Fans Reconstruct Heroes: Modeling Fictional Characters in Participatory Culture

This manuscript was submitted as 'full paper' and should be reviewed along the usual dimensions for research contributions which include (1) originality, (2) significance of the results, and (3) quality of writing. Please also assess the data file provided by the authors under “Long-term stable URL for resources”. In particular, assess (A) whether the data file is well organized and in particular contains a README file which makes it easy for you to assess the data, (B) whether the provided resources appear to be complete for replication of experiments, and if not, why, (C) whether the chosen repository, if it is not GitHub, Figshare or Zenodo, is appropriate for long-term repository discoverability, and (4) whether the provided data artifacts are complete. Please refer to the reviewer instructions and the FAQ for further information.

(1) This research presents an original effort in modeling fictional characters within participatory culture. While the methodology and workflow follow standard practices, they are executed thoughtfully.

(2) The paper promises to provide an interesting model and resource for fanfiction studies, and may serve as an exemplar for modeling characters in other domains.

(3) The writing is very clear and easy to follow.

General Comments
The authors have addressed most of my previously raised comments, and I have no further major concerns to add. My only minor question, which is not a critical issue, concerns the semantic similarity matching used in Steps 6 and 9.

Regarding Step 6, the manuscript states:
"[i]n Step 6, we evaluated the alignment of raw character attribute values from our 141-character
dataset against the constructed vocabulary using a two-stage pipeline: rule-based normalisation ..., followed by semantic
similarity matching with Sentence-BERT at a cosine threshold of 0.8. Across all evaluated fields, 167 out of 238 raw values
(70.2%) were successfully matched to a vocabulary concept."

You mention that the similarity matching in Step 6 was achieved via Sentence-BERT, which computes similarity distance at the sentence level. In this case, what exactly are you computing against within the controlled vocabulary? I assume this would be the description rather than the concept names or labels. A brief clarification in the text would be helpful.

In Step 9, you note:
"For each candidate value ...compute its cosine similarity against all concepts in the controlled vocabulary using Sentence-BERT embeddings. If the highest similarity score is above 0.85, the candidate is accepted as a valid instance of the corresponding vocabulary concept..."

I wonder whether a heavy reliance on non-domain-specific Sentence-BERT embeddings might lead to a loss of nuance or a misreading of nuanced semantics in your dataset. Given that the attributes you are trying to identify seem fairly unambiguous and distinct in nature, this may not be a major issue. However, it is always worth keeping in mind that off-the-shelf embeddings can sometimes dilute the interpretability of a bespoke ontology or semantic model. It would be helpful to evaluate these results to verify whether threshold-based acceptance is fully reliable in this context.

The paper is well-written, and the first two sections are a pleasure to read. The paper has improved, and several criticisms that I and the other reviewers raised in the first round of reviews have been addressed. However, some relevant problems remain. In particular, I will leave it to the editors to evaluate the strong overlap with Pianzola, F., Cheng, L., Yang, X., & Panache, F. (2025) to determine whether the submitted paper meets the SWJ requirements for novelty. Additionally, some ontological issues remain. My evaluation of a "minor revision" assumes that the editor believes there is sufficient new material in the submitted paper. I will leave the decision to them.

As I mentioned in my initial review, I am primarily concerned about the novelty of the proposal. The paper, "Pianzola, F., Cheng, L., Yang, X., & Panache, F. (2025). The GOLEM ontology for narrative and fiction," cited in the manuscript, is now available, and I have read it. The paper provides a thorough overview of the GOLEM ontology, including the character module mentioned in the manuscript. First, as reviewer #2 pointed out in the first round of reviews, I think this comprehensive presentation provides a much better understanding of the GOLEM ontology and the character module (several links between this module and others in the submitted manuscript are only partially considered). Second, The ontological analysis of the notions discussed in the character module of the manuscript does not substantially expand upon the analysis presented in the published paper. Furthermore, it still has some issues and avoids deep discussion about the nature of characters. See the details below.
The novel contributions of the manuscript are limited to: (1) some additional methodological considerations (in Controlled Vocabulary Construction of Character Features); I find these interesting, even though they reflect standard practices; (2) a comparison with other ontologies of characters (in Existing Ontological Models for Fictional Characters); this comparison is comprehensive, but quite superficial in terms of ontological analysis; and (3) a case study about greek mythology in fanfiction; I find this useful to illustrate the ontological choices, but quite simplistic for validating the ontology (similarly for the proposed CQs that honestly seem quite trivial; furthermore, the authors claims that "These CQs thus served as design specifications" (p.15) while in (Pianzola et al. 2025) it is claimed that the development of GOLEM has been guided by theory-driven requirements, see section 3.3).
Personally, I think one could have included the case study in the 2025 publication or expanded it significantly in a new manuscript to provide a thorough evaluation (from the applicative perspective) of the entire ontology.

The integration with, and reuse of, DOLCE do not preserve the original meaning of the reused categories and relations. The authors appear to consider only the light versions of DOLCE (e.g., DUL), which omit important constraints present in the original FOL version. These constraints were introduced to better characterize the concepts in DOLCE, but they are absent from the light versions due to the limited expressive power of DLs. First, according to what is written in the manuscript and in the cited (Scotti et al. 2026), G1_Character is a subclass of DOLCE agentive-social-object which is a subclass of DOLCE non-physical-endurant. (1) In DOLCE non-physical-endurants depend on physical-endurants, but this dependence is neither made explicit nor discussed in the paper. (2) non-physical-endurants have only abstract-qualities, yet the authors suggest including physical descriptions, such as height or weight, among the features of characters. Second, G17_Character_Feature is a subclass of DOLCE region. In DOLCE, different endurants can be mapped to the same region through their qualities. Individual qualities, but not regions, inhere in a single endurant. However, the authors claim that "Biographical and physical characteristics (e.g., “blond hair”, “physician”) could be modeled as classes, but doing so would treat them as types shared between characters, rather than as particularized attributes tied to specific character versions. For example, a canonical Harry Potter and a fanfiction “Female Harry Potter” would both be linked to the same type “blond hair”, losing the ability to record that each version has its own occurrence of that feature. Therefore, we need instance-level particulars to model this diversity." This is not the idea behind DOLCE's regions. It is closer to the idea of individual qualities, which the authors explicitly discarded. Additionally, note that standard tropes and DOLCE's individual qualities are not the same: while trope theory views change as trope substitution, DOLCE views it as a change in the location of a quality in the corresponding space.

The authors claim: "Zollner-Weber (2009) examined the use of ontologies and logic reasoning in the humanities, identifying the tension between the rigidity of formal knowledge representation and the interpretative openness required by humanistic analysis—a tension GOLEM addresses by separating fixed ontological axioms from extensible controlled vocabularies."
I think this claim, as well as (for example) the decision to represent different types of characters as instances of E55_Types rather than as subclasses of G1_Character, deserves further discussion. The exact advantages of having character types reified in the domain in terms of reasoning and updating (of the ontology) are unclear to me. In all cases, if one needs to formally characterize the difference between two subtypes of G1_Character, some axioms are necessary. When types are reified, these axioms must involve individual constants, i.e., specific instances introduced in the ontology. Please provide a better justification for your choice.

As I mentioned in my first review, the main novelty of the character module, the G0_Character-Stoff concept, is not adequately characterized ontologically. Similar concepts are present in philosophical discussions about character identity, but the authors do not consider them. In particular, the difference with respect to the notion of "character type" is not discussed. The guidelines for understanding what makes a character in fan fiction to be considered Harry Potter (intended as an instance of Character-Stoff) remain totally obscure. Saying "community-recognition-based approach to cross-work character identity" avoids the ontological analysis of this interesting yet difficult notion. From an ontological perspective, this remains a weak point of the paper.

----

- The prefix gd: is not used in the Figures.

- (p.9) "Within this conceptual framework, the Character Module defines G1 Character as a subclass of CIDOC-CRM’s E28 Conceptual Object and dlp:social-object incorporating DOLCE’s social object taxonomy (Bottazzi and Ferrario 2009). This dual classification acknowledges characters "
In what sense this is a dual classification? According to (Scotti et al. 2026), crm:E28_Conceptual_Object is a subclass of dlp:social-object. Actually, the paper does not directly discuss the dual nature of characters that have features of real/concrete entities as well as fictional entities (e.g., they have an author and are created at a certain time).

- The inclusion of "blond hair" among the features of characters deserves more attention. The problem is linked to whether features are holistic. Is "blond hair" a feature of a character, or is it better to say that "blond" is a feature of a part of a character, i.e., its hair? The problem becomes more pronounced when considering distributive features like "striped."

- (p.12) "The SHACL shapes enforce two main constraints...". Perhaps the motivation behind the constraint (i) "hierarchy constraints" merits some explanation.