=Paper=
{{Paper
|id=Vol-3834/paper57
|storemode=property
|title=Admiration and Frustration: A Multidimensional Analysis of Fanfiction
|pdfUrl=https://ceur-ws.org/Vol-3834/paper57.pdf
|volume=Vol-3834
|authors=Mia Jacobsen,Ross Deans Kristensen-McLachlan
|dblpUrl=https://dblp.org/rec/conf/chr/JacobsenK24
}}
==Admiration and Frustration: A Multidimensional Analysis of Fanfiction==
https://ceur-ws.org/Vol-3834/paper57.pdf
Admiration and Frustration: A Multidimensional
Analysis of Fanfiction
Mia Jacobsen1,∗ , Ross Deans Kristensen-McLachlan1,2
1
Center for Humanities Computing, Aarhus University, Denmark
2
Department of Linguistics, Cognitive Science, and Semiotics, Aarhus University, Denmark
Abstract
Why do people write fanfiction? How, if at all, does fanfiction differ from the source material on which
it is based? In this paper, we use quantitative text analysis to address these questions by investigating
linguistic differences and similarities between fan-produced texts and their original sources. We ana-
lyze fanfiction based on Lord of the Rings, Harry Potter, and Percy Jackson and the Olympians. Working
with a corpus of around 250,000 texts containing both fanfiction and sources, we draw on Biber’s Mul-
tidimensional Analysis [4], scoring each text along six dimensions of functional variation. Our results
identify both global and community-based preferences in the form and function of fanfiction. Crucially,
fan-produced texts are found not to diverge from their source material in statistically meaningful ways,
suggesting that fans mimic the writing style of the original author. Nevertheless, fans as a whole prefer
stories with less focus on narrative and greater emphasis on character interactions than the source text.
Our analysis supports the notion proposed by qualitative studies that fanfiction is motivated both by
admiration for and frustration with the canon.
Keywords
quantitative text analysis, fanfiction, multidimensional analysis, style and genre, statistical modelling
1. Introduction
In 1992, Henry Jenkins published a seminal book in fan research, Textual Poachers. Contrary to
the received opinion that fan cultures comprise misfits, degenerates, and mindless consumers
to be ridiculed, fans are “active producers and manipulators of meaning” [16]. Drawing on the
concept of textual ‘poaching’ developed by Michel de Certeau [7], Jenkins argues that fans ac-
tively transform the consumption of a given media into a participatory culture. This behavior
is a product of both adoration and frustration with the media, motivating fans to explore and
articulate the ways in which the narrative was unsatisfying and could be ‘salvaged’. A cen-
tral aspect of this metaphorical salvation is the creation and dissemination of cultural products
within fan communities - also known as fandoms. These products cover a wide range of media,
such as videos, art, and playlists. For many people outside of these communities, though, the
prototypical example of these fan productions is likely to be written texts. These examples of
fan writing are both novel and derivative, allowing fans to create narratives of their favorite
CHR 2024: Computational Humanities Research Conference, December 4–6, 2024, Aarhus, Denmark
∗
Corresponding author.
£ miaj@cas.au.dk (M. Jacobsen); rdkm@cc.au.dk (R. D. Kristensen-McLachlan)
ȉ https://orcid.org/0009-0003-3720-3418 (M. Jacobsen); https://orcid.org/0000-0001-8714-1911
(R. D. Kristensen-McLachlan)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
93
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
media in opposition to the original creators’ intentions. These writings, also known as fanfic-
tion (or fanfic), were popularized within fan circles mainly through the creation of fan zines in
the 1950’s and 1960’s [15], but are now much more commonly posted to online fanfiction sites,
with notable examples being Fanfiction.net and Archive of Our Own (AO3) [10].
Fanfiction as a textual genre is commonly defined as stories involving characters and worlds
taken from a preexisting storyworld [24, 27, 2]. However, this definition neglects to mention
the influence the fan communities have on fanfiction as a medium, despite individual fictions
being rooted in specific fandoms. Indeed, fanfiction writers themselves report that the com-
munity is the main motivator for producing and disseminating these texts [1]. Moreover, the
norms within the communities mean that writers often receive supportive feedback on their
fanfiction stories [8]. This feedback is often responded to by later incorporation of the wishes
of commenters into the fanfiction [9, 5]. In this way, fanfiction and the fan community have
a reciprocal, even dialectical, relationship. The fans feed their wishes into the fanfiction, and
the fanfiction becomes the main medium for co-creation and distribution of the norms and
values within the community. As Busse also puts it in Framing Fanfiction: “fan fiction loses its
meaning if removed from its context. Fan fiction thus offers insight into the fan community –
its conversations, its tropes, and its members’ discussions and concerns.” [6]
Any proper consideration of fanfiction must therefore take into consideration both the lin-
guistic structure of the texts and the unique context provided by the process of co-production
within fan communities. In this paper, we hence set out to address the following research
question: what does the linguistic structure of fanfiction tell us about the motivations and con-
cerns of the fandoms that produce them? Is fanfiction as a genre monolithic or do different
communities have different preferences?
1.1. Related Works
It is clear that the structure of individual fandoms and the texts they produce are potentially
of great interest, insofar as they provide insight into the genesis of interpretive communities
[12]. Nevertheless, there is a relative scarcity of scholarly research into fandoms and fanfiction,
despite the sheer volume of data available online. Fanfiction research has traditionally been
developed from a qualitative and ethnographic perspective [2]. However, given the volume
of online text available through platforms such as AO3 and the prominence of these texts in
online spaces, there is an increasing interest in the computational analysis of fanfiction [29].
Computational studies of fanfiction can be split into two main groups: those interested in
the traits of popular fanfiction; and those interested in the character and gender dynamics of
fanfiction more generally. For example, previous studies have found both gender and char-
acter disparities in fanfiction texts, with fanfiction being more likely to deprioritize the main
characters in favor of the secondary characters and devote more attention to female characters
[18]. Another study have found fanfiction pertaining to Greek mythology to be more likely
to contain violence when the story is about a heterosexual couple compared to other couple
constellations [19].
Concerning the textual features of successful or popular stories, these fanfics are found to
have a simpler syntactic structure, a plainer writing style, but also a wider vocabulary [17, 20].
The features that pertain to direct speech are also more prevalent in popular fanfics compared
94
to other fanfics [17]. A different study comparing the emotional arcs and characters graphs of
fanfiction found that fans preferred fanfictions with emotional arcs that were dissimilar to the
source text’s emotional arc, indicating a preference for stories with a different turn of events
[26]. On the other hand, from the perspective of character networks, no clear global preference
can be found regarding similarity or dissimilarity to the networks found in the source texts [26].
1.2. Multidimensional Analysis
Something which is currently missing from these quantitative analyses is an explicit linking
between form and function. In other words, the distribution of individual linguistic features
does not strictly tell us anything meaningful about the structure of the texts as texts or about the
effect of these quantitative differences on readers. While this issue of interpretation is arguably
a more fundamental problem in quantitative text analysis, it is nonetheless true that specific
methods of analysis lend themselves more naturally to less speculative kinds of interpretation.
This is particularly true of fields such as corpus stylistics, where a range of inferential statistics
and null-hypothesis tests are integrated with stylistic analysis of authorial choices to explain
variation in texts and across corpora.
One such approach is Biber’s Multidimensional Analysis (MDA) [4]. MDA has been widely
adopted across multiple different textual registers and genres and across multiple different
languages. The core component of MDA involves analyzing the distribution of specific
grammatical-semantic linguistic features which are argued to be functionally motivated. These
features are grouped into metacategories, allowing us to describe the structure of texts along
a number of dimensions of variation. This gives us a way of comparing the linguistic structure
of texts and to explain what that variation means in terms of what those features do in a text.
To our knowledge, there are no studies that use MDA in the study of fanfiction. This study
thus takes a novel approach to the study of fanfiction, one focused on the usage of linguistic
features across text types to investigate the motivation and desires of fanfiction readers and
writers
2. The Corpus
We chose to work with three particular fandoms, each of which are based on literary works of
fantasy. Specifically, J.K. Rowling’s Harry Potter series (HP), Rick Riordan’s Percy Jackson and
the Olympians series (PJ), and J.R.R. Tolkiens’ Lord of the Rings trilogy (LOTR) were chosen.
These three groups constitute some of the biggest fandoms based on literary, fantasy novels.
Limiting the scope in this way arguably limits the generalizability of our study but it also allows
for a clearer comparison between fanfiction and source text, as well as a more controlled com-
parison across fandoms. We prioritized the robustness of the comparisons specifically because
the pre-existing literature in the field is so limited.
The corpus of fanfiction was collected from the online fanfiction site AO3. This particular
site is one of the largest repositories of fanfiction with over 13 million works and simultane-
ously functions as an online archive for fanfiction sites which no longer exist such as LiveJour-
nal [28].
95
Table 1
Corpus summary
Harry Potter Percy Jackson LOTR Total
Fanfiction 224,039 17,823 12,523 254,385
Source texts 202 69 84 355
On AO3, fanfics are split into fandoms through the use of tags. Authors add fandom tags
to their fanfics when posting them, and they are used to specify which fictional universes a
given story is connected to. A team of volunteers called tag wranglers make sure that tags are
appropriately aggregated, so that, for example, fanfics with misspelled tags are additionally
tagged with the correct one [23]. These tags were therefore used for retrieving the fanfiction
texts, specifically, the tags “Harry Potter – J. K. Rowling”, “Percy Jackson and the Olympians
– Rick Riordan”, and “Lord of the Rings – J. R. R. Tolkien”, since these tags pertain to the
original, literary installments of the source texts. Although the stories might also be tagged
with other fandoms such as “Harry Potter (movies)”, they are at least to some degree about the
original texts, denoted by the use of these tags. Additionally, only fanfics written in English and
fanfics which had no crossovers – meaning characters or worlds from other fandoms – were
included. Besides that, all maturity ratings, lengths, and completion statuses were included in
the scraping process.
We modified an already existing web scraper to collect data from AO3,1 with minor adjust-
ments adapted to fit the current study. The texts themselves were scraped over the course of
late 2023 through early 2024. The latest text included in this study was published January 3rd,
2024, and the earliest is stated as being published on January 1st, 1950 (although this is almost
surely due to errors in the archival process). Along with the texts, we also collected engage-
ment metadata from AO3, such as the number of ”hits” and ”kudos” - i.e. how many times the
story had been read and liked by the community. The fanfics were scraped in accordance with
AO3’s terms of service and stored in compliance with GDPR.
In line with the tags that were used to collect the fanfiction, we limited our data to only the
”core” texts in the original series. For LOTR this meant that only the three books were included:
The Fellowship of the Ring, The Two Towers, and The Return of the King. The Silmarillion and The
Hobbit were excluded even though they are also written by Tolkien and take place in the same
fictional universe, since the use of fandom tags had effectively excluded fanfics pertaining to
only The Silmarillion or only The Hobbit. For PJ, the study only includes the five original books:
The Lightning Thief, the Sea of Monsters, the Titan’s Curse, the Battle of the Labyrinth, and the
Last Olympian. Finally, HP included only the original seven books: The Philosopher’s Stone,
the Chamber of Secrets, the Prisoner of Azkaban, the Goblet of Fire, the Order of the Phoenix, the
Half-Blood Prince, and the Deathly Hallows.
Before feature extraction and modeling, some data cleaning measures were implemented,
these are detailed in Appendix A.1. A summary of the number of texts can be seen in Table 1.
1
https://github.com/radiolarian/AO3Scraper
96
Table 2
Summary of dimensions of variation established using MDA
Summary Short Description
Dimension 1 Involved versus In- Informational: Dense and careful information inte-
formational gration.
Involved: Verbal style with a focus on the here-and-
now.
Dimension 2 Narrative Concern Distinguishes between texts with a narrative focus
from others
Dimension 3 Context- Context-dependent: Receiver must use context to in-
(in)dependent fer what time and place is being referred to.
referents Context-independent: The referents in the text are
made explicit and thus not dependent on the con-
text
Dimension 4 Overt expression of The degree to which the sender’s opinion is overtly
persuasion expressed and/or overt attempts to persuade the re-
ceiver are made
Dimension 5 Abstract style Distinguishes between informational discourse
that is abstract and technical from informational
discourse that is not
Dimension 6 On-line information Indicates whether the information presentation is
elaboration made carefully or is more fragmented
3. Experiment
3.1. Feature Extraction
We extracted MDA features using the Multidimensional Analysis Tagger (MAT) [21]. The tag-
ger creates grammatically annotated version of the texts by using a combination of the Stanford
Tagger2 , as well as a series of rules for identifying the patterns of linguistic features described
in Biber’s original study [4]. This allows the user to input either a single text or a whole cor-
pus and receive both a tagged version of the text(s) and the different dimension scores for that
text. In other words, the MAT scores each of the texts in the new register on the already es-
tablished dimensions of variation within the English language. This means that the corpus of
texts provided by a user is described relative to other prominent registers in English.
Each text is thus given a score for each of the six dimensions of variation. The different
dimensions and their interpretations can be seen in Table 2. For a full description see [4, 21].
2
https://nlp.stanford.edu/software/tagger.html
97
Figure 1: Distribution of dimension scores across text types and fandoms.
3.2. Statistical Models
Based on the output from the MDA performed by the MAT, we conducted two statistical anal-
ysis. For both analyses we used linear mixed effects models. This type of statistical model
has the advantage of accounting for two of the study’s most prominent challenges: Imbalance
of sample sizes and repeated measures. Both of these challenges are addressed in the model
formulation. Through the specification of fixed and random effects, the hierarchical structure
of the data is built into the model. It thus accounts both for repeating authors and produces
robust results when faced with imbalanced datasets [11, 14].
The first model was set up to test whether there is a statistical difference between fanfiction
and source texts when it comes to the dimension scores extracted by the MAT. For each dimen-
sion, a linear mixed effects model was created which sought to predict the dimension scores
from the text type (fanfiction / source text) and the fan group (HP/LOTR/PJ). A random inter-
cept for author was included to control for the repetition of authors in the dataset. Because of
98
the great imbalance in the number of source texts compared to fanfic texts, we applied a set
of weights to down-weight the fanfics and up-weight the source texts. These are specified in
Appendix A.4. The model for this analysis is described as follows:
𝐷𝑖𝑚𝑒𝑛𝑠𝑖𝑜𝑛 ∼ 𝑡𝑒𝑥𝑡 𝑡𝑦𝑝𝑒 + 𝑓 𝑎𝑛𝑑𝑜𝑚 + (1|𝑎𝑢𝑡ℎ𝑜𝑟) (1)
The second analysis included only the fanfiction texts and investigated differences in how
readers respond to the different dimensions across fandoms. First, we defined an engagement
metric inspired by Pianzola et al [23]. The engagement metric is computed as the number of
kudos (i.e., likes) divided by the number of hits times 100 to get a percentage. In other words,
it can be thought of as the percentage of people who read a fanfic and also decided to give it a
like. Despite this metric not accounting for re-reads and updates,3 we deemed it to be suitably
representative as an engagement metric. We again created a linear mixed effects model for
each dimension. These models sought to predict the dimension score based on an interaction
between the engagement metric and the fan group, with the standardized word count and
publishing date included as control variables. Similarly to the first analysis, a random intercept
for author was included. Due to the large amount of data in each group, it was not deemed
necessary to include weights in this model. The formula for the second analysis was as follows:
𝐷𝑖𝑚𝑒𝑛𝑠𝑖𝑜𝑛 ∼ 𝑒𝑛𝑔𝑎𝑔𝑒𝑚𝑒𝑛𝑡 ∗ 𝑓 𝑎𝑛𝑑𝑜𝑚 + 𝑝𝑢𝑏𝑙𝑖𝑠ℎ𝑒𝑑 𝑑𝑎𝑡𝑒 + 𝑤𝑜𝑟𝑑 𝑐𝑜𝑢𝑛𝑡 + (1|𝑎𝑢𝑡ℎ𝑜𝑟) (2)
After fitting the models, the assumptions of linear mixed effects regression were checked,
and they were deemed to not be violated (see Appendices A.2 and A.3).
4. Results
The overall distribution of scores across each dimension can be seen in Figure 1. Most strikingly,
these distributions are closely aligned across all dimensions for all fanfics, indicating a marked
uniformity of linguistic style across each of the dimensions of variation. However, there are
subtle differences which can be teased apart through the mixed effects models described above.
The results from the two models are presented in Tables 3 and 4 respectively.
Table 3 shows that there is no significant difference between fanfics and source texts across
the different dimensions of variation. Instead, there are only differences in these scores be-
tween individual fandoms. With regards to informational/involved discourse (D1), LOTR has
the greatest degree of informational discourse and PJ has the greatest degree of involved dis-
course, while HP is in between the two groups. Additionally, LOTR has the greatest narrative
concern (D2) of the three groups as well as the greatest context-independence (D3), while PJ
has the least narrative concern (D2) and the greatest context-dependence (D3). Together, these
three dimension indicate that fanfiction authors might be mimicking the style of the original
author. The prevalence of abstract style (D5) also support this interpretation, since LOTR has
the greatest degree of abstract style, while PJ has the least abstract style among these three
groups. These differences in fandoms for D1, D3, and D5 are similarly found in the second
analysis (see Table 4).
3
Users can open a fanfic multiple times - when there are updates, for instance - but they can only like it once.
99
Table 3
Estimates for model (1) for each dimension of variation
Dimension 1 𝛽 SE t-value p-value
Text type -3.12 2.44 -1.28 0.20
LOTR -2.30 0.08 -28.78 <0.001*
PJ 1.47 0.064 23.16 <0.001*
Dimension 2 𝛽 SE t-value p-value
Text type -0.58 1.2 -0.5 0.62
LOTR 0.24 0.037 6.29 <0.001*
PJ -0.45 0.03 -15.25 <0.001*
Dimension 3 𝛽 SE t-value p-value
Text type -0.32 0.71 -0.44 0.66
LOTR 0.14 0.025 5.71 <0.001*
PJ -0.23 0.02 -11.3 <0.001*
Dimension 4 𝛽 SE t-value p-value
Text type -1.36 0.98 -1.38 0.17
LOTR -0.033 0.034 -0.96 0.34
PJ 0.25 0.028 9.22 <0.001*
Dimension 5 𝛽 SE t-value p-value
Text type -0.46 0.57 -0.81 0.42
LOTR 0.13 0.02 6.45 <0.001*
PJ -0.095 0.016 -5.76 <0.001*
Dimension 6 𝛽 SE t-value p-value
Text type -0.19 0.36 -0.53 0.6
LOTR 0.098 0.013 7.66 <0.001*
PJ -0.081 0.01 -7.89 <0.001*
In contrast to the other dimensions, the model for overt expression of persuasion (D4) finds
no difference in scores between HP and LOTR, but PJ texts generally have a greater degree of
overt persuasion. There is no immediately apparent reason for this pattern. One interpretation
concerns the domain of authorial point-of-view and modality [25]. While it is not possible to
explore this result in detail in this paper, it does suggest that future research may need to pay
greater attention to the rhetorical aspects of fanfiction than has previously been afforded to
the genre.
When it comes to on-line information elaboration (D6) another unexpected pattern of findings
emerges. We find that PJ is generally more careful and planned in its information presentation,
whereas LOTR seems to have a more fragmented information presentation. Since both abstract
style (D5) and informational/involved discourse (D1) indicate the complete opposite pattern, it is
unusual that PJ is significantly more planned than LOTR. The distributions of dimension scores
illustrated on Figure 1 also show some unusual patterns when it comes to on-line information
elaboration (D6). When looking at Figure 3, which shows the result of the second analysis, the
100
Table 4
Estimates for model (2) for each dimension of variation
Dimension 1 𝛽 SE t-value p-value
LOTR -2.59 0.16 -15.90 <0.001*
PJ 1.76 0.14 12.18 <0.001*
engagement 0.13 0.0057 22.06 <0.001*
engagement:LOTR 0.027 0.021 1.28 0.2
engagement:PJ -0.047 0.020 -2.42 <0.05*
Dimension 2 𝛽 SE t-value p-value
LOTR 0.15 0.076 1.94 0.052
PJ -0.66 0.068 -9.77 <0.001*
engagement -0.048 0.0027 -18.07 <0.001*
engagement:LOTR 0.032 0.0098 3.24 <0.01*
engagement:PJ 0.052 0.0092 5.67 <0.001*
Dimension 3 𝛽 SE t-value p-value
LOTR -1.43 0.051 2.81 <0.01*
PJ -0.26 0.046 -5.64 <0.001*
engagement -0.0037 0.0018 -2.05 <0.05*
engagement:LOTR -0.0012 0.0067 -0.18 0.86
engagement:PJ 0.0016 0.0063 -0.25 0.8
Dimension 4 𝛽 SE t-value p-value
LOTR 0.22 0.067 0.035 0.97
PJ 0.046 0.061 0.77 0.44
engagement 0.040 0.0024 16.71 <0.001*
engagement:LOTR -0.0056 0.0088 -0.64 0.52
engagement:PJ 0.038 0.0083 4.59 <0.001*
Dimension 5 𝛽 SE t-value p-value
LOTR 0.0024 0.041 5.29 <0.001*
PJ -0.10 0.038 -2.74 <0.01*
engagement -0.011 0.0015 -7.63 <0.001*
engagement:LOTR -0.0098 0.0054 -1.80 0.071
engagement:PJ -0.013 0.0051 -0.26 0.80
Dimension 6 𝛽 SE t-value p-value
LOTR 0.097 0.026 3.81 <0.001*
PJ -0.074 0.023 -3.21 <0.01*
engagement -0.0082 0.00092 -8.91 <0.001*
engagement:LOTR -0.0028 0.0033 -0.85 0.39
engagement:PJ -0.037 0.0031 -1.17 0.24
findings are again worth questioning. The y-axis for D6 is on a tiny range (from -1.1 to -1.4),
and the confidence intervals of the regression lines are quite wide. As D6 was dropped in later
iterations of the MDA [3], we would argue that omitting it within the context of this study
101
makes more sense than including it.
Table 4 describes the relationship between engagement and dimension scores across fan-
doms and shows a yet more nuanced picture. We found significant interaction effects for infor-
mational/involved discourse (D1), narrative concern (D2), and overt expression of persuasion (D4).
This means that for these three dimensions, there is a significant difference in how fans re-
spond to the dimension scores across groups. For context-(in)dependent referents (D3), abstract
style (D5), and on-line information elaboration (D6) there was no differences in how fans across
groups respond to these linguistic features.
For informational/involved discourse (D1), there is a positive relationship between engage-
ment and dimension scores, meaning that across fandoms there is a preference for fanfics that
are more involved. Although this effect persists in all fandoms, it is smaller for the PJ fan-
dom, while there is no significant difference in how fans from HP and LOTR respond to this
dimension.
For narrative concern (D2), the difference between HP and LOTR is no longer present when
purely looking at the main effect of fandom, but PJ has a significantly lower degree of narrative
concern compared to the other groups. For the engagement score there is a negative main
effect, meaning that less narrative concern is associated with greater engagement. This effect
is different across groups, with HP having a stronger effect than the other two groups.
The differences in expression of persuasion (D4) across fandoms also disappears in the sec-
ond analysis. Nevertheless, there is a significant, positive main effect of engagement on overt
persuasion across groups, and a significant, positive interaction effect of engagement for PJ
fanfics specifically. In other words, across fandoms there is a general preference for more
overt expression of persuasion, but this effect is especially strong for PJ.
For both context-(in)dependent referents (D3) and abstract style (D5) there is an association
between greater engagement and a lower dimension score, meaning that across the three fan-
dom groups, they all respond with a preference for texts that have a greater context-dependence
and less abstract information. It is worth noting that the effect for D3 is vanishingly small (𝛽 =
0.004), but that taken together with the other findings, it could again illustrate that fans prefer
texts that are more here-and-now oriented and less technical, thus exhibiting the same patterns
as seen throughout this analysis.
5. Discussion
What do these results mean for the study of fanfiction and the fandoms that produce them? We
consistently find no statistical difference in dimension scores between fanfiction and their orig-
inal source texts, only between fan groups. Some might argue that this is unsurprising since
both fanfiction and their source texts fall into the more general category of ’fiction’, which
limits the actual variation that might occur compared to other registers. This view neglects,
however, that different genres of fiction have been found to differ in MDA [4, 3]. Additionally,
despite finding no variation between the two text types, the analysis does pick up on arguably
more subtle differences in writing style between the original authors as exhibited by the differ-
ences across fandoms.
Despite the lack of a difference between the text types, fanfiction does seem to be more ho-
102
Figure 2: Mean dimension scores across text types and fandoms.
mogeneous than expected. Although previous research has found fans to be a heterogeneous
group, fanfiction exhibits a quite consistent linguistic style as compared to the source texts, as
seen in the distributions of dimension scores in Figure 1 and the regression analysis illustrated
in Figure 2. Despite the plot indicating a difference in means between the fanfics and source
texts, the source texts’ confidence intervals are much wider than the fanfics’, which probably
drives the statistically insignificant results. The analysis thus indicates that fanfiction in gen-
eral has a distinct style that is more consistent than the styles across source texts, but that this
style within communities is influenced by the writing style of the source text. Since we did not
find a statistical difference between fanfiction and source texts, further research is needed to
identify with certainty if this fanfiction style exists and how it integrates the style of its source
material.
Our second experiment foregrounds the way different communities of fans respond to the
prevalence of different textual features in their fanfics. By investigating the degree to which
fans appreciate the prevalence of different linguistic features in the texts, we find that fans write
fanfiction that generally looks similar across dimensions and mimic the style of the original
author, but they might not appreciate the same traits. All the fandoms studied prefer less
narrative concern, less abstract information, more conversational style, and discourse focused
103
Figure 3: Regression lines between engagement metric and dimension scores across fandoms.
on the here-and-now. These preferences illustrate that although fanfiction imitates the original
writing style of the source author, fans across groups still have a preference for stories that are
less imitative and more focused on the core aspects of fanfiction, namely character interaction
and emotional experiences [2, 24, 15].
When looking at the local, community-specific preferences as illustrated on Figure 3, LOTR
fans seem to prefer greater narrative concern (D2) as compared to the other two fandoms. This
preference is, simultaneously, a prominent aspect of the style of writing that sets LOTR apart
from the others. In contrast, PJ fans have no preference for narrative concern, but strongly
prefer greater overt expression of persuasion (D4) when compared to the other two fandoms.
Again, PJ texts in general were found to score higher on this dimension - something that put
those texts apart from the others. HP, the fandom in the middle of the spectrum, shows a strong
preference for less narrative concern. One interpretation could be that these three fandom
groups have distinct preferences for the degree of narrative in their fanfics. HP fans thus
write fanfics with a greater variety in narrative concern but have a strong tendency to prefer
works with less narrative. Meanwhile, LOTR fans are more inclined to prefer greater narrative
concern and thus also write works that fit this – a trait likely inherited from their source text.
Lastly, for PJ fans, the flat regression line visible on Figure 3 could be an indication that they do
not respond to narrative concern at all. In other words, they appreciate fanfics with a tendency
104
for either end of this dimension, but writers are more inclined to write with less narrative
concern, which could also be an inherited trait from their source text.
So, while global preferences do exist for involved discourse, non-abstract information, and
here-and-now focus, community-specific preferences seem to arise from the linguistic features
that set specific fandoms apart from other texts. This analysis supports the idea that fans write
fanfiction due to both admiration and frustration with the source material. The admiration
is seen in the imitation of the original author’s style, whereas the frustration is seen in the
preference for fanfics that break with the mold. This idea of admiration and frustration is
exactly the argument put forth by Jenkins [16] and later echoed through other studies [2, 24].
6. Limitations
One limitation concerning the data collection is the fact that all the analyzed fandoms are based
on fantasy novels. This was necessary due to the scope of the study where the simplicity of
the study design and robustness of findings were prioritized. This made the inclusion of other
types of source material not feasible. It does, however, impact the generalizability of these
findings, which might only be present in fanfiction based on fantasy novels.
Another limitation concerns the comparison between the source texts and the fanfiction
texts. All the included fandoms had prior to the scraping of the texts already been adapted
to TV and/or film. As it was near impossible to make sure that fanfiction based on only the
actual texts were included, all fanfiction stories which had the specified tags were included.
This limits the robustness of the comparison, as fans might not have read the source texts
before writing their fanfic, basing their fanfiction entirely on the adaptation. The idea of only
including fanfiction based on the actual source texts is, however, not as sensible as it might
seem. Fans are known to write fanfiction based on shows or stories they have not consumed
themselves [24]. Thus, knowing whether a piece of fanfiction is based on just its source text,
an adaptation, or even another fanfiction is impossible.
Finally, we operationalize engagement as a kudos/hits ratio meaning that fanfics are effec-
tively ‘punished’ for being revisited multiple times. Moreover, the metric is quite crude, es-
pecially if the goal is to deeply understand reader preferences when it comes to fanfiction.
However, as very little previous research has taken a computational approach to fanfiction
reader appreciation, this study is a step in the direction of a more nuanced understanding of
fanfiction as a phenomenon.
7. Conclusion
Despite the dynamic, dialectical co-production of fanfiction by specific fandoms, the resulting
texts are not significantly different from their source material, focusing instead on mimick-
ing the voice of the original author. While fans do mimic the voice of the original author,
across communities we find that fans prefer fanfiction stories that are more conversational
and here-and-now oriented, meaning a preference for fanfiction stories that are different than
their source material. Fans appear to be more interested in character interactions than in plot.
This trend is only to some degree, however, as the evidence also suggests that fans prefer the
105
linguistic features that set their source text apart from other groups of fans. Our analyses thus
support the conclusion that fans write fanfiction both due to admiration and frustration with
the source material – similar to what previous, qualitative studies have found.
Our study hence has a two-fold contribution. Firstly, it shows that these qualitative findings
are replicated when taking a quantitative approach, thereby providing additional support for
the reliability of these arguments. Secondly, our experiments illustrates how one can answer
the why of fanfiction writing by inferring it from an analysis of the how. Specifically, we find
that the imitation of the original author’s writing style could be an expression of admiration,
while the greater appreciation for fanfiction stories that are less imitative and perhaps more
generically fanfiction could be an expression of the tension, frustration, and resistance to the
source material.
Acknowledgments
Part of the computation done for this project was performed on the UCloud interactive HPC
system, which is managed by the eScience Center at the University of Southern Denmark.
References
[1] K. Bahoric and E. Swaggerty. “Fanfiction: Exploring in-and out-of-school literacy prac-
tices”. In: Colorado Reading Journal 26 (2015), pp. 25–31.
[2] J. L. Barnes. “Fanfiction as imaginary play: What fan-written stories can tell us about the
cognitive science of fiction”. In: Poetics 48 (2015), pp. 69–82. doi: https://doi.org/10.1016
/j.poetic.2014.12.004.
[3] D. Biber. “A typology of English texts”. In: Linguistics 27.1 (1989), pp. 3–43.
[4] D. Biber. Variation across speech and writing. Cambridge: Cambridge University Press,
1988.
[5] R. W. Black. “Language, culture, and identity in online fanfiction”. In: E-learning and
Digital Media 3.2 (2006), pp. 170–184. doi: 2https://doi.org/10.2304/elea.2006.3.2.170.
[6] K. Busse. Framing fan fiction: Literary and social practices in fan fiction communities. Iowa
City: University of Iowa Press, 2017.
[7] M. de Certeau. The Practice of Everyday Life. Berkeley: University of California Press,
1984.
[8] J. S. Curwood, A. M. Magnifico, and J. C. Lammers. “Writing in the wild: Writers’ moti-
vation in fan-based afÏnity spaces”. In: Journal of Adolescent & Adult Literacy 56.8 (2013),
pp. 677–685. doi: https://doi.org/10.1002/JAAL.192.
[9] S. Evans, K. Davis, A. Evans, J. A. Campbell, D. P. Randall, K. Yin, and C. Aragon. “More
than peer production: Fanfiction communities as sites of distributed mentoring”. In: Pro-
ceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social
Computing. Portland, Oregon, 2017, pp. 259–272.
106
[10] J. Fathallah. “Digital fanfic in negotiation: LiveJournal, Archive of Our Own, and the
affordances of read–write platforms”. In: Convergence 26.4 (2020), pp. 857–873. doi: http
s://doi.org/10.1177/1354856518806674.
[11] A. Field, J. Miles, and Z. Field. Discovering Statistics Using R. Sussex: Sage, 2012.
[12] S. Fish. Is There A Text in This Class. Cambridge: Harvard University Press, 1980.
[13] L. Hansen, L. R. Olsen, and K. Enevoldsen. “TextDescriptives: A Python package for cal-
culating a large variety of metrics from text”. In: arXiv preprint arXiv:2301.02057 (2023).
[14] S. Huber, E. Klein, K. Moeller, and K. Willmes. “Comparing a single case to a control
group–applying linear mixed effects models to repeated measures data”. In: Cortex 71
(2015), pp. 148–159.
[15] A. Jamison. Fic: Why fanfiction is taking over the world. Dallas, Texas: BenBella Books,
Inc., 2013.
[16] H. Jenkins. Textual poachers: Television fans and participatory culture. London & New
York: Routledge, 1992.
[17] A. Mattei, D. Brunato, and F. Dell’Orletta. “The Style of a Successful Story: a Com-
putational Study on the Fanfiction Genre”. In: Computational Linguistics CLiC-it 2020.
Bologna, 2020.
[18] S. Milli and D. Bamman. “Beyond canonical texts: A computational analysis of fanfic-
tion”. In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language
Processing. Austin, Texas, 2016, pp. 2048–2053.
[19] J. Neugarten and R. Smeets. “MythFic Metadata: Exploring Gendered Violence in Fanfic-
tion about Greek Mythology”. In: (2023).
[20] D. Nguyen, S. Zigmond, S. Glassco, B. Tran, and P. J. Giabbanelli. “Big data meets story-
telling: using machine learning to predict popular fanfiction”. In: Social Network Analysis
and Mining 14.1 (2024), p. 58.
[21] A. Nini. “The multi-dimensional analysis tagger”. In: Multi-dimensional analysis: Re-
search methods and current issues (2019), pp. 67–94.
[22] J. Ooms and D. Sites. “cld2: Google’s compact language detector 2”. In: Retrieved Feburary
7 (2018), p. 2019.
[23] F. Pianzola, A. Acerbi, and S. Rebora. “Cultural accumulation and improvement in online
fan fiction”. In: CEUR Workshop Proceedings. Vol. 2723. CEUR-WS. org. Amsterdam, 2020.
[24] S. Pugh. The democratic genre: Fan fiction in a literary context. Brigend: Seren, 2005.
[25] P. Simpson. Language, Ideology and Point of View. London & New York: Routledge, 1993.
[26] Z. Sourati Hassan Zadeh, N. Sabri, H. Chamani, and B. Bahrak. “Quantitative analysis of
fanfictions’ popularity”. In: Social Network Analysis and Mining 12.1 (2022), p. 42.
[27] B. Thomas. “What is fanfiction and why are people saying such nice things about it??”
In: Storyworlds: A Journal of Narrative Studies 3 (2011), pp. 1–24. doi: https://doi.org/10
.5250/storyworlds.3.2011.0001.
107
[28] C. Tosenberger. “Mature poets steal: children’s literature and the unpublishability of
fanfiction”. In: Children’s Literature Association Quarterly 39.1 (2014), pp. 4–27. doi: 1
0.1353/chq.2014.0010.
[29] K. Yin, C. Aragon, S. Evans, and K. Davis. “Where No One Has Gone Before: A Meta-
Dataset of the World’s Largest Fanfiction Repository”. In: Proceedings of the 2017 CHI
conference on human factors in computing systems. Denver, 2017, pp. 6106–6110.
108
A. Appendix
A.1. Data Cleaning
In the usual MDA analysis pipeline, the first 400 words of a text are extracted and tagged for
linguistic features. However, many posts made to AO3 are not written stories but are instead
picture collages, playlists, poems, audiostories, or other kinds of fan creations. Additionally,
fanfiction stories often have so-called author notes at the beginning and end of a chapter, which
are not part of the story itself. In an effort to minimize artefacts but keep representativeness,
fanfics with less than 600 words were excluded, and the MDA was run on the middle 500 words
of each fanfic.
After the snippets had been extracted, we utilized the textdescriptives package [13] to assess
the quality of the snippets. Specifically the quality pipeline component which calculates the
quality of the text based on both heuristic quality metrics and repetitious text metrics was
utilized. We used the default quality check settings and filtered out texts that did not pass this
quality check. The quality check was implemented to make sure that only fanfics which could
be described as written stories were included, and not posts such as lists or picture collages.
For the source texts, we also deviated from the typical approach in MDA. Since there, in
some sense, were only three source texts, it would amount to too little data if only 500 words
were extracted from each. Therefore, to have a comparable corpus of source texts that were
tagged, we extracted 500 words every 5000 words for each of the full source texts.
Before the statistical analysis, it was also deemed necessary to perform outlier removal. Not
only is there a great variability in document length, but hits, kudos, and dimension scores also
had quite extreme values. Outliers were defined as data points lying in the top and bottom
0.5% of a given distribution, in this case all of the dimension scores as well as hits, kudos, and
word count. This method was chosen as it minimized the data that was excluded without
compromising on the robustness of the statistical analysis compared to other methods (e.g.,
removing extreme outliers as defined by a boxplot). It also ensured that the cut-offs were as
explicit as possible. A total of 7,026 fanfics and 67 source text snippets were excluded based on
outlier removal. Additionally, as mentioned earlier, fanfics that did not pass the quality check
from the textdescriptives package were excluded, which constituted 38,307 fanfics.
Finally, it was also necessary to perform additional language detection on the text snippets,
as some texts were written in other languages than English. Using the cld2 package in R [22],
319 fanfics were excluded as they were detected as a language other than English. Finally,
since the second analysis included publishing date and word count as a control variables, the
12 fanfics that were set as published before January 1st, 2000 were removed, and the word count
was standardized to aid in model convergence.
A.2. Model (1) assumption checks
There are the following five assumptions of linear mixed effects modeling: Linear relationship
between predictor and output variable(s), no multicollinearity, independence of data points,
homoscedasticity, and multivariate normality.
The first three assumptions can be addressed jointly for all six models. Since the models
compare the means of categorical groups, one might have run an ANOVA instead of linear
109
Figure 4: Test of homoscedasticity
regression. However, a mixed effects model was necessary due to repeated authors. Since
ANOVA’s can be conceptualized as a specific case of the general linear model, the linear rela-
tionship is built into the model formulation. Independence of data points is achieved through
the random effects. Since all six models have the same predictors, we calculated the variance
inflation factor (VIF) for the Dimension 1 model. Both predictors had a VIF of 1, indicating no
multicollinearity.
For the fourth assumption, homoscedasticity, we have plotted the residuals against the fitted
values for each model below. Although the plots show some downward trend in the residuals,
due to the lack of a clear cone shape or other extreme heteroscedasticity the models were
deemed to be not violating this assumption.
The fifth assumption is also tested by plotting the residuals in a qq-plot to ensure they are
normally distributed. These are plotted below. All plots indicate that the residuals are gener-
ally normally distributed. As such, the assumption of multivariate normality was deemed not
violated.
110
Figure 5: QQ-plot of the residuals for each model
A.3. Model (2) assumption check
The first and third assumptions can be addressed at once for all models. Due to the sheer
number of data points, it would be infeasible to plot them on a point plot to assess whether there
is a linear relationship. Instead, we take special notice that all regression lines on figure 2 mostly
have narrow confidence intervals, meaning that the line is quite confident in its placement. we
would therefore argue that this assumption is not violated, however, it is worth exploring if
other relationships than linear might explain the data better.
As with the previous model, the assumption of independence of data points is accounted for
by the random effects. The VIF was calculated for the predictors in the model for Dimension
1. All VIF scores were below 5, meaning no violation of multicollinearity – except for the
111
Figure 6: Test of homoscedasticity and multivariate normality for each dimension score model
interaction terms. However, collinearity of interaction terms has been argued to be inevitable,4
and as such, we deem the assumption not violated.
Finally, the tests for homoscedasticity and multivariate normality for each dimension’s
model are presented above. As with the previous models, no model residuals indicate any
extreme deviations from homoscedasticity or normality, and as such the assumptions are not
violated.
A.4. Weights applied to model (1)
Since there are 1000 times the number for HP fanfics as there are HP sources, and around
150 times the amount of PJ and LOTR fanfics as PJ and LOTR source texts, we accounted for
this imbalance using the following weights. HP fanfics were weighted with 1/total number of
fanfics (= 0.00000393). PJ fanfics and LOTR fanfics were weighted with 1*15/total number of
fanfics (= 0.0000597), to account for the fact that there are around 15 times the number of HP
fanfics as other fanfics. All source texts were weighted with 1/total number of source texts (=
0.00282). Although different weights can quite substantially change the outcome of the models,
these weights were decided upon since they most accurately describe the different imbalances
in the data.
4
see: https://www.statalist.org/forums/forum/general-stata-discussion/general/1359532-is-multicollinearity-
between-interaction-terms-a-problem
112