Add 1 files

2510/2510.20926.md

@@ -0,0 +1,478 @@
+Title: FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction
+
+URL Source: https://arxiv.org/html/2510.20926
+
+Markdown Content:
+Natasha Johnson 1 Amanda Bertsch 1 Maria-Emil Deal 2 Emma Strubell 1
+
+1 Language Technologies Institute, Carnegie Mellon University 
+
+2 School of Library and Information Studies, University of Oklahoma 
+
+nmj@alumni.stanford.edu abertsch@cs.cmu.edu
+
+###### Abstract
+
+As language models become capable of processing increasingly long and complex texts, there has been growing interest in their application within computational literary studies. However, evaluating the usefulness of these models for such tasks remains challenging due to the cost of fine-grained annotation for long-form texts and the data contamination concerns inherent in using public-domain literature. Current embedding similarity datasets are not suitable for evaluating literary-domain tasks because of a focus on coarse-grained similarity and primarily on very short text. We assemble and release FicSim, a dataset of long-form, recently written fiction, including scores along 12 axes of similarity informed by author-produced metadata and validated by digital humanities scholars. We evaluate a suite of embedding models on this task, demonstrating a tendency across models to focus on surface-level features over semantic categories that would be useful for computational literary studies tasks. Throughout our data-collection process, we prioritize author agency and rely on continual, informed author consent.1 1 1 Dataset can be accessed at [https://huggingface.co/datasets/ficsim/ficsim](https://huggingface.co/datasets/ficsim/ficsim). Additional documentation can be found at [https://github.com/natashamariejohnson330/FicSim](https://github.com/natashamariejohnson330/FicSim)
+
+FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction
+
+Natasha Johnson 1 Amanda Bertsch 1 Maria-Emil Deal 2 Emma Strubell 1 1 Language Technologies Institute, Carnegie Mellon University 2 School of Library and Information Studies, University of Oklahoma nmj@alumni.stanford.edu abertsch@cs.cmu.edu
+
+1 Introduction
+--------------
+
+![Image 1: Refer to caption](https://arxiv.org/html/2510.20926v1/x1.png)
+
+Figure 1: Similarity between literary texts can be defined along many dimensions. Computational literary studies scholars generally seek to measure specific, semantic types of similarity such as similarity in tone, but embedding models over-index on more obvious features such as the author’s writing style.
+
+The last few years have been a time of immense progress in long-context processing in NLP. Several language models now support context lengths in excess of a million tokens. Embedding models for 32k context inputs abound. While challenges remain in long context modeling, successive approaches have made strong progress on benchmarks and have found applications downstream Kapoor et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib26)); Godbole et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib19)); Nie et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib39)).
+
+In parallel, there is interest in applying NLP methods within digital humanities (DH), particularly within computational literary studies. Many DH scholars have incorporated NLP methods such as topic modeling, sentiment analysis, and semantic textual similarity (STS) tasks such as clustering and measuring similarity into their research (Kleymann et al., [2022](https://arxiv.org/html/2510.20926v1#bib.bib28); Algee-Hewitt and Fredner, [2023](https://arxiv.org/html/2510.20926v1#bib.bib1)).
+
+Source Long Publicly available Not publicly analyzed Author labeled Multiple axes of similarity
+Recent novels✓✗✓✗✗
+Project Gutenberg✓✓✗✗✓
+MTEB STS tasks✗✓✓✗✗
+AO3 Fanfiction✓✓✓✓✓
+
+Table 1:  Comparison of data sources for semantic textual similarity tasks. Fanfiction represents our approach. 
+
+Yet evaluations of these approaches have been limited, particularly with regards to STS tasks. Many NLP models which have been (or have the potential to be) applied in DH research are not evaluated on literary applications. When they are, it is often on digital texts made accessible through public repositories such as Project Gutenberg (Kohlmeyer et al., [2021](https://arxiv.org/html/2510.20926v1#bib.bib30); Kryściński et al., [2019](https://arxiv.org/html/2510.20926v1#bib.bib31); Underwood et al., [2018](https://arxiv.org/html/2510.20926v1#bib.bib49); Bamman et al., [2024](https://arxiv.org/html/2510.20926v1#bib.bib8); Kočiský et al., [2018](https://arxiv.org/html/2510.20926v1#bib.bib29); Xu et al., [2022](https://arxiv.org/html/2510.20926v1#bib.bib55)). However, these texts—alongside related data and analyses from other commonly-scraped sites like Wikipedia—are included in the pretraining data of most models (Elazar et al., [2024](https://arxiv.org/html/2510.20926v1#bib.bib14)), which could cause direct or indirect contamination and skew evaluation results (Palavalli et al., [2024](https://arxiv.org/html/2510.20926v1#bib.bib40); Zhang et al., [2024b](https://arxiv.org/html/2510.20926v1#bib.bib58)).
+
+Furthermore, while existing literary datasets evaluate model suitability for tasks such as summarization Kryściński et al. ([2019](https://arxiv.org/html/2510.20926v1#bib.bib31)), question answering Kočiský et al. ([2018](https://arxiv.org/html/2510.20926v1#bib.bib29)); Xu et al. ([2022](https://arxiv.org/html/2510.20926v1#bib.bib55)), and identifying literary co-reference Bamman et al. ([2020](https://arxiv.org/html/2510.20926v1#bib.bib9)), DH scholars have expressed the need for embedding methods that capture semantic textual similarity within novel-length texts along several axes such as plot, tone, and setting Sobchuk and Šeļa ([2024](https://arxiv.org/html/2510.20926v1#bib.bib46)).
+
+In response to this gap, we present FicSim, an evaluation dataset for fine-grained semantic textual similarity (STS), constructed of long-form human-written narratives that are unlikely to appear in pretraining data, are accompanied by author-labeled metadata, and are included in this dataset with author consent. We describe our processes for selecting text not included in CommonCrawl scrapes, for obtaining and maintaining author consent for the use of their works, and for constructing pairwise similarity measurements corresponding to 12 different facets of fictional texts, in consultation with both literary scholars and authors (§[3](https://arxiv.org/html/2510.20926v1#S3 "3 Constructing FicSim ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")). We then describe the resulting dataset and its use (§[3.3](https://arxiv.org/html/2510.20926v1#S3.SS3 "3.3 FicSim Statistics ‣ 3 Constructing FicSim ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")). We evaluate existing approaches on this multi-faceted STS task (§[4](https://arxiv.org/html/2510.20926v1#S4 "4 Evaluation ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")). Models struggle to capture salient characteristics of long-form texts, only weakly disambiguate between categories, and over-index on surface features of the text (§[5](https://arxiv.org/html/2510.20926v1#S5 "5 Results: Can existing models perform fine-grained literary STS? ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")). We conclude by discussing the relevance of our results for both literary studies scholars and NLP researchers (§[7](https://arxiv.org/html/2510.20926v1#S7 "7 Conclusion ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")). We hope that FicSim enables more focus on narrowing the gap between models’ general capabilities and their applicability to literary domain tasks.
+
+2 Sourcing Data
+---------------
+
+### 2.1 Desiderata
+
+To effectively measure similarity in long-form fiction, we need a corpus of stories with several characteristics, summarized in Table[1](https://arxiv.org/html/2510.20926v1#S1.T1 "Table 1 ‣ 1 Introduction ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction"): The stories must be (1) long, coherent narratives (2) publicly available for evaluation (3) not publicly analyzed online, in order to prevent potential contamination (4) well-annotated, ideally by an expert, capturing multiple axes of similarity, beyond superficial similarities that are easy to detect but limited in the value for literary scholars.
+
+Recently published novels offer one compelling solution Karpinska et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib27)); Duarte et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib13)) but limit public release of the dataset.2 2 2 Copyright laws do permit physical copies of books to be purchased, scanned, and shared with certain provisions. However, digitizing physical texts can be a labor and cost intensive. Furthermore, the suitability of such novels for evaluation purposes decays over time, as summaries and analyses of the texts become increasingly likely to have been incorporated into model training. Public domain literature, such as the texts made available through Project Gutenberg, satisfies the length and public availability requirements, but is deeply present both in pretraining corpora and in public culture—with many analyses online of the themes, plot, and character arcs of each story, it is unclear whether a model identifying these characteristics is doing so through memorization Palavalli et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib40)).
+
+### 2.2 Our Approach
+
+In response, we turn towards fanfiction–fictional texts inspired by existing media, often sharing the characters or setting of the source work. Many fanfiction texts are complex long-form narratives, reflecting both the fandom subculture and major cultural movements of the time. In recent decades, fan studies has become an active subfield within literary, media, and cultural studies.
+
+Significantly for our purposes, popular fanfiction websites allow users to assign tags to their fanfiction (e.g. Figure[2](https://arxiv.org/html/2510.20926v1#S2.F2 "Figure 2 ‣ 2.2 Our Approach ‣ 2 Sourcing Data ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")) to help with fanfic discoverability and categorization. Tags range from purely descriptive to analytical to conversational and identify important elements of a fanfic from the author’s perspective. They are intended to help readers find the content that is interesting to them among millions of stories. We use such tags as the basis for our gold-standard similarity scores.
+
+![Image 2: Refer to caption](https://arxiv.org/html/2510.20926v1/bolder_tags.png)
+
+Figure 2: Example of fic tagging and metadata on AO3. Colored annotations mark data that inform similarity scores. 
+
+### 2.3 Story Selection
+
+#### Archive of Our Own.
+
+We source our fanfiction from [Archive of Our Own](https://archiveofourown.org/) (AO3), a digital repository hosting over 15M works. We selected this venue for two main reasons: First, AO3 has made significant efforts to discourage web scraping, including blocking Common Crawl scraping in 2022 (Works, [2023](https://arxiv.org/html/2510.20926v1#bib.bib54)) and implementing aggressive rate limiting policies.3 3 3 While these do not guarantee that no fanfics posted after this date are used in pretraining corpora, these restrictions, along with AO3’s lack of an official API, make these texts relatively unlikely to appear in pretraining web scrapes. Second, due to the site’s construction and norms, many AO3 stories feature highly detailed tagging (e.g. Figure[2](https://arxiv.org/html/2510.20926v1#S2.F2 "Figure 2 ‣ 2.2 Our Approach ‣ 2 Sourcing Data ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")), which we leverage to compute similarity along various axes.
+
+#### Requirements for stories.
+
+We only consider stories that are written in English, exceed 10,000 words, and feature detailed tagging. Furthermore, because AO3 restricted web crawling in December 2022, we only consider stories that were started and completed after this date. Focusing on texts over 10,000 words ensures that the stories in our dataset are similar in length to fictional texts commonly studied by literary scholars, including short stories, novelettes, novellas, and novels Gioia and Gwynn ([2006](https://arxiv.org/html/2510.20926v1#bib.bib18)); Harmon ([2003](https://arxiv.org/html/2510.20926v1#bib.bib22)). We do not screen for or remove explicit content, although the majority of the stories in the dataset are not marked as explicit.4 4 4 Explicit texts (which may be tagged as such because of violent or sexual content) are an active area of research in media studies (e.g. Popova ([2018](https://arxiv.org/html/2510.20926v1#bib.bib42)); Fazekas ([2022](https://arxiv.org/html/2510.20926v1#bib.bib15))). Additionally, when identifying stories that we hoped to include in our dataset, we looked for a variety of tropes, settings, and writing styles.
+
+### 2.4 Author Consent
+
+Many artistic communities have strong negative feelings towards the machine learning community. Writers have described the AI community as exhibiting “complete lack of respect” for artistic work (Gero et al., [2025](https://arxiv.org/html/2510.20926v1#bib.bib17)). One survey found that 96% of authors were against the use of their work for AI training without their explicit consent (The Authors Guild, [2023](https://arxiv.org/html/2510.20926v1#bib.bib48)). Fanfiction authors are no exception: fan communities have expressed dismay over learning that some fanfiction has been included in CommonCrawl datasets, and Archive of Our Own lawyers went before the U.S. Copyright Office to argue that fan authors should be able to opt-out their work from model pretraining corpora Archive of Our Own ([2023](https://arxiv.org/html/2510.20926v1#bib.bib4)).
+
+Therefore, although Archive of Our Own explicitly permits metadata collection done by academic researchers (Works, [2023](https://arxiv.org/html/2510.20926v1#bib.bib54)), we additionally received each author’s permission to include their work(s) in our dataset using an IRB-approved process. We reached out to each author individually through AO3 to explain our project and share guarantees about author privacy, story usage, and withdrawal of consent. Then, we invited authors to sign a consent form and provided mechanisms to ask questions or withdraw consent at any time. Furthermore, in consultation with several members of fan communities, we decided to commit to not using FicSim to train models; to access the dataset, we will require other researchers to agree to the same terms of usage. For more information on the consent process and the full text of the outreach documents, see Appendix[C](https://arxiv.org/html/2510.20926v1#A3 "Appendix C Author Consent Process ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction").
+
+### 2.5 Fanfiction versus Traditional Literature
+
+Over the past decade in particular, the line between fanfiction and published literature has been blurred in terms of both story content and writing style. Many stories originally written as fanfiction have been published as standalone books (with minimal edits, such as changing the characters’ names) and have seen commercial success Arzbaecher ([2023](https://arxiv.org/html/2510.20926v1#bib.bib5)). Many tropes created or popularized by fan communities have been adopted outside of fan spaces by authors and consumers of contemporary genre fiction Majnaric ([2024](https://arxiv.org/html/2510.20926v1#bib.bib36)); Jerasa and Boffone ([2021](https://arxiv.org/html/2510.20926v1#bib.bib25)). Furthermore, when performing stylistic comparison between fanfiction stories and their inspiration texts (e.g. Harry Potter novels), researchers have found that fanfiction stories do not stylistically diverge from their source material in statistically significant ways Jacobsen and Kristensen-McLachlan ([2024](https://arxiv.org/html/2510.20926v1#bib.bib24)). Thus, we believe fanfiction texts are suitable for evaluating model performance on literature, particularly contemporary fiction.
+
+### 2.6 Use of Fanfiction Tags
+
+We derive gold-standard annotations for document similarity from the tags each author has ascribed to their story, which may be either canonical 5 5 5 That is, standardized tags that are internally linked to synonymous and related tags by AO3’s tag-wrangling team. or user-authored.
+
+Our use of user-generated tags as the basis for our similarity calculations was informed by the work of Lyons and Tappeiner ([2008](https://arxiv.org/html/2510.20926v1#bib.bib35)), who explored using user-generated tags to expand access to library resources. They discuss how user-based tagging offers a form of natural language keyword categorization that can help capture particular narrative features which are not standard subject headings for narrative works. In this way, user-generated tags do a better job of helping guests find materials that relate to their interests. This is essentially the purpose that tags serve on AO3—its tagging system, handling of user-generated tags, and user norms have allowed readers to find the content that is interesting to them among millions of other stories.
+
+Though AO3 tags might seem like a limited source from which to construct similarity scores, within DH work, genre labels are often used as ground-truth labels for evaluating literary embedding, topic modeling, and clustering methods (Sobchuk and Šeļa, [2024](https://arxiv.org/html/2510.20926v1#bib.bib46); Schöch, [2017](https://arxiv.org/html/2510.20926v1#bib.bib44); Allison et al., [2012](https://arxiv.org/html/2510.20926v1#bib.bib2)). Within fan spaces, fanfiction tags serve many of the same purposes and are held in the same common understanding (Hellekson and Busse, [2014](https://arxiv.org/html/2510.20926v1#bib.bib23)) as genre labels. In fact, many tags created or popularized by fan communities such as “Enemies to Lovers” and “Slow burn” have notably been adopted outside of fan spaces by authors and consumers of contemporary genre fiction (Majnaric, [2024](https://arxiv.org/html/2510.20926v1#bib.bib36)). These tags are commonly used in advertisements, recommendations, and reviews of books (Jerasa and Boffone, [2021](https://arxiv.org/html/2510.20926v1#bib.bib25)). Goodreads has even added "Enemies to Lovers" as a book genre on their site ([goodreads,](https://arxiv.org/html/2510.20926v1#bib.bib20)).
+
+Thus, our process calculates story similarity according to a well-established framework of what is important in a story, as developed and refined by fanfiction authors and readers, and as adopted by many traditional authors, publishers, and readers.
+
+3 Constructing FicSim
+---------------------
+
+### 3.1 Tag Categorization
+
+We place tags into 12 different categories, corresponding with various types of similarity we hope to measure. Some categories, such as “Plot,” “Theme,” and “Time” capture general qualities of fictional literature and align with projects that discuss narrative similarity on the basis of actions, subjects, themes, and temporal setting (Algee-Hewitt and Fredner, [2023](https://arxiv.org/html/2510.20926v1#bib.bib1); Kleymann et al., [2022](https://arxiv.org/html/2510.20926v1#bib.bib28); Sobchuk and Šeļa, [2024](https://arxiv.org/html/2510.20926v1#bib.bib46); Piper and Bagga, [2022](https://arxiv.org/html/2510.20926v1#bib.bib41)). Others, such as “Fanfiction Tone and Content Tags” capture fanfiction-specific qualities and serve to identify whether models can identify similarity based on genre-specific conventions. To determine the categories, we relied on the aforementioned DH scholarship, as well as the annotators’ analysis of the tag set. When applicable, we place tags into multiple categories. Table[2](https://arxiv.org/html/2510.20926v1#S3.T2 "Table 2 ‣ 3.1 Tag Categorization ‣ 3 Constructing FicSim ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction") describes the categories with tag examples from each.
+
+Category Description Example Tags
+Plot Narrative actions and concrete subjects; describes what happens in a story and what is in a story.Letters Blood and Torture Artificial Intelligence
+Character States The emotions, attributes, roles, and physical characteristics of characters in the text.Trans Woman Character Is Bad at Feelings Dissociative Identity Disorder
+Relationship Dynamics Key characteristics of both platonic and romantic relationships in the text.Established Relationship Possibly slowest ever burn F/F
+Theme Abstract ideas explored throughout the story.Racism Found Family What Is The Impact Of A Mother
+Time Temporal setting.Alternate Universe - 19th Century Modern Retelling Post-Apocalypse
+Style Features of the writing style or narrative technique.POV Third Person Omniscient Epistolary Dialogue Heavy
+Fanfiction Tone and Content Tags Fan-community language for the type of story; often relates to both tone and plot (e.g. “fluff” generally involves lighthearted and domestic scenes).Angst Fluff and Hurt/Comfort Tooth-Rotting Fluff
+Fandom-specific Tags that involve settings, characters, or events from a canon text, and thus reveal information about the fandom the story belongs to.Phantom of the Opera AU Yule Ball (Harry Potter) Capitano/Mavuika
+Overall(Fandom-Agnostic)An aggregate grouping of tag categories 1-7; captures similarity between all tags that do not reveal the fandom or author identity.[any of the above tags, excluding fandom-specific]
+Overall(Fandom-Specific)An aggregate grouping of tag categories 1-8; captures similarity between all tags[any of the above tags]
+Fandom Captures whether two texts are inspired by the same piece(s) of media (e.g. books, movies, television shows).Genshin Impact Grey’s Anatomy Star Wars
+Author Captures whether two texts are written by the same fan author.[author IDs]
+
+Table 2:  Types of similarity and example tags that align with each category. Some of the above tags were placed into multiple categories within our dataset, but they nonetheless serve as useful references for the particular category they represent above. 
+
+#### Tag cleaning.
+
+In order to increase tag interpretability and allow for clearer tag comparison, we clean and standardize the tags prior to constructing similarity scores. The cleaning process involves removing unnecessary punctuation, standardizing capitalization, and correcting obvious spelling errors. Our tag standardization involves replacing tags with their canonical counterparts and rephrasing or restructuring tags to increase their semantic interpretability.
+
+#### Fandom-agnostic tags.
+
+In order to allow our dataset to evaluate model suitability for applications outside of fan studies, we want all tag categories except Fandom, Fandom Specific, and Overall (fandom-specific) to contain only fandom-agnostic terms which could be used to compare story similarity across fandoms. For the nearly 50% of the tags that contained fandom-specific references, we create a fandom-agnostic version of the tag by removing or replacing fandom-specific references (e.g. “Protective Cristina Yang” becomes “protective character”). We then place this fandom-agnostic tag into the appropriate fandom-agnostic category(/ies) while also placing the original tag into the Fandom and Fandom Specific category.
+
+#### Annotation process.
+
+Our tag categorization and rewriting was performed by two authors who are experts in fan studies, one with a background in computational literary studies and the other with a background in library and information science. Following the qualitative methods tradition of interpretative analytical process Lincoln et al. ([2011](https://arxiv.org/html/2510.20926v1#bib.bib34)); Strauss and Corbin ([1990](https://arxiv.org/html/2510.20926v1#bib.bib47)), the annotating authors arrived at annotations through a series of consensus-building discussions. In a few instances where tags contained fandom-specific terminology for a particular fandom that neither annotator had specific expertise in, the annotators reached out to fanfiction authors within that fandom to confirm their interpretation of the tag prior to categorization. In these cases, the authors were acquaintances that each annotator knew through their own engagement with fan communities.
+
+### 3.2 Similarity Score Calculation
+
+After tags are cleaned and categorized, we calculate category-specific similarity scores between stories. We embed each tag using Gemini Embedding Lee et al. ([2025](https://arxiv.org/html/2510.20926v1#bib.bib32)) . Then, given category-specific tag lists from two fanfics, we calculate the similarity between tag lists A A and B B as the average of the pairwise tag cosine similarities:
+
+sim​(A,B)=avg⁡sim a∈A,b∈B​(a,b)\text{sim}(A,B)=\operatorname{avg}\text{sim}_{a\in A,b\in B}(a,b)
+
+#### Gemini embedding.
+
+We selected Gemini Embedding as our embedding model for the following reasons: (1) It has the overall highest score on the MTEB leaderboard Muennighoff et al. ([2023](https://arxiv.org/html/2510.20926v1#bib.bib38)); (2) Among the best-performing models on the MTEB English-language STS tasks 6 6 6 As of April 2025., Gemini Embedding is the highest ranking model that is not from the same model family (or built upon the same model family) as the models we evaluate below. Furthermore, because Gemini Embedding has an input limit of 2048 tokens, selecting it as our tag embedding model does not then preclude us from including it in our evaluation section.
+
+#### Gold score validation.
+
+We validate our tag handling and similarity score construction in two steps. Prior to tag cleaning, one author annotated a set 330 story triplets (thirty 3-way story comparisons in each of the 11 non-author categories), identifying whether the first story in a triplet was more similar to the second or third story from that triplet in a given category.7 7 7 Story triplets were drawn from a randomly-generated set of 100 story triplets in each category. The author identified the first 30 non-ambiguous comparisons in each category, skipping triplets where the author thought an argument could be made for either ranking. Tag cleaning, standardization, and embedding processes were then adjusted to align the resulting gold similarity scores with the author’s annotations. Then, to further validate the scores, two other authors each annotated identical sets of 220 story triplets (20 stories x 11 categories) on the same task. Since the two annotators were not both experts in fan studies or literature, they were allowed to skip comparisons for which they did not think they could identify the more similar story. Of the 158 triplets that neither annotator skipped we measured annotator agreement of 82% (Cohen’s κ\kappa = 0.65). Our gold truth labels were then evaluated against the 129 triplets for which both authors provided the same rating, demonstrating 80% alignment with the annotations.
+
+### 3.3 FicSim Statistics
+
+Our final dataset includes 90 stories 8 8 8 Literary datasets with expert annotations are often comparable in magnitude Sims et al. ([2019](https://arxiv.org/html/2510.20926v1#bib.bib45)); Bamman et al. ([2020](https://arxiv.org/html/2510.20926v1#bib.bib9)). and gold similarity scores along 12 axes, for a total of 33,790 pairwise comparisons. The stories range from 10,001 to 488,772 words and span 46 fan communities (fandoms). The cleaned tagset has 9448 total tags (2133 unique) across 12 categories. For more detailed dataset statistics and license information, see Appendices[A](https://arxiv.org/html/2510.20926v1#A1 "Appendix A Additional dataset documentation ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction") and [D](https://arxiv.org/html/2510.20926v1#A4 "Appendix D License ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction").
+
+### 3.4 Similarity scores
+
+For every story pair, FicSim presents up to 12 similarity scores, described in detail in Table[2](https://arxiv.org/html/2510.20926v1#S3.T2 "Table 2 ‣ 3.1 Tag Categorization ‣ 3 Constructing FicSim ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction"). We divide these scores into three evaluation groups:
+
+#### Fine-grained similarity:
+
+Plot, character states, relationship dynamics, theme, time, style, fanfiction tone & content tags.
+
+#### Broader notions of similarity:
+
+Overall (fandom-agnostic and fandom-specific). The overall (fandom-specific) category captures a mix of both superficial and more integrated narrative elements, while the overall (fandom-agnostic) category primarily targets the latter.
+
+#### Superficial similarity:
+
+Fandom-specific tags, fandom, and author. Because these are always provided (even when a story is not otherwise tagged well) and generally obvious to deduce from the text, we do not consider these axes of similarity to evaluate embedding quality, only to measure how embeddings capture these in contrast to the more fine-grained features above.
+
+4 Evaluation
+------------
+
+For each model, we compute cosine similarity between embeddings for all story pairs. Following Muennighoff et al. ([2023](https://arxiv.org/html/2510.20926v1#bib.bib38)), we measure Spearman’s ρ\rho between the model-induced ranking of story similarities and our tagset-derived gold ranking. We report ρ\rho out of 100 instead of 1.0 for readability, and highlight values that are significant (p<0.05)(p<0.05).
+
+Model Plot Char-acter States Rela-tionship Dynamics Theme Time Style Fanfiction Tone &Content All(Fandom Agnostic)All(Fandom-Specific)Fandom Specific Fandom Author
+Linq-Embed\cellcolor blue 18.65 8.63 5.52 0.14 28.85 3.33 9.95 16.59 34.61 34.99 30.47 40.91
++ SW 10.17 5.69 9.12 6.87 21.51 7.03 6.09 7.37 15.17 18.02 29.25 40.73
+GTE-Qwen2 15.84 4.36 9.62 13.92 31.23 13.43 20.87 24.38 40.50 45.48 45.87 40.82
++ SW 13.00-1.01 10.74 7.52 24.84 12.96 1.74 11.02 20.49 20.65 26.73 40.73
+SFR-Embed 17.06 0.98 5.83 4.96 24.34-5.06 8.45 18.07 40.65 36.62 34.35 38.76
++ SW 10.42 7.14 10.73 7.98 22.66 9.28 8.05 10.47 17.88 19.29 31.49 40.64
+GTE-ModernBERT 13.49 11.87 15.39 0.17-0.30 8.20 13.28 16.77 42.70 48.50 50.08 36.33
++ SW 18.07 4.68 23.82 6.71 40.24 16.43 5.96 22.13 41.37 42.03 46.57 37.60
+m2-BERT-32k 5.95 8.06 17.08 2.66-0.44-1.66 14.30 9.99 14.11 13.49 14.40 17.47
++ SW 8.96 7.77 13.54 6.35 0.90 5.00 19.65 12.81 20.68 20.02 22.87 20.18
+Voyage-3-large 9.87 4.60 9.93 10.81 27.62 22.42 13.71 20.01 43.66 42.21 31.72 37.92
++ SW 10.80 4.25 8.29 11.01 22.13 19.32 9.22 19.81 43.15 42.06 35.99 37.72
+Claude+SW-1.19-0.74 5.94 11.64 27.54 17.05-0.43 8.99 39.27 49.41 44.95 40.19
+
+Table 3:  Spearman correlation of embedding cosine similarity to our tagset similarity measures, for several representative open-source and API-based embedding models. All models struggle at category similarity and overindex on authorial style. +SW denotes the use of a sliding window; Statistical significance highlighted. 
+
+#### Models.
+
+We select the 3 best-performing open-weights models on the Hugging Face Wolf et al. ([2020](https://arxiv.org/html/2510.20926v1#bib.bib53)) MTEB leaderboard with 32k context lengths: Linq-Embed-Mistral Choi et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib12)), GTE-Qwen2-7B-instruct Li et al. ([2023](https://arxiv.org/html/2510.20926v1#bib.bib33)), and SFR-Embed-Mistral Meng et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib37)). These are all 7B models; however, computational literary studies scholars often have access to limited computational resources. Thus, we consider a much smaller model– GTE-ModernBERT-base Zhang et al. ([2024a](https://arxiv.org/html/2510.20926v1#bib.bib57))– and two API-based solutions: m2-BERT-80M-32k-retrieval Fu et al. ([2023](https://arxiv.org/html/2510.20926v1#bib.bib16)) through the TogetherAI API and using Voyage-3-large Voyage AI ([2025](https://arxiv.org/html/2510.20926v1#bib.bib50)) through the Voyage.ai API. Finally, we consider whether large language models could perform this task. We use Claude-3.7-Sonnet Anthropic ([2025](https://arxiv.org/html/2510.20926v1#bib.bib3)) to summarize each story, then embed the much shorter summary documents with Voyage.9 9 9 Anthropic does not have its own embedding model; we use Voyage on the Claude outputs because this is the embedding model Anthropic recommends in [their documentation](https://docs.anthropic.com/en/docs/build-with-claude/embeddings). We follow each model’s default strategy for constructing embeddings; for additional details on models and pooling methods, see Appendix[B](https://arxiv.org/html/2510.20926v1#A2 "Appendix B Additional documentation of models ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction").
+
+#### Prompt.
+
+Linq-Embed, SFR-Embedding, and both GTE models support providing an instruction in a special format at the beginning of a text to be embedded. We experiment with using this to produce category-specific embeddings, by providing instructions to focus on each type of similarity in turn. We provide the prompt (see Appendix[B](https://arxiv.org/html/2510.20926v1#A2 "Appendix B Additional documentation of models ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")) to each embedding model at the start of the document.10 10 10 For the other models that do not support a custom instruction, we simply prepend the prompt without special formatting. For Claude, we use a modified form of category-specific prompt for summarizing and omit the prompt for the embedding stage.
+
+#### Sliding window.
+
+There are stories in FicSim that exceed the context length of every model evaluated. When stories exceed the maximum context length, we consider two options. As a baseline, we naively truncate each story to the maximum length. We then consider a sliding-window approach where we chunk the text to windows of the maximum context length and then pool embeddings across windows Wang et al. ([2019](https://arxiv.org/html/2510.20926v1#bib.bib51)). For the sliding window approach, we overlap windows by 2048 characters (depending on the model and story, approximately 500 words).
+
+5 Results: Can existing models perform fine-grained literary STS?
+-----------------------------------------------------------------
+
+We present results using a single general-purpose embedding from each model in Table[3](https://arxiv.org/html/2510.20926v1#S4.T3 "Table 3 ‣ 4 Evaluation ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction") and results on category-specific embeddings for each model in Table[5](https://arxiv.org/html/2510.20926v1#S5.T5 "Table 5 ‣ Models overindex on surface features. ‣ 5 Results: Can existing models perform fine-grained literary STS? ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction"). All models struggle to perform fine-grained STS; unexpectedly, the larger models are not consistently better than the small embedding models. We discuss overall trends below.
+
+#### Truncation and sliding windows show marginal differences.
+
+Surprisingly, naively truncating to the first segment of the story and taking the mean of sliding window embeddings are similarly successful across the board. Though the mean Spearman correlation for truncation is slightly higher than the mean sliding window correlation (11.70 vs 11.17), when comparing individual correlations for a given model and category, truncation beats sliding window 51% of the time. 11 11 11 Note that we do not compare between sliding window and truncation for the Claude results because only four stories exceed Claude’s 200k context window, so there is not enough data to make a meaningful comparison between the methods. Despite allowing a model to consider more of the text, sliding window embeddings may fail to increase model performance because models are not generally trained for this approach; because the pooling strategy needs to be adjusted when pooling across more data; or simply because some of the features that capture story similarity can be extracted from the start of the text alone. That being said, while sliding windows do not improve correlation scores in any of the fandom-agnostic categories, they do decrease correlations in the confounding categories (mean ρ\rho of 30.47 as opposed to 35.69), suggesting that their embeddings weigh surface-level features less heavily.
+
+Gold score category Overall(Fandom Specific)Fandom Specific Fandom Author
+Plot 31.85 7.56 8.43 12.14
+Character States 4.52 0.09 15.10 13.55
+Relationship Dynamics 36.11 10.58 17.25 17.53
+Theme 10.30 1.95 3.98 5.10
+Time 12.53 14.36 34.65 21.28
+Style 24.18 8.90 11.21 13.05
+Fanfic Tone & Content 00.00-3.61-1.05 5.19
+Overall (Fandom-Agnostic)48.01 14.04 18.90 19.20
+
+Table 4:  Gold category-specific scores range in correlation with the overall (fandom-specific), fandom-specific, fandom, and author categories. The comparatively low scores in the latter three categories indicate that the gold scores weigh surface-level features far less heavily than our embedding models. Statistical significance highlighted
+
+#### Models overindex on surface features.
+
+Across all models, Spearman’s ρ\rho is higher for the four “confounder” categories than any of the fine-grained similarity categories. Notably, the author category (which is computed solely from exact match of author IDs) has the highest score more frequently than any other category. This indicates that embeddings are much more sensitive to author-specific stylistic factors than to the fine-grained semantic factors captured in the remaining categories.
+
+Some sensitivity to author and fandom is expected—some authors will focus on different types of stories, and some conventions or styles will be more common in one fandom than another. The overall (fandom-specific) category takes this into account by considering similarity based on both fandom-specific and non-fandom tags. However, 77% of the models in Table [3](https://arxiv.org/html/2510.20926v1#S4.T3 "Table 3 ‣ 4 Evaluation ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction") score higher in the fandom-specific, fandom, or author category than in the overall (fandom-specific). This indicates that they are not only capturing the fingerprints of certain authors and fandoms, but that they are furthermore failing to capture other narrative elements. In contrast, Table[4](https://arxiv.org/html/2510.20926v1#S5.T4 "Table 4 ‣ Truncation and sliding windows show marginal differences. ‣ 5 Results: Can existing models perform fine-grained literary STS? ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction") shows how the gold scores correlate with fandom and author-based categories. Many of the gold scores are positively correlated with the overall (fandom-specific) category, reflecting the fact that this category incorporates all other narrative features into its ranking. However, the majority are not as strongly correlated with the fandom-specific, fandom, and author scores, indicating that the information they capture extends beyond these categories.
+
+Model Plot Character States Relationship Dynamics Theme Time Style Tone &Content Overall(Fandom-Agnostic)
+Linq-Embed 8.36 6.00 4.94 4.38 24.96 6.63 9.90 16.59
++ SW 10.30 5.90 9.36 7.03 21.67 7.30 6.26 7.37
+GTE-Qwen2 15.44 5.36 9.43 14.43 33.19 14.01 20.35 24.38
++ SW 13.06-1.30 10.41 7.27 24.75 12.78 1.86 11.02
+SFR-Embed 18.82-0.18 5.79 2.59 24.30-5.28 7.52 18.07
++ SW 10.46 7.33 10.91 8.12 22.50 9.62 8.22 10.47
+GTE-ModernBERT 11.64 12.78 12.98 0.74 5.07 8.82 16.20 16.77
++ SW 18.32 6.24 23.59 7.72 38.21 15.65 6.48 22.13
+m2-BERT-32k 5.82 7.88 16.62 2.88-2.14-2.54 12.98 9.99
++ SW 8.82 7.11 13.07 6.58 2.28 4.61 18.17 12.81
+Voyage-3-large 14.96 7.66 16.93 11.53 21.84 21.91 15.79 20.01
++ SW 13.99 7.04 12.51 15.24 23.54 19.04 11.61 19.81
+Claude+SW 6.23 8.16 7.33 13.07 21.77 5.71 9.94 8.99
+
+Table 5:  When using category-specific instructions, rank-correlation does not show notable improvement and is still quite poor on average. Statistically significant results are highlighted.
+
+While capturing author and fandom information is not inherently harmful, the outsized impact of these (trivially computable from metadata) features on embedding-based similarity scores limits their applicability to analysis looking for more subtle phenomena like theme or trope similarity.
+
+#### Category-specific embeddings show minimal impact.
+
+Table[5](https://arxiv.org/html/2510.20926v1#S5.T5 "Table 5 ‣ Models overindex on surface features. ‣ 5 Results: Can existing models perform fine-grained literary STS? ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction") shows the performance with category-specific instructions across each model and category. When comparing correlations for a given model and context-handling method, category-specific embeddings outperform non-specific embeddings exactly 50% of the time.
+
+6 Related Work
+--------------
+
+#### Long-context and embedding evaluation.
+
+A number of datasets for long-context evaluation have included literary texts. BookSum (Kryściński et al., [2019](https://arxiv.org/html/2510.20926v1#bib.bib31)) involves summarization over public domain books. LongBench, LongBenchv2, and HELMET (Bai et al., [2023](https://arxiv.org/html/2510.20926v1#bib.bib6), [2024](https://arxiv.org/html/2510.20926v1#bib.bib7); Yen et al., [2025](https://arxiv.org/html/2510.20926v1#bib.bib56)) include question answering over NarrativeQA (Kočiský et al., [2018](https://arxiv.org/html/2510.20926v1#bib.bib29)); (Zhang et al., [2024c](https://arxiv.org/html/2510.20926v1#bib.bib59)) introduces summarization and QA tasks over a set of novels with entity names replaced to reduce the impact of potential contamination. Embedding-focused datasets include STS tasks but focus primarily on very short inputs Muennighoff et al. ([2023](https://arxiv.org/html/2510.20926v1#bib.bib38)); LongEmbed Zhu et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib60)), which evaluates long-context embedding but not on STS, instead using QA tasks over NarrativeQA and SummScreen screenplays Chen et al. ([2022](https://arxiv.org/html/2510.20926v1#bib.bib11)). None of these benchmarks measure performance on long-context STS tasks, which are of particular interest to digital humanities and literary scholars Sobchuk and Šeļa ([2024](https://arxiv.org/html/2510.20926v1#bib.bib46)).
+
+#### NLP tools for literary studies.
+
+A number of works have studied the applicability of NLP methods to digital humanities tasks on public-domain literature. Bamman et al. ([2024](https://arxiv.org/html/2510.20926v1#bib.bib8)) compare LLMs to traditional supervised methods on a wide range of tasks within literary studies. Other works propose novel computational approaches to analyze elements of fictional texts that are of interest to literary scholars, such as character mobility (Wilkens et al., [2024](https://arxiv.org/html/2510.20926v1#bib.bib52)), emotional arc (Öhman et al., [2022](https://arxiv.org/html/2510.20926v1#bib.bib61)), and narrative pacing (Bamman et al., [2014](https://arxiv.org/html/2510.20926v1#bib.bib10)). Kohlmeyer et al. ([2021](https://arxiv.org/html/2510.20926v1#bib.bib30)) propose lib2vec, a method for representing facets of fictional texts using multiple embeddings; because our similarity categories differ, direct application of their method to FicSim is challenging. The (in)applicability of NLP systems to downstream uses has also been studied in other domains, including law (Kapoor et al., [2024](https://arxiv.org/html/2510.20926v1#bib.bib26)) and materials science (Gururaja et al., [2025](https://arxiv.org/html/2510.20926v1#bib.bib21)).
+
+7 Conclusion
+------------
+
+We present FicSim, a dataset of stories and similarity labels for benchmarking model performance on long-context STS tasks within fictional texts. Using FicSim, we show that there is no single model that performs well across all types of similarity—and there are types of similarity for which no model performs well. In corpora with strong superficial similarities, like author or fandom overlap, embeddings may capture this information at the expense of other types of similarity. For this specific type of task, bigger (or more expensive) models are not uniformly better than their smaller, cheaper alternatives. Our evaluation of sliding window attention and category-specific embeddings also demonstrates that sensible modifications to the model to adapt to long-form or literary texts have a minimal impact on performance. We call for the careful evaluation of models on the particular task they are applied to, with annotation or validation by subject-matter experts.
+
+The poor performance of otherwise strong models on FicSim highlights the substantial gap that exists between current models and their utility for literary applications. Our models fail to capture finegrained literary similarity and overindex on superficial features of the text in their embeddings. We expect that clever system design or additional domain-specific training could improve performance within this generation of embedding models, and we encourage the evaluation on literary tasks for future embedding model releases.
+
+We hope FicSim will help digital humanities researchers make informed decisions about model selection for tasks relating to story similarity, encourage more evaluation of embedding models on DH tasks, and serve as an example of how creative works can be used for academic research without circumventing creators’ rights and wishes.
+
+Limitations
+-----------
+
+#### Data Collection
+
+While we were originally informed by an AO3 support member that leaving comments on fanfics would be an appropriate method for soliciting consent, our account was later temporarily suspended on the basis that we were leaving spam messages. (We had used the same introductory message to reach out to each author, in alignment with IRB protocols.) Our attempt to appeal the suspension was unsuccessful, despite our explanation that we were following instructions we had been given by another AO3 team member. When the temporary suspension was lifted, we decided not to attempt further data collection, because we ultimately did not want to be using AO3 in a way that further increased tension between machine learning researchers and fanfiction writers.
+
+Thus, while we were able to assemble a dataset using the methods outlined in this project, an exact replication of our process would not be appropriate.
+
+#### Embedding methods
+
+It is not possible to consider every possible means of constructing embeddings; while we aimed to capture a set of models and methods that were representative of those applied in digital humanities works with embeddings, it is possible that there exist other methods that would outperform those presented as baselines here. In particular, computing similarities using multiple-embedding strategies is likely to improve performance. We leave devising better embedding strategies for literary domain text to future work.
+
+#### Language
+
+We consider only stories written in English because of our need to reach out to each author individually. While we believe the fanfictions within FicSim represent an interesting selection of works across these similarity dimensions, differences exist between literary corpora. It is possible that models that excel at similarity on FicSim would nevertheless struggle on STS tasks for 19th century English literature, short-form satirical poetry from social media, or any other number of specialized literary domains. We see FicSim as an initial step towards improved literary-domain evaluation.
+
+Acknowledgments
+---------------
+
+We extend our deepest gratitude towards the fanfiction authors who have given us permission to include their works in our dataset and without whom this project would not have been possible.
+
+We would also like to thank David Mimno, Maarten Sap, Chap Morack, Suguru Ishizaki, and our reviewers for their helpful feedback on our work.
+
+AB was supported by a grant from the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE2140739. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the sponsors.
+
+References
+----------
+
+*   Algee-Hewitt and Fredner (2023) Mark Algee-Hewitt and Erik Fredner. 2023. [Typicality](https://litlab.stanford.edu/projects/typicality/). 
+*   Allison et al. (2012) Sarah Allison, Ryan Heuser, Matthew Jockers, Franco Moretti, and Michael Witmore. 2012. Quantitative formalism: An experiment. _N+1_, 13:81–108. 
+*   Anthropic (2025) Anthropic. 2025. [Claude 3.7 sonnet and claude code](https://www.anthropic.com/news/claude-3-7-sonnet). Accessed: 2025-05-19. 
+*   Archive of Our Own (2023) Archive of Our Own. 2023. [A statement on ai and fanworks](https://archiveofourown.org/admin_posts/25888). Accessed: 2025-05-20. 
+*   Arzbaecher (2023) Lauren Arzbaecher. 2023. [10 books you didn’t know started out as fan fiction, from ’Twilight’ to ’Star Wars’-inspired stories](https://www.businessinsider.com/guides/learning/fanfiction-books). 
+*   Bai et al. (2023) Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. 2023. [LongBench: a bilingual, multitask benchmark for long context understanding](https://arxiv.org/abs/2308.14508). 
+*   Bai et al. (2024) Yushi Bai, Shangqing Tu, Jiajie Zhang, Hao Peng, Xiaozhi Wang, Xin Lv, Shulin Cao, Jiazheng Xu, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. 2024. [LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks](https://doi.org/10.48550/arxiv.2412.15204). _arXiv (Cornell University)_. 
+*   Bamman et al. (2024) David Bamman, Kent K. Chang, Li Lucy, and Naitian Zhou. 2024. [On Classification with Large Language Models in Cultural Analytics](https://doi.org/10.48550/arxiv.2410.12029). _arXiv (Cornell University)_. 
+*   Bamman et al. (2020) David Bamman, Olivia Lewke, and Anya Mansoor. 2020. [An annotated dataset of coreference in English literature](https://aclanthology.org/2020.lrec-1.6/). In _Proceedings of the Twelfth Language Resources and Evaluation Conference_, pages 44–54, Marseille, France. European Language Resources Association. 
+*   Bamman et al. (2014) David Bamman, Ted Underwood, and Noah A. Smith. 2014. [A Bayesian Mixed Effects Model of Literary Character](https://doi.org/10.3115/v1/p14-1035). _Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics_, Volume 1: Long Papers. 
+*   Chen et al. (2022) Mingda Chen, Zewei Chu, Sam Wiseman, and Kevin Gimpel. 2022. [SummScreen: A dataset for abstractive screenplay summarization](https://doi.org/10.18653/v1/2022.acl-long.589). In _Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_, pages 8602–8615, Dublin, Ireland. Association for Computational Linguistics. 
+*   Choi et al. (2024) Chanyeol Choi, Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, Minkyung Cho, and Jy yong Sohn. 2024. [Linq-embed-mistral technical report](https://arxiv.org/abs/2412.03223). _Preprint_, arXiv:2412.03223. 
+*   Duarte et al. (2024) André Duarte, V, Xuandong Zhao, Arlindo L. Oliveira, and Lei Li. 2024. [DE-COP: Detecting copyrighted content in language models training data](https://doi.org/10.48550/arxiv.2402.09910). _arXiv (Cornell University)_. 
+*   Elazar et al. (2024) Yanai Elazar, Akshita Bhagia, Ian Magnusson, Abhilasha Ravichander, Dustin Schwenk, Alane Suhr, Pete Walsh, Dirk Groeneveld, Luca Soldaini, Sameer Singh, Hanna Hajishirzi, Noah A. Smith, and Jesse Dodge. 2024. [What’s in my big data?](https://arxiv.org/abs/2310.20707)_International Conference on Learning Representations_. 
+*   Fazekas (2022) Angela Fazekas. 2022. [_Creative Becomings: Explicit Fanfiction, Reinventing Adolescence, and Queer Relationality_](https://www.proquest.com/dissertations-theses/creative-becomings-explicit-fanfiction/docview/2743629194/se-2). Ph.D. thesis. Copyright - Database copyright ProQuest LLC; ProQuest does not claim copyright in the individual underlying works; Last updated - 2024-02-16. 
+*   Fu et al. (2023) Daniel Y Fu, Simran Arora, Jessica Grogan, Isys Johnson, Sabri Eyuboglu, Armin W Thomas, Benjamin Spector, Michael Poli, Atri Rudra, and Christopher Ré. 2023. Monarch mixer: A simple sub-quadratic gemm-based architecture. In _Advances in Neural Information Processing Systems_. 
+*   Gero et al. (2025) Katy Ilonka Gero, Meera Desai, Carly Schnitzler, Nayun Eom, Jack Cushman, and Elena L. Glassman. 2025. [Creative writers’ attitudes on writing as training data for large language models](https://doi.org/10.1145/3706598.3713287). In _Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems_, CHI ’25, page 1–16. ACM. 
+*   Gioia and Gwynn (2006) Dana Gioia and R.S. Gwynn. 2006. _The art of the short story_. Longman Publishing Group. 
+*   Godbole et al. (2024) Aditi Godbole, Jabin Geevarghese George, and Smita Shandilya. 2024. [Leveraging long-context large language models for multi-document understanding and summarization in enterprise applications](https://arxiv.org/abs/2409.18454). _Preprint_, arXiv:2409.18454. 
+*   (20) goodreads. [Enemies to Lovers Books](https://www.goodreads.com/shelf/show/enemies-to-lovers). 
+*   Gururaja et al. (2025) Sireesh Gururaja, Nupoor Gandhi, Jeremiah Milbauer, and Emma Strubell. 2025. [Beyond text: Characterizing domain expert needs in document research](https://arxiv.org/abs/2504.12495). _Preprint_, arXiv:2504.12495. 
+*   Harmon (2003) William Harmon. 2003. _A handbook to literature_, 9 edition. Prentice Hall. 
+*   Hellekson and Busse (2014) Karen Hellekson and Kristina Busse. 2014. [_The Fan Fiction Studies Reader_](https://doi.org/10.2307/j.ctt20p58d6). University of Iowa Press. 
+*   Jacobsen and Kristensen-McLachlan (2024) Mia Jacobsen and Ross Deans Kristensen-McLachlan. 2024. [Admiration and Frustration: A Multidimensional Analysis of Fanfiction](https://ceur-ws.org/Vol-3834/paper57.pdf). _CHR 2024: Computational Humanities Research Conference_. 
+*   Jerasa and Boffone (2021) Sarah Jerasa and Trevor Boffone. 2021. [BookTok 101: TikTok, Digital Literacies, and Out‐of‐School Reading Practices](https://doi.org/10.1002/jaal.1199). _Journal of Adolescent & Adult Literacy_, 65(3):219–226. 
+*   Kapoor et al. (2024) Sayash Kapoor, Peter Henderson, and Arvind Narayanan. 2024. [Promises and pitfalls of artificial intelligence for legal applications](https://arxiv.org/abs/2402.01656). _Preprint_, arXiv:2402.01656. 
+*   Karpinska et al. (2024) Marzena Karpinska, Katherine Thai, Kyle Lo, Tanya Goyal, and Mohit Iyyer. 2024. [One thousand and one pairs: A “novel” challenge for long-context language models](https://doi.org/10.18653/v1/2024.emnlp-main.948). In _Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing_, pages 17048–17085, Miami, Florida, USA. Association for Computational Linguistics. 
+*   Kleymann et al. (2022) Rabea Kleymann, Andreas Niekler, and Manuel Burghardt. 2022. [Conceptual Forays: A corpus-based study of “Theory” in Digital Humanities Journals](https://doi.org/10.22148/001c.55507). _Journal of cultural analytics_, 7(4). 
+*   Kočiský et al. (2018) Tomáš Kočiský, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, Gábor Melis, and Edward Grefenstette. 2018. [The NarrativeQA reading comprehension challenge](https://doi.org/10.1162/tacl_a_00023). _Transactions of the Association for Computational Linguistics_, 6:317–328. 
+*   Kohlmeyer et al. (2021) Lasse Kohlmeyer, Tim Repke, and Ralf Krestel. 2021. [Novel views on Novels: embedding multiple facets of long texts](https://hpi.de/fileadmin/user_upload/fachgebiete/naumann/publications/PDFs/2021_kohlmeyer_novel.pdf). _WI-IAT ’21: Web Intelligence and Intelligent Agent Technology_, page 8. 
+*   Kryściński et al. (2019) Wojciech Kryściński, Nazneen Rajani, Divyansh Agarwal, Caiming Xiong, and Dragomir Radev. 2019. [BookSum: A Collection of Datasets for Long-form Narrative Summarization](https://arxiv.org/abs/2105.08209). _arXiv.org_. 
+*   Lee et al. (2025) Jinhyuk Lee, Feiyang Chen, Sahil Dua, Daniel Cer, Madhuri Shanbhogue, Iftekhar Naim, Gustavo Hernández Ábrego, Zhe Li, Kaifeng Chen, Henrique Schechter Vera, Xiaoqi Ren, Shanfeng Zhang, Daniel Salz, Michael Boratko, Jay Han, Blair Chen, Shuo Huang, Vikram Rao, Paul Suganthan, Feng Han, Andreas Doumanoglou, Nithi Gupta, Fedor Moiseev, Cathy Yip, Aashi Jain, Simon Baumgartner, Shahrokh Shahi, Frank Palma Gomez, Sandeep Mariserla, Min Choi, Parashar Shah, Sonam Goenka, Ke Chen, Ye Xia, Koert Chen, Sai Meher Karthik Duddu, Yichang Chen, Trevor Walker, Wenlei Zhou, Rakesh Ghiya, Zach Gleicher, Karan Gill, Zhe Dong, Mojtaba Seyedhosseini, Yunhsuan Sung, Raphael Hoffmann, and Tom Duerig. 2025. [Gemini Embedding: Generalizable Embeddings from Gemini](https://arxiv.org/abs/2503.07891). _arXiv.org_. 
+*   Li et al. (2023) Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023. Towards general text embeddings with multi-stage contrastive learning. _arXiv preprint arXiv:2308.03281_. 
+*   Lincoln et al. (2011) Yvonna S Lincoln, Susan A Lynham, and Egon G Guba. 2011. Paradigmatic controversies, contradictions, and emerging confluences, revisited. _The Sage handbook of qualitative research_, 4:97–128. 
+*   Lyons and Tappeiner (2008) Catherine Lyons and Elisabeth Tappeiner. 2008. [Cataloging 2.0: Metadata research and initiatives at a community college library](https://doi.org/10.1080/10911360802087358). _Journal of Library Metadata_, 8(2):155–157. 
+*   Majnaric (2024) Lucija Majnaric. 2024. [Novel in the Time of the Internet: A Closer Look at the Fanfiction Phenomenon](https://repozitorij.ffzg.unizg.hr/islandora/object/ffzg:12322). 
+*   Meng et al. (2024) Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. 2024. [Sfr-embedding-mistral:enhance text retrieval with transfer learning](https://www.salesforce.com/blog/sfr-embedding/). Salesforce AI Research Blog. 
+*   Muennighoff et al. (2023) Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. [MTEB: Massive text embedding benchmark](https://doi.org/10.18653/v1/2023.eacl-main.148). In _Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics_, pages 2014–2037, Dubrovnik, Croatia. Association for Computational Linguistics. 
+*   Nie et al. (2024) Yuqi Nie, Yaxuan Kong, Xiaowen Dong, John M. Mulvey, H.Vincent Poor, Qingsong Wen, and Stefan Zohren. 2024. [A survey of large language models for financial applications: Progress, prospects and challenges](https://arxiv.org/abs/2406.11903). _Preprint_, arXiv:2406.11903. 
+*   Palavalli et al. (2024) Medha Palavalli, Amanda Bertsch, and Matthew R. Gormley. 2024. [A taxonomy for data contamination in large language models](https://arxiv.org/abs/2407.08716). 
+*   Piper and Bagga (2022) Andrew Piper and Sunyam Bagga. 2022. [Toward a Data-Driven Theory of Narrativity](https://doi.org/10.1353/nlh.2022.a898332). _New Literary History_, 54(1):879–901. 
+*   Popova (2018) Milena Popova. 2018. [_"Slight dub-con but they both wanted it hardcore": Erotic fanfiction as a form of cultural activism around sexual consent_](https://uwe-repository.worktribe.com/output/869400). Ph.D. thesis. 
+*   Reimers and Gurevych (2019) Nils Reimers and Iryna Gurevych. 2019. [Sentence-BERT: Sentence embeddings using Siamese BERT-networks](https://doi.org/10.18653/v1/D19-1410). In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_, pages 3982–3992, Hong Kong, China. Association for Computational Linguistics. 
+*   Schöch (2017) Christof Schöch. 2017. [Topic Modeling Genre: An Exploration of French Classical and Enlightenment Drama](https://www.digitalhumanities.org/dhq/vol/17/2/000691/000691.html). _Digital Humanities Quarterly_, 11(2). 
+*   Sims et al. (2019) Matthew Sims, Jong Ho Park, and David Bamman. 2019. [Literary Event Detection](https://doi.org/10.18653/v1/p19-1353). _Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics_. 
+*   Sobchuk and Šeļa (2024) Oleg Sobchuk and Artjoms Šeļa. 2024. [Computational thematics: comparing algorithms for clustering the genres of literary fiction](https://doi.org/10.1057/s41599-024-02933-6). _Humanities and Social Sciences Communications_, 11(1). 
+*   Strauss and Corbin (1990) Anselm Strauss and Juliet Corbin. 1990. _Basics of qualitative research_. Sage publications. 
+*   The Authors Guild (2023) The Authors Guild. 2023. [New authors guild ai survey reveals that authors overwhelmingly want consent and compensation for use of their works](https://authorsguild.org/news/ag-ai-survey-reveals-authors-overwhelmingly-want-consent-and-compensation-for-use-of-their-works/). Accessed: 2025-05-20. 
+*   Underwood et al. (2018) Ted Underwood, David Bamman, Sabrina Lee, Boris Capitanu, Hoyt Long, Richard Jean So, Teddy Roland, Laura Mandell, Allen Riddell, Andrew Piper, and Stephen J. Downie. 2018. [The transformation of Gender in English-Language fiction](https://doi.org/10.22148/16.019). _Cultural Analytics_. 
+*   Voyage AI (2025) Voyage AI. 2025. [voyage-3-large: the new state-of-the-art general-purpose embedding model](https://blog.voyageai.com/2025/01/07/voyage-3-large/). Accessed: 2025-05-19. 
+*   Wang et al. (2019) Zhiguo Wang, Patrick Ng, Xiaofei Ma, Ramesh Nallapati, and Bing Xiang. 2019. [Multi-passage BERT: A globally normalized BERT model for open-domain question answering](https://doi.org/10.18653/v1/D19-1599). In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_, pages 5878–5882, Hong Kong, China. Association for Computational Linguistics. 
+*   Wilkens et al. (2024) Matthew Wilkens, Elizabeth F. Evans, Soni Sandeep, David Bamman, and Andrew Piper. 2024. [Small Worlds. Measuring the Mobility of Characters in English-Language Fiction](https://tuprints.ulb.tu-darmstadt.de/27523/). _Journal of Computational Literary Studies [preprint]_. 
+*   Wolf et al. (2020) Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. [Huggingface’s transformers: State-of-the-art natural language processing](https://arxiv.org/abs/1910.03771). _Preprint_, arXiv:1910.03771. 
+*   Works (2023) Organization For Transformative Works. 2023. [AI and Data Scraping on the Archive](https://archiveofourown.org/admin_posts/25888). 
+*   Xu et al. (2022) Ying Xu, Dakuo Wang, Mo Yu, Daniel Ritchie, Bingsheng Yao, Tongshuang Wu, Zheng Zhang, Toby Li, Nora Bradford, Branda Sun, Tran Hoang, Yisi Sang, Yufang Hou, Xiaojuan Ma, Diyi Yang, Nanyun Peng, Zhou Yu, and Mark Warschauer. 2022. [Fantastic questions and where to find them: FairytaleQA – an authentic dataset for narrative comprehension](https://doi.org/10.18653/v1/2022.acl-long.34). In _Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_, pages 447–460, Dublin, Ireland. Association for Computational Linguistics. 
+*   Yen et al. (2025) Howard Yen, Tianyu Gao, Minmin Hou, Ke Ding, Daniel Fleischer, Peter Izsak, Moshe Wasserblat, and Danqi Chen. 2025. Helmet: How to evaluate long-context language models effectively and thoroughly. In _International Conference on Learning Representations (ICLR)_. 
+*   Zhang et al. (2024a) Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialong Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, et al. 2024a. mgte: Generalized long-context text representation and reranking models for multilingual text retrieval. In _Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track_, pages 1393–1412. 
+*   Zhang et al. (2024b) Xinhao Zhang, Olga Seminck, Pascal Amsili, and Lattice. 2024b. [Remember to Forget: A study on verbatim memorization of literature in large language models](https://ceur-ws.org/Vol-3834/paper96.pdf). In _CHR 2024: Computational Humanities Research Conference_. 
+*   Zhang et al. (2024c) Xinrong Zhang, Yingfa Chen, Shengding Hu, Zihang Xu, Junhao Chen, Moo Khai Hao, Xu Han, Zhen Leng Thai, Shuo Wang, Zhiyuan Liu, and Maosong Sun. 2024c. [∞\infty bench: Extending long context evaluation beyond 100k tokens](https://arxiv.org/abs/2402.13718). _Preprint_, arXiv:2402.13718. 
+*   Zhu et al. (2024) Dawei Zhu, Liang Wang, Nan Yang, Yifan Song, Wenhao Wu, Furu Wei, and Sujian Li. 2024. [LongEmbed: Extending embedding models for long context retrieval](https://doi.org/10.18653/v1/2024.emnlp-main.47). In _Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing_, pages 802–816, Miami, Florida, USA. Association for Computational Linguistics. 
+*   Öhman et al. (2022) Emily Öhman, Yuri Bizzoni, Pascale Moreira, and Kristoffer Nielbo. 2022. EmotionArcs: Emotion Arcs for 9,000 Literary Texts. _Proceedings of LaTeCH-CLfL_. 
+
+Appendix A Additional dataset documentation
+-------------------------------------------
+
+Tag standardization Much of our tag categorization and standardization process was inspired by AO3’s own practices. AO3 has a team dedicated to tag wrangling, which is the task of maintaining a database of canonical tags, sorting and organizing those tags, and linking tags to their canonicalized form. Thanks to this high standard of organization, many tags can be mapped back to canonical tags. Non-canonical tags often come in the forms of meta commentary, merging of multiple canonical tags, or a modification of canonical tags to include fandom-specific references. These tags still contain valuable information about their stories, and looking at them in conjunction with similar canonical tags sometimes helped us determine appropriate categorizations.
+
+Comparisons by category. Not all stories have a similarity score along every axis. Table[6](https://arxiv.org/html/2510.20926v1#A1.T6 "Table 6 ‣ Appendix A Additional dataset documentation ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction") lists the number of comparisons possible in each category.
+
+Comparison axis Pair count
+Plot 4005
+Character States 3240
+Relationship Dynamics 2278
+Theme 1431
+Time 105
+Style 1431
+Fanfiction Tone & Content Tags 1275
+Overall (All fandom-agnostic tags)4005
+Overall (not fandom-agnostic tags)4005
+Fandom-Specific Tags 4005
+Fandoms 4005
+Author 4005
+Total 33,790
+
+Table 6:  The number of pairwise comparisons in FicSim by category. We exclude stories from pairwise comparisons in categories where they lack tagging. 
+
+Length Figure[3](https://arxiv.org/html/2510.20926v1#A1.F3 "Figure 3 ‣ Appendix A Additional dataset documentation ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction") shows the length distribution of texts in FicSim.
+
+![Image 3: Refer to caption](https://arxiv.org/html/2510.20926v1/x2.png)
+
+Figure 3: Story lengths in FicSim range from 10k to >400k words.
+
+Additional metadata. In addition to the similarity scores and full texts, FicSim contains many other metadata fields about each story, including chapter splits, author IDs, author-written summaries (where available), and a number of AO3-imposed classifications (e.g. the genders of the characters in the primary relationship in the story). While we do not explicitly clean data for additional non-tag categories, we make these data available in the hope that they will be useful to other researchers working on literary applications.
+
+Appendix B Additional documentation of models
+---------------------------------------------
+
+Abbreviated name HF or API name Max context Pooling strategy Param count
+Linq-Embed Linq-Embed-Mistral 32,768 Last-token 7B
+SFR-Embedding SFR-Embedding-Mistral 32,768 Last-token 7B
+GTE-Qwen2 gte-Qwen2-7B-instruct 32,768 All tokens 7B
+GTE-ModernBERT gte-modernbert-base 8,192 CLS 149M
+m2-BERT-32k m2-bert-80M-32k-retrieval 32,768 CLS 80M
+Voyage-3-large voyage-3-large 32,768 Unknown Unknown
+Claude-3.7-Sonnet claude-3-7-sonnet-20250219 200,000 n/a Unknown
+
+Table 7:  Additional details on the models ran. 
+
+This section contains additional details for reproducing the embedding methods.
+
+#### Models.
+
+We evaluate on 7 models, described in detail in Table[7](https://arxiv.org/html/2510.20926v1#A2.T7 "Table 7 ‣ Appendix B Additional documentation of models ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction").
+
+Pooling. We follow each model’s default strategy for constructing embeddings: Linq-Embed and SFR-Embedding use last-token pooling, GTE-Qwen2 uses mean pooling of all tokens, and GTE-ModernBERT uses a CLS token. In cases where we obtain multiple embeddings (i.e., when using a sliding window), we average all embeddings to produce a single embedding for each document. When mean-pooling multiple embeddings, we average after pooling all token embeddings instead of averaging a single pooled embedding from each window. In mean-pooled embeddings, we include prompt tokens but take only one embedding of each overlapped token in sliding windows with overlap.
+
+#### Software.
+
+For single-window (truncation) approaches, we use sentence_transformers Reimers and Gurevych ([2019](https://arxiv.org/html/2510.20926v1#bib.bib43)). For sliding window approaches, we use Hugging Face transformers Wolf et al. ([2020](https://arxiv.org/html/2510.20926v1#bib.bib53)). We call Voyage and Claude through their respective APIs, which (at the time of writing) do not retain user data for model training. A limited amount of language model assistance was used for writing simple data processing scripts; all code was verified by the authors.
+
+#### Computational resources.
+
+All local models were run on a mixture of L40S and H100 GPUs; we estimate that the total compute time in development and running the final embedding methods did not exceed 200 GPU-hours. The total cost of development and running the API-based methods was approximately $80, of which $78 was the cost of running Claude.
+
+#### Prompts.
+
+We use the same prompt for all models except Claude; the category-specific prompts are in Table[8](https://arxiv.org/html/2510.20926v1#A2.T8 "Table 8 ‣ Prompts. ‣ Appendix B Additional documentation of models ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction"). For Claude, we use the system prompt “Below is a long-form fanfiction written in English. You will be asked to summarize this story.” We provide the full text of the story as a user message, then provide an additional user message with instructions. The instruction message always begins “Please write a detailed summary of this story, using up to 5,000 words.” It optionally also has a category-specific instruction; these instructions are listed in Table[9](https://arxiv.org/html/2510.20926v1#A2.T9 "Table 9 ‣ Prompts. ‣ Appendix B Additional documentation of models ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction").
+
+In the rare (3) cases where a story exceeds Claude’s context window, we summarize as much of the story as possible in a first API call and provide the summary plus the remainder of the story in a second API call. In this second call, the system prompt is changed to “Below is a long-form fanfiction written in English. The first section is a summary of the first portion of the story, and then the remainder of the story follows. You will be asked to summarize this story.” The two sections of the input are labeled “Summary:” and “Remainder of the story:” and separated by a line of dashes. The instruction is changed to “Please write a detailed summary of the full story, using up to 5,000 words. You may copy the summary of the first portion of the story exactly, or modify it as you wish.” along with any category-specific instruction.
+
+Category Prompt
+Plot Identify the main plot arc of the fanfiction based on the text.
+Character state Identify the main character states of the fanfiction based on the text.
+Relationship dynamic Identify the main relationship dynamics of the fanfiction based on the text.
+Theme Identify the main themes of the fanfiction based on the text.
+Time Identify the main time period of the fanfiction based on the text.
+Style Identify the main literary style of the fanfiction based on the text.
+Tone & content Identify the main fanfiction-specific tone and content descriptors of the fanfiction based on the text.
+Overall[no prompt]
+
+Table 8: Prompt for all embedding models. The prompt (with any applicable model-specific formatting) is prepended to the beginning of the text and the start of every sliding window.
+
+Category Prompt
+Plot In your summary, pay particular attention to the plot of the text.
+Character state In your summary, pay particular attention to the attributes of the characters in the text.
+Relationship dynamic In your summary, pay particular attention to the relationship dynamics of the characters in the text.
+Theme In your summary, pay particular attention to the themes of the text.
+Time In your summary, pay particular attention to the temporal setting of the text.
+Style In your summary, pay particular attention to the literary style of the text.
+Tone & content In your summary, pay particular attention to any fanfiction-specific tone or tropes exhibited in the text.
+Overall[no additional prompt]
+
+Table 9: Prompt for Claude summarization. This is appended as part of the last user message, after the system message and a user message containing the full text of the story.
+
+Appendix C Author Consent Process
+---------------------------------
+
+Archive of Our Own does not have a private messaging feature, and authors do not generally post contact information (or real names) on their fanfictions. After consulting with the AO3 policy team, we agreed to reach out to authors by leaving a comment on the stories we would like to use. This comment then directs them to our main Reddit post, which links to the project’s webpage, explains the study and its terms, and offers a locale for authors to ask questions directly of the authors. This process was approved as Carnegie Mellon University IRB Study 00000260.
+
+Revoking consent. We maintain a Google Form for requesting removal of a story at any time, with no questions asked. We commit to monitoring this form in perpetuity and removing fanfiction promptly if authors choose to revoke consent. For replication of results on the dataset, we will clearly label the dataset on Hugging Face and the repository with a version number, and ask that anyone using the dataset report the evaluation version.
+
+### C.1 Outreach process documents
+
+We provide the exact text of the comments to reach out to authors (Figure[4](https://arxiv.org/html/2510.20926v1#A3.F4 "Figure 4 ‣ C.1 Outreach process documents ‣ Appendix C Author Consent Process ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")) and the text of the Reddit post (Figure[5](https://arxiv.org/html/2510.20926v1#A3.F5 "Figure 5 ‣ C.1 Outreach process documents ‣ Appendix C Author Consent Process ‣ FicSim: A Dataset for Multi-Faceted Semantic Similarity in Long-Form Fiction")). 12 12 12 Our original post cites 30k as our desired lower word limit for fanfiction contributions, but after seeing the volume and quality of fanfic contributions we received below this threshold, we decided to include texts above 10k words in our dataset, provided they had adequately detailed tagging.
+
+Figure 4: Sample comment on fanfiction
+
+Figure 5: Post on Reddit with information on how to consent
+
+Appendix D License
+------------------
+
+Copyright 2025, the original author of each fanfiction (used with permission).
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+No part of the text of any story in the dataset will be used in training of any machine learning model, or in any system that involves a model retaining memory, knowledge, or other influence from the story text.
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.