di昀erentiate their work from what has come before, leading to an overall increase in innovation over time. This observation, therefore, highlights the strong connection between movie production, creativity, and the in昀氀uence of past cinematic trends. Filmmakers draw from the past while attempting to break away from prevailing storylines, resulting in the observed increase in innovation in the real dataset as opposed to the simulated ones.
Our analysis of over 19,000 movies spanning more than a century has yielded fascinating insights into the trajectory of cinematic innovation. We observed a signi昀椀cant increase in innovation throughout the 20th century, underscoring the era’s reputation as a period of rapid creative evolution. Thus, contrary to the o昀琀en-voiced lament that cinema is losing its innovative edge, our study suggests that the level of innovativeness in cinema is not in decline. In fact, according to our model, the level of innovation today is as high as it was during the golden era of cinema in the 1950s. This implies that the use of formulaic plots is not more prevalent now than it was in the past.