Development of efficient image processing methods in recent years brought rising interest in application of quantitative methods to television and film works. As evidenced by papers presented during meetings of ADHO Special Interest Group Audio Visual Material in Digital Humanities, this concerns especially their visual layer. Hitherto-conducted studies range from cinemetrics (e.g. works of Mike Baxter and Yuri Tsivian) to applications of high-level computer vision methods to the study of moving images (e.g. the Distant Viewing TV project: Arnold and Tilton 2017) to whole systems dedicated to archiving and analyzing of videos (CLARIAH Media Suite). Unsurprisingly, the analysis of the audio layer is yet to be undertaken on a larger scale – state of the art solutions to problems such as speaker diarization are still far from useful, and, like computer vision, this type of analysis is greedy in terms of computational power. Comprehensive studies examining all multimodal aspects of films and TV series still belong to the future.

While the above-mentioned studies are of great importance, they align more with the term of “distant watching” or “distant viewing,” and in this paper I aim to draw attention to another aspect of audiovisual creations: to the textual layer of dialogue. Sarah Kozloff (2000) noted that film dialogue is often overlooked and considered of little importance and interest to both film scholars or linguists. She argued in favor of the usefulness of such studies, providing relevant examples and encouraging further ventures into linguistic aspects of audiovisual creations. Later years brought noteworthy projects on both film and television dialogue, most notably numerous works in corpus linguistics and discourse analysis, produced, among others, by Monika Bednarek, whose 2010 and 2018 books provide the most comprehensive review of the research that combines linguistics and television studies.

In the latter book, Bednarek noticed the rise of popularity of more formalized research on TV dialogue, including quantitative studies using corpora of transcribed dialogues, and pointed to the examples of TV series studied within various linguistic frameworks, including sociolinguistics (discussion in Richardson 2010), corpus linguistics (multiple examples for this and other linguistic branches in the extensive bibliography by Bednarek and Zago 2018), multimodal discourse analysis (Bateman and Schmidt 2011, Bateman and Wildfeuer 2014, and Wildfeuer 2014) and audiovisual translation (Hołobut et al. 2016, 2018a, 2018b, Bruti and Vignozzi 2016). TV series in their textual layer also become an object of research on narratives and themes, e.g. topic modeling-based plot explorations conducted by Benjamin M. Schmidt (2015) on a corpus of 80,000 film and television series scripts.

Significantly for this study, the research project conducted by Hołobut and her team is also one of the first to apply stylometry to the exploration of stylistic similarities in audiovisual data. In one of their first papers, Hołobut, Rybicki and Woźniak (2016) discovered that, in the case of television series, genre was likely to overshadow stylistic features specific to the author. In their further research, namely a stylometric analysis of Pride and Prejudice and its adaptations, Hołobut and Rybicki found that television series were “imitating Jane Austen’s verbal style more suggestively than the cinema productions” [Hołobut and Rybicki 2018, 8]. Another attempt at applying stylometric methods to television series dialogues was made by van Zyl and Botha (2016), whose study of The Big Bang Theory dialogue lists confirmed the existence of clear character idiolects in television discourse.

Due to the multimodal nature of audiovisual works, in contrast to typical practice in literary and linguistic studies, the concept of authorship is very difficult to attach to one person in television. In fact, the issue of authorship “has been rarely associated with television” at all, as authorship “was diffused among many individuals and may more properly have been claimed by a studio, company, or channel than by a person” [Hartley 2004, 401]. Collaborative storytelling and the number of people involved in the process of writing, rewriting and changes introduced during shooting and even in post-production, are often emphasized by both television creators and scholars.

While since the 1950s, film directors have been widely accepted as authors (or auteurs) of their works, and much attention is given to analyzing the works of such as directorial starts as Quentin Tarantino or David Lynch, the directors of television series have had a much smaller influence and typically only work on one up to a few episodes per season. Mittel also notes that while

the literary writer is ultimately celebrated for authorship by origination, assumed (however erroneously) to have written every word as an individual, a film's director is granted authorship by responsibility, ultimately making the necessary creative choices by supervising and guiding all of the film's many collaborators, even if much of the specific work was done by others.  [Mittell 2017, 36]

This perspective on authorship by responsibility is also often linked to television, however, in relation not to directors but to showrunners (e.g. Steiner 2015, Jensen 2017, Lavik 2015). A showrunner is an informal term used to describe a person who holds most creative power in the show, usually combining the duties of main scriptwriter and executive producer. The concept is used mostly in relation to American culture (e.g. David Chase, Joss Whedon, Shonda Rhimes, David Simon), but there are also examples of showrunners in the UK (e.g. Russell T Davies, Charlie Brooker) and Scandinavia (Hans Rosenfeldt). Detailed origin and characteristics of this job have so far been subject to few studies, with the most comprehensive analyses by Brett Martin (2013) and Erland Lavik (2015). In his book, Martin looks into the phenomena associated with the popularity of series authored by some of the earliest openly-appointed showrunners and, based on conversations with them and their collaborators, he details existing forms of this mode of work. With all the information he collected, Martin does not hesitate to refer to TV showrunners (such as David Chase) as authors, comparing the conditions of their work to serial Victorian writers, such as Charles Dickens, Anthony Trollope or George Eliot. He also points out that given their decision-making over “story direction to casting to the color of seemingly insignificant characters’ shirts” [Martin 2013, 8] they seem “all-knowing deit[ies].” Martin also observes that showrunners are typically writers (with few examples of showrunners holding only producer credits) and that they tend to emphasize their power over words and world-shaping. While typically there are many writers involved in the writing of a series, there are various ways their work can be organized: collaboration can include a “writers’ room,” where a showrunner meets with writers to pass/develop ideas together in detail before designating a team member to transform such detailed vision into actual script and again to discuss it in the room; in another manner, writers contribute episodes based on a general season narrative arch developed by a showrunner, who supervises execution and introduces corrections and changes to maintain consistency. In turn, in his book Lavik compares American, Dutch and Norwegian styles of showrunning and argues that outside of the American perspective detailed by Martin, showrunners in Europe give their writers much more authorial freedom and have themselves more creative than administrative responsibilities. Interestingly, Lavik also distinguishes between focused (one showrunner for the whole production of the show) or distributed (showrunners changing every (few) seasons) overarching authorship of showrunners, and actual authorship of writers.

The exact scope of influence of a showrunner is unclear and varies from series to series. A script is usually rewritten many times at various points before, during and after production, and, as noted by Jensen:

other writers work to support the vision founded by the showrunner, or creator. That, however, does not mean that these writers (...) do not make a difference in the final outcome of a series (they most certainly do), but they do so within a paradigm laid down by the creator(s) of the show.  [Jensen 2017, 37]

Jensen also notes that despite this complex authorship of television productions, the works of particular showrunners bear similarities across various series, both in style and in specific interests of the authors [Jensen 2017, 33]. He emphasizes interest in “seeing how the various works across an oeuvre depict a topic in different ways, maybe even in contradictory ways” [Jensen 2017, 34] – this is to be expected from an author able to create well-developed characters that differ in their ideas and nature as expressed in their dialogue lines, and can be used to evaluate character-building skills of the author. One method to do so may consist in contrastive analysis discussed later in this paper.

The concept of authorship of audiovisual works as adapted from literature poses even more problems, as discussed in detail by Steiner (2015) on the example of Game of Thrones and its “showrunner-auteur troika:” of author of the novels adapted into the series and series’ two showrunners, all participating in writing of the stories. Determining the scope of influence of any of them is a task posing a great challenge to quantitative and qualitative approaches alike. Again, I will argue that one of stylometric methods, sequential analysis, can facilitate this task, at least when it comes to the authorship of the textual layer.

The British television series Doctor Who is a phenomenon for a number of reasons, and a perfect subject for a study applying a variety of quantitative methods. Broadcast since 1963 (with a break between 1989 and 2005), it is one of the most popular science fiction series, critically acclaimed, with generally a mainstream status (at least in Britain), and with a cult following.

The basic assumptions of the show are quite unique:

Its main character, the Doctor, travels with his companions in the TARDIS (Time and Relative Dimension in Space), a ship capable of traveling through space and time that takes the exterior form of a 1930s British police booth, but is bigger on the inside. (...) the Doctor is not consistently portrayed by the same actor; periodically the Doctor “dies” and regenerates in a new humanlike form with a new personality [Edwards 2014, 375].

This form certainly contributes to the popularity of the show. Frequent changes allow to keep it fresh and likely to attract both long-term fans as well as people interested in only following one specific regeneration of the Doctor. Rebuilding the show around new incarnation of the main character enables viewers to start watching the show from various episodes and seasons without normal difficulty related to following TV series from a random point, that is having to figure out previous plot twists relevant for its understanding.

The show can be, and is, divided in numerous ways both by producers, fans, and scholars. The most basic division line is that between the “Classic” (1963-89) – and the “New,” sometimes spelled “Nu” (produced from 2005 onwards). The 1996 feature film is considered a part of the Classic Who.

Another division line is defined by the respective regenerations of the Doctor, thirteen as of 2018; this does not include the so-called Doctor 8.5, also known as the War Doctor, appearing in the 50th anniversary special. Since the Doctor is the main axis of the show and usually the one character who delivers most of the dialogue (in the New series, an average of 30-40% of all lines belong to the Doctor, i.e. twice as much as any of his companions), their[3] change strongly influences the overall shape of the show.

The 2005 revival introduced many innovations, primarily in narrative style and in the introduction of the position of a showrunner. This has been discussed as both a factor ensuring the success of the refreshed series and a reason behind a major change in narrative style [Hills 2010] [Hills 2013]. As of 2018, the show has recently adopted its third showrunner, Chris Chibnall, with the first, Russell T Davies, serving in the years 2005-2010, and the second, Steven Moffat, between 2010 and 2017. Each of them had an established successful series writing career at the moment of taking over and was well-known by the British public. The showrunners were generally seen as bringing their individual style and vision to the show, and the years of their work on Doctor Who were marked respectively as the Davies and the Moffat eras (see e.g. in Chapman 2014) – a quick search reveals that a “Chibnall era” is now also used in the Internet, though with as of yet small frequency.

This perspective of various strong influences offers an opportunity to examine not only the authorship of the episodes, consistency of the show and character stability, but also the relevance of particular divisions.

Stylometry is grounded in the observation that the distribution of certain features, most notably most frequent words, is a strong bearer of stylistic ‘signal’, which can be used to predict authorship, chronology, genre, or to lesser extent topical similarities. The earliest attempts to compare the use of words across texts date back to Lorenzo Valla in the 15th century and Augustus de Morgan and Wincenty Lutosławski in the 19th century, with the latter considered to be the first to compare top most frequent words, using this method to analyze works of Plato. The approach became more popular after Mosteller & Wallace’s famous study of The Federalist Papers (1964), and started to develop intensely with the availability of more powerful computers and Burrows’ study of Jane Austen’s works (1987), in which he introduced a more reliable method of comparing not just individual words but whole texts with his Delta measure (2002).

A stylometric study requires a few basic items: a corpus of texts which will be turned into a frequency table detailing use of all available words, a method for calculating a distance between their use across the texts (so called distance measure) and a classification algorithm. An optional, but convenient addition is a visualization method, allowing for easier interpretation of the final outcome. All analyses in this study were performed using stylo, an R package for computational text analysis [Eder et al. 2016], which includes algorithms for all these steps and offers a wide selection of types of distance measures and classification algorithms, not requiring more computational skills than preparing the corpus and making an informed choice between appropriate options.

One important step of preparing the experiment is choosing the right features to analyze. While Mosteller & Wallace opted for using most frequent words (MFWs); later studies experimented with the use of combinations (the so-called n-grams) of subsequent characters, subsequent words or even grammatical features. While character n-grams show nice performance for some types of texts, most notably historic and poetic texts [Kjell 1994] [Stamatatos 2009] [Kestemont 2014], n-grams of words typically perform worse [Eder 2011], as do grammatical features, e.g. POS tags [Górski et al. 2014] [Cafiero and Camps 2019], or lemmas. Given that the above studies seem to show that, in contemporary settings, most frequent words perform best, I decided to use them as features. Dismaying as such an approach may seem to the traditional humanities scholar, empirical evidence supporting it is overwhelming. The possibility of using such simple evidence for such large purposes rests upon the fact that words do not function as discrete entities. Since they gain their full meaning only through the different sorts of relationships they form with each other, they can be seen as markers of those relationships and, accordingly, of everything that those relationships entail [McKenna et al. 1999].

The first and mainly exploratory test was hierarchical clustering, one of the basic but most reliable algorithms of grouping similar texts, in its network variant of bootstrap consensus analysis. This variant is based on conducting a series of hierarchical clustering analyses with different counts of features and comparing their inter-agreement with a set threshold, so that only the texts classified as similar a set number of times are taken into account (e.g. for a threshold of 0.5, i.e. 50% and four iterations, two texts need to be classified as each other’s nearest neighbors in at least two iterations for the connection to be considered valid). It presents a more objective and reliable view of the relations in the dataset than regular cluster analysis, decreasing the risk of cherry-picking by not relying on just one best outcome, and can be visualized in a graphical form of a network, which facilitates interpretation of the results in the case of large datasets, like the one in this study.

Network data produced by stylo was used in Gephi to create visualizations and to conduct further analyses, e.g. community detection with the Louvain modularity algorithm [Blondel et al. 2008]. Community detection is a method facilitating identification of naturally occurring groups within networks, often difficult to observe in detail with human eye, allowing for understanding their large-scale structures. The method divides a network into groups of nodes sharing strong connections between them. The Louvain algorithm is commonly used in many disciplines and was found to be the most reliably performing community detection in stylometric studies [Ochab et al. 2019] as it is well adjusted to the analysis of networks that may contain small structures, which can be expected also in the case of TV dialogue data.

In the contrastive analysis of Doctor regenerations I used the algorithm of Craig's Zeta [Burrows 2007] [Craig and Kinney 2009], which compares which features are used significantly and consequently more often across works in each of the two examined sets.

For a more detailed study of authorial influence of showrunner supervising scriptwriters I used Rolling Delta algorithm [Eder et al. 2016], a method that looks for “signals” of different authors within a single collaborative text. It conducts sequential classification of “slices” of such a text and allows for a fairly precise identification of authorial takeovers within one work (e.g. the case of Harper Lee in Choiński et al. 2019).

Dialogues in television series differ from both literary texts to unscripted spoken language. Awareness of this is crucial for proper performance and interpretation of a quantitative analysis. Functions of dialogue in television writing related to anchorage of the narrative, building continuity and helping viewers remember characters’ names, locations etc. [Kozloff 2000, 33] make for a higher percentage of proper names both in the overall show and in particular episodes. As observed in the frequency tables of word usage in the corpus used in this study, at least in the case of Doctor Who, especially episodes introducing new characters tend to repeat names more often, and, as evidenced by the count of number of words uttered by companion characters in Doctor Who, introductory episodes are usually also the ones in which respective characters speak the most within the whole set of their appearances in the show.

This nature of TV language, combined with the thematic variety of episodes, calls for implementing some sort of limit to the influence of features that may score very high in individual texts (and this way also in the list of most frequent words) but be rather non-important for the general stylistic profile of the corpus. One such solution, commonly known as culling, has long been in use in stylometry, although as of yet there are no studies determining exact conditions of its use, thus enforcing special carefulness in its application. Culling is a procedure of removing such features from the list of analyzed features that only appear in a determined percentage of analyzed texts.

In this study I chose to first conduct a leave-one-out cross-validation classification test on the whole corpus (classifying each text using the rest of the dataset as the training model) and discovered that culling of 20% increased the performance of my chosen classifier (Support Vector Machine) performance for all examined numbers of MFWs (100-500). While such procedure might be useful, factors to be taken into consideration before adopting it to other studies are size of the corpus, thematic variation of the texts, and their length. Finally, one should remember that the use of culling heavily decreases the number of features that can be used in the study (e.g. in case of texts composed solely of lines spoken by the Doctor, only 395 shared words remain).

As for other settings, I explored possibilities of using various distance measures and found Cosine Delta [Evert et al. 2017], which was proven to be the most reliable in a number of recent studies, to produce most stable results across settings, with Burrows’s Classic Delta also providing fairly good results.

The first experiment was set to examine the direction of the development of the show and see whether a division into Classic and New Who, or into the showrunner eras, result in the change of the stylometric “signals.” For this test I used all the texts in the corpora and considered two scenarios: first, examining the whole texts; second, examining just the extracted lines of the Doctor part. For each of the scenarios I conducted two analyses: with full text, and with sampling (random sampling of four 1,000-word samples per text). The second solution was adopted as an additional safety measure due to my concerns about the uneven length of the texts: while New Who usually has 4,500-6,000 words per episode, some Classic Who stories go up to 20,000 words.

For this part of study I focused on the New Who exclusively. I started with a network analysis including only the whole texts from this time period, which was enough to form a number of observations. The first basic one is the high visibility of contributions of Steven Moffat. Most of his works cluster together, be it the ones written under supervision of Davies or during his own era, with a distinction of a few texts that cluster into other communities with a few other writers he supervised, which could be hypothesized as his influence over them (see Figure 8).

The situation seems quite different for the other two showrunners. Episodes written by Russell T Davies appear in all communities related to his tenure, mixing with other writers and showing strong narrative connections (in Figure 8, the group marked with light green connects mostly more dramatic two-part episodes and finales, whereas the brown one gathers more regular episodes from the middles of the seasons). Preliminary tests conducted for this study, not including the last two seasons, showed this even more sharply, with separate clusters for episodes opening and closing seasons: his writing and supervision seem to strictly control the development of the show, and while some influence of topics (e.g. episodes with traveling back in time) can be seen, the authorial control of a showrunner over writers, exhibited by strong ties between them, seems to be the main driving power behind the similarities shared by the texts.

Interestingly, as presented in Figure 9, the season created by Chris Chibnall clusters not only with his previous works on the show in the eras of his predecessors but also with the episodes written by “usual” scriptwriters earlier on, mostly from the Moffat era. This can be interpreted in a number of ways. Undoubtedly, Chibnall has a strong authorial presence, but the fact that other writers who largely contributed to the show and also built their own prominent careers such as Toby Whithouse (creator of Being Human) and Matthew Graham (co-creator of Life on Mars) cluster together with him could be interpreted as the community of writers who grew up in the Doctor Who writing room now finding their own voice (albeit somewhat similar when contrasted to Davies and Moffat).

The analysis of sole lines of the Doctor revealed a level of distinction between the Thirteenth and other New Who Doctors. The easiest hypothesis would be the impact of gender – the new female Doctor versus all men – or distinct impact of the showrunner. However, at least the Davies era episodes written by Chibnall show the latter’s stylistic “submission” rather than his own vision of the Doctor.

To verify this, I conducted contrastive analysis with Craig’s Zeta [Craig and Kinney 2009] to compare the Thirteenth Doctor against her four direct predecessors, considered together and separately. Additionally, I also ran another test comparing Doctors created by Davies and by Moffat. Craig’s Zeta allows for comparing two groups of texts to find words used consistently throughout texts in one of the groups, and consistently avoided by the other, and vice versa. Obviously, the most frequent words disappear, while content medium-frequency words come to the surface; this is also the reason why this version of quantitative analysis becomes (at least slightly) more similar to the traditional/qualitative approach.

As contrastive analysis was performed without culling, meaning that all words were included as they were, one immediate feature of lists of words characteristic to the particular regenerations is the high position of the proper names relating to companions – which is caused by frequent addressing of the given character or talking about them. This is true in most cases, but reveals various relations between the Doctor and their sidekicks: the names of all companions of the Ninth and Thirteenth Doctor are at the top of the list, but other Doctors show their preference towards a particular companion, e.g. Tenth to Rose over Martha and Donna, Eleventh to Amy over Rory and Clara.

On the first look, it would be easy to interpret the results as presenting a gendered portrayal of characters – the new female Doctor says “love,” “woman,” “together,” and “please” more often than other Doctors. However, a deeper analysis of preferred words of her and other Doctors shows that it is not that simple. “Love” is more characteristic for the Eleventh Doctor, known for his expressive and emotional attitude to the world, and while “please” is high on the Thirteenth’s most used list, its only seems high in comparison to Ninth and Twelfth Doctors, arguably more stereotypically male and harsher in behavior than the other two male Doctors. “Please” is also a word preferred by Moffat Doctors over Davies Doctors.

One characteristic feature of the language of the Thirteenth Doctor is a high ratio of elliptic adverbs such as “obviously,” “presumably,” “definitely,” “possibly,” as well as the adjective “sure.” Also, regenerations written by Moffat use them more often than Davies, which may influence the pattern of relative frequencies in the corpus enough to partly explain the fact that, in the network studies presented above, Louvain Modularity recognizes Chibnall as closer to Moffat than to Davies. When analyzed in context, this use of adverbs reveals an interesting conclusion for the first female Doctor: compared to her predecessors who were largely explaining the world to their companions, the Thirteenth Doctor is more likely to engage in a discussion that may lead to her questioning hypotheses of her team or confirming their intuition and reasoning.

This is where knowledge of the audiovisual material and the whole corpus is helpful in the interpretation of the results: of the five New Who Doctors who struggle with posttraumatic stress disorder [Gibbs 2013] caused by their regeneration, loss of home planet and subsequent companions, the Thirteenth Doctor is the one who healed best and looks forward to a major change in her attitude. In the past, through travels with various companions, the Doctor often relied on their help, but at the end of the day the higher (intellectual) position of the Doctor was always clear. The last season shows the Doctor who chooses to engage into a team relationship rather than create a strong bond with an individual or duo of characters, and to give them more space for their own agenda. She takes on three companions (and only one of them is a young woman, which was a given for her predecessors), and while at various points of her introductory season some of them seem to be more important than others, overall they have about the same number of lines per episode and (which was not measured quantitatively) seem to have about the same amount of screen time. Altogether the companions speak about 30% of all lines per episode, almost matching the part of the Doctor (~35%), and this is the highest ratio of participation of the companions in the New Who, the most talkative of whom rarely surpassed half of the lines of the Doctor during the Davies era, and 25% under Moffat.

The keywords obtained in the analysis also show the Thirteenth Doctor to be quite tech-savvy, even compared to the Tenth Doctor who was famous for his technobabble such as “This is my timey-wimey detector. It goes ding when there's stuff” (“Blink”, 3x10). She has “readings,” “signals,” etc. Finally, her list includes “kill,” “killing,” and “weapons,” and while “fighting” and “killing” are characteristic for Davies when contrasted with Moffat, the Thirteenth Doctor is the one whom these words distinguish the most. All Doctors are pacifist and opposing violence, and the show generally avoids showing death of people, even though the Doctor has to die to regenerate. Already the Moffat tenure showed some relaxation of this rule (there is a recurring fan joke saying he was never able to permanently kill off a character), frequently using death by sacrifice as a plot development tool, only to later cancel it by introduction of a workaround. The Thirteenth Doctor deals with death much more often – it concerns both major characters and random people, and it is not justified by sacrifice; in fact, it may seem almost accidental, as when a Tzim-Sha warrior kills an elderly security guard peacefully doing his watch in 11x01. The most recent Doctor and her “gang” fight with cruel creatures and “people killing” on a regular basis.

Combined, observed differences between the most recent female and the earlier male regenerations seem to be related more to the change of the mode from “lone genius” to “team leader” rather than building the character according to traditional gender roles. Considering social-emotional keywords, Thirteenth Doctor seems to be most similar to Tenth and Eleventh. Interestingly, contrastive analysis of Davies and Moffat Doctors also revealed that the former has a preference towards social sphere of life (words such as “mother,” “family,” but also “alone,” “human”), while the latter is distinguished by more verbs (e.g. “feel,” “shut,” “forget,” “talking”), and expressive interjections.