Introduction

The evolution of online reading platforms, from early content-centered libraries like Project Gutenberg and the Internet Archive to user-centered platforms such as Fanfiction.net, Archive of Our Own, and Wattpad, highlights the growing importance of online reading. These platforms not only provide a space for collective reading but also foster social interaction and community building, offering writers interactive opportunities [7] and readers direct engagement with authors [20]. Online reading communities are crucial for young readers and writers, fulfilling emotional and social needs [25], and providing a space for marginalized voices [5,3]. Many researches [10,21] have explored the various aspects of Digital Social Reading (DSR), and one of the many focuses is the reader response [11,19]. However, current research on storytelling and reader response has overlooked cross-cultural comparisons, with only a few exceptions [15].This research gap has been identified not only in the field of DSR but also in comparative literary studies, which primarily focus on understanding the cultural influences behind literature by examining authors as transcultural readers rather than investigating the perspectives of readers. Consequently, this focus on literary production leaves a gap in our understanding of how readers from different cultural backgrounds interpret and engage with literature. Questions such as whether cultural settings influence the understanding of topics, characters, and plots, or if culture shapes reading in certain dimensions but not others, have not been extensively investigated [29]. At the same time, the most frequently studied books and readerships remain those with distinguished popularity, commercial success, social impacts, and scholarly prestige in the Anglophone world, which are subject to historical and social-cultural biases such as classism, sexism, racism, colonialism [2,27,15], and might not be inclusive enough to understand reader response more broadly.

Hu et al. [15] conducted a comparative analysis on the book list and tags across Goodreads and Douban -an international and Chinese platform, respectively -and noticed divergences in readers'collective understanding of classics. Following up on that study, we are interested in exploring what differences there might be in the readers' response to the same stories when they are read in different languages. To this end, we conducted a comparative study of different aspects of reader response, using topic modelling to focus on the aspects that are mentioned in the comments left on two reading platforms publishing the same stories in Chinese and English translation. The research question that we address is: do readers of different languages and cultural backgrounds have a different focus when commenting on a story, such as the plot, the characters, or the setting? Do such reader responses have recognisable features that are specific to one language and culture? The dataset we worked on are from two DSR platforms: Qidian.com (original Chinese web novels) and Webnovel.com (English translation).

Digital Social Reading Platforms for Chinese Stories

Since its birth in the 1990s, Chinese webnovels have grown rapidly to become a new form of Chinese literature. In recent years, Chinese webnovels have become increasingly popular worldwide, emerging as a significant form of participatory transcultural storytelling. They not only have a large number of loyal fan-readers in China, but have also become increasingly popular among international readers by being translated into many languages and circulating in different countries [16]. As defined by Michel Hockx, Chinese online literature is"Chineselanguage writing, either in established literary genres or in innovative literary forms, written especially for publication in an interactive online context and meant to be read on-screen" [14]. The spread of Chinese literature thanks to digital technology provides opportunities for the cultural influence of this literature abroad, but at the same time, his reach dilutes the national attributes of Chinese online literature, loosening the boundaries between different cultural spheres [16].

Some scholars [26,16,22] have pointed out that Chinese online literature has reached such a wide audience due to its foundation in a rich literary tradition spanning classical, modern, and contemporary China, and has inherited and integrated Western fantasy elements and Hollywood narrative techniques while maintaining connectivity with other popular literature and then localizing and recreating them in the Chinese context [16]. The imaginative worlds constructed in online fiction encompass history, military, war, romance, recent actual events, and the future, all expressed in a traditional and secular Chinese style, while at the same time allowing readers from all over the world to find familiar elements in them. This has fostered its connectivity with other successful popular cultures, enabling it to attract readers from other cultural backgrounds while giving them a sense of familiarity with the narrated world. An-other explanation for the success of Chinese novels is that these online novels were created for hedonistic reading, or "popcorn literature"(Shuang). The process of reading such novels is supposed to be light, pleasurable, and exciting. For example, many of these Chinese novels may cover elements of Taoism, the Three Thousand Worlds, etc., but they are not conveyed in a very orthodox or obscure way, but rather in a simple and clear way, without requiring relevant knowledge or a specific cultural background to understand them. Readers from any cultural background can easily get into the story. Following Hollywood movies, Japanese animation, and Korean dramas, Chinese web novels have become the fourth largest cultural phenomenon in the present world [18,16].

Qidian is one of the earliest online reading platforms in mainland China, founded in 2002, with an innovative pay-to-read model and works covering urban, fantasy, romance, science fiction, mystery, sports, games etc., extending to more than 200 genres. Nowadays Qidian is one of the largest online reading platforms with more than 30 million registered readers and more than one million stories.

WebNovel, a platform owned by the same corporation as Qidian, was ofÏcially launched in 2017, and is the first ofÏcial platform for the overseas dissemination of Chinese online literature. It started by providing English translations of stories that were originally published on Qidian, and later also expanded to encourage authors to freely create and upload their own stories on this platform. According to the 2023 China Online Literature Overseas Trend Report [1], WebNovel has launched translations of about 3,600 Chinese stories, with 238 translated works that have been read by more than 10 million people, and 9 that have exceeded 100 million readers. The gender distribution of WebNovel consists of 66% male readers and 34% female readers. The top ten countries in terms of percentage of visits by country/region are the United States, India, the United Kingdom, Indonesia, Thailand, Venezuela, Ghana, Vietnam, and Egypt, with the United States dominating the percentage of visits to the website, accounting for 20.7% by October 2020.

Parallel Corpus

The corpus creation process involved a manual search for translated novels available on Webnovel.com within the NOVEL category, only targeting completed works.1 Subsequently, each identified translated novel was mapped with its original counterpart on Qidian.com. This process yielded a total of 120 novels for inclusion in the corpus. However, the copyright for 10 of these novels on Qidian.com had expired and it was difÏcult to locate comments of these novels on the Chinese platform, rendering them ineligible for inclusion. The final corpus consists of 110 stories [28]. According to WebNovel' s categorization visible on the website interface, these 110 stories consist of 103 Male Lead and 7 Female Lead. We scraped the metadata from each platform and the respective comments and replies for each story, including the user profile of the readers who left the comments. Table 1 has shown the metadata for comments and replies, based on the information provided, we could also map the interactions between comments and their replies. The length of the novels exhibits a wide range, spanning from 288 to 3,588 chapters, with corresponding word counts varying between 1,027,000 and 8,448,900 (measured in Chinese characters). Notably, there is a slight difference in the genres and categories of stories on Qidian and WebNovel.

We also collected the user profiles of readers who has left comments or replies on the stories, as shown in Table 2. Unlike Qidian where the readership is mainly native Chinese speakers or overseas Chineseusing groups, readers on WebNovel do not necessarily come from the same region. We looked into the location distribution of readers in the dataset, using the location available on the users profiles. For the WebNovel dataset, we have the readers from 244 countries, and the top 10 locations with the most users are shown in Table 3. We have also taken into account that book comments left on WebNovel are not necessarily in English, so we looked into the language distribution of comments and replies using automatic language detection.

The results (Table 4) show that English comments accounted for 72.7% of the total replies, and English replies accounted for 68.2% of the total replies. All other languages only account for no more than 2% of the comments/replies each. Given that our research is focused on comparing Chinese and English readerships, we only consider replies and comments written in English.

Methodology

To identify which aspects of the stories Chinese-and English-speaking readers focus on when commenting online, we employed topic modeling. We considered both Latent Dirichlet Allocation (LDA) and embeddings-based modeling. LDA has the advantage of being able to assign several different topics to one document, by generating models with multinomial distribution over topics. However, its efÏcacy in analyzing social media data has been highly criticized [9,24]. Noisy and sparse datasets are unsuitable for LDA [6] due to a lack of enough textual features for statistical learning [4,8]. BERTopic [13] employs a clustering embeddings approach and extends it by incorporating a class-based variant of TF-IDF for creating topic representations. It has been proved that its effectiveness to generate insights from short and unstructured text offers the most potential [8]. However, with BERTopic each document is only assigned to a single topic. Even though topic probabilities can be extracted, they are not equivalent to an actual topic distribution [8], meaning that we may lose the ability to analyze the complexity of each document, especially for longer comments where various aspects of a story might be commented on.

We conducted preliminary evaluation on the topic modeling of comments in both languages, evaluating the performance of LDA and BERTopic with various core configurations and pretrained transformer models.

Data Preparation

Preprocessing is a critical step in ensuring data quality and consistency. Although BERTopic does not require preprocessing of the input text, a study in comparison of LDA and BERTopic [17] shows that topic diversity and coherence is higher in both cases with fully preprocessed text. So we preprocessed our data as well.

We randomly examined a sample of comments and found that there were comments with unusual expressions. For example, "XpXPPXPXPXPXPPXPXPPXPZP", where "xp"refers to the experience points a user can gain to level up and get rewards on both platforms). There are also some random typing (e.g. "F f f f f f. F f g. F t t f. T t. T t t. T t r. R r. R e e w ") and sequences of emojis. We removed such comments with unconventional and inconsistent spelling, as well as punctuation, numbers, and stopwords. We also performed lemmatization and stemming to enhance the coherence of the analysis.

Topic Modeling Evaluation

We first applied LDA to the comments of both platforms separately. Initially, we assessed the 𝐶 𝑉 [23] coherence of the topics by using a range of topic numbers for both language dataset. The results, illustrated in 1, suggest that 30 topics could provide a good balance between the number of topics and their coherence on both dataset. Then we conduct an evaluation including a close reading of the topic words, to better understand the underlying meanings of the topics [8]. The LDA generated topics were quite difÏcult to interpret, so we decided to try BERTopic2 , using the same number of Topics. BERTopic-generated topics were easier to interpret. Moreover, when integrated with the KeyBERT-inspired approach [12], the model performed better in generating overall coherent yet diverse topics, even though it did not achieve the highest 𝐶 𝑉 score (0.49).

Based on this evaluation, we chose to use BERTopic.

Results

The results of the topic modeling (Table 5) suggest that readers who leave comments in both languages pay attention to characterisation, story development, reading experience, but they also leave comments without any specific meaning, just with the intention to gain account experience (WebNovel Topic 5; Qidian Topic 0). Additionally, some readers of both platforms were able to recognise elements of the novels' setting borrowed from other popular cultures (WebNovel Topic 16 and 17; Qidian Topic 2). However, beside these commonalities, the topics also reflect some features that are exclusive to comments in each language. For example, Chinese-speaking readers on Qidian use many formulaic sentences (Qidian Topic 11,17,18,20,22,26), which are unique phrases or sentences that can only be understood in the same cultural context. The literal meaning of these expressions may seem irrelevant to the story text, but their derived meaning can be captured by other readers and invite a discussion. For example, "缝缝补补又三年。 "means "After mending, the worn-out cloth can last for three more years."and it is an expression used to refer to a recurring thing. In the comment, it refers to a similar event/ pattern that keeps repeating in the story. Chinese readers tend to leave comments (Qidian Topic 7) when implicit sexual depictions appear in the text, as a playful way of proving that they had noticed. This may be due to content censorship, as the platform is aimed at an all-ages audience and the terms of service forbid explicit sexual depictions to appear in the text. For Example,"我怀疑你在开车但我没 有证据。 "means "I suspect you are implying a sexual scene but I don't have any evidence."

As for the reader response on WebNovel, we can identify some unique topics, for example, the push for updates of the story (WebNovel Topic 3), and comments left as bookmarks, like "Plan to read"(WebNovel Topic 26). We also observed readers express gratitude towards the translator when they approved the good translation quality of the work (WebNovel Topic 24).

There are also comments related to sensitive themes (WebNovel Topic 27) that did not show up in the Qidian topics. These comments (Appendix A) suggest that the stories in question may include themes or narratives that are racially insensitive or sexist. Readers might find these themes offensive and respond strongly to them. Similarly, when the content of a story does not meet the readers' expectations, readers who are deeply invested in it may have strong negative reactions.

When retrieving topics and corresponding comments, we performed simultaneous keyword searches on the dataset. By manually querying topic-related terms, we identified related comments and found that the actual number of comments belonging to a given topic exceeded those clustered by topic modeling. For instance, WebNovel topic 27 clustered 15 comments annotated as Sensitive/Violence. However, a keyword search for "racist, " which frequently appeared in these comments, yielded 564 entries. Additionally, we observed that topic modeling for both languages indicated -1 (unassigned) as the dominant cluster. This might suggest that many comments may belong to other topics within the clustering. The presence of such a large number of unassigned clusters warrants further investigation.

Conclusion

This study highlights the differences and commonalities in reader responses to Chinese web novels across two platforms: Qidian and WebNovel. The findings indicate that readers of both languages focus on characterization, story development, and the overall reading experience. Additionally, interactions extend beyond the narrative, as readers leave comments to gain account experience, a behavior prompted by the platforms'affordances rather than by a direct engagement with the story. Distinctive patterns also emerged in the comments of both languages. Chinese-speaking readers on Qidian frequently use culture-specific formulaic sentences and comment on implicit sexual scenes. These kinds of comments not directly related to the story create a shared understanding among readers, fostering a collective awareness of the subtleties in the narrative. It builds a sense of community where readers acknowledge the same hidden layers of meaning, and might enhance the collective reading experience. Readers who leave comments in English on WebNovel emphasize pushing for story updates, recommending the story (or not), using comments as bookmarks, and expressing gratitude towards translators in case of high-quality translations. Notably, WebNovel also sees a higher occurrence of sensitive themes, with comments often criticizing perceived racism and sexism in the novels. Overall, these insights into readers' responses underscore the importance of cultural context and platform-specific dynamics in shaping reader interactions.

Limitations and future work

Online platforms offer valuable resources for comparative analysis of how individuals engage with fiction. This corpus provides an important opportunity to observe how readers from diverse cultural backgrounds perceive and interact with fiction in a highly interactive manner.

While this study provides valuable insights into reader response in different language settings, several limitations need to be acknowledged. First, we only performed topic modeling on comments, as comments are more directly related to the story and less likely to deviate from the main topic compared to replies to comments, which are the next level of discussion. However, excluding replies means we may have overlooked a richer spectrum of reader response behaviors. Integrating replies into future analyses could provide a more comprehensive understanding of the dynamics and nuances in reader interactions. Second, when performing topic modeling, we set the same number of topics for both language datasets. This uniform approach does not account for the possibility that the inherent number of topics might differ between the two datasets. As a result, setting a fixed number of topics might lead to the aggregation of distinct topics, thereby limiting the completeness of our exploration of reader responses. For example, great amount of comments was classified to the "Unassigned" the topic. Additionally, the scope of our corpus was largely limited to Male Lead stories -this being the category with more translated stories -which may not fully represent the diversity of online reading across different cultural and social contexts. Lastly, while our analysis aimed to compare reader response across different language settings, it did not explicitly account for cultural differences that might influence these responses. Future studies will focus on refining topic modeling results for a deeper analysis of topics. We plan to incorporate replies into the model to examine the overall topics in reader responses and conduct close reading to the topic samples including the comments that are in unsigned topics. Additionally, we will add more metrics to the modeling, such as focusing on differences for specific stories when analyzing the topics of comments and replies on both platforms. Business reports [1] on Chinese online stories indicate that more Chinese stories are being translated into English. Consequently, we will continue to expand our parallel corpus to include more platforms and cover a broader range of genres and categories, thereby enhancing the generalizability of our findings. In our effort to compare across cultures rather than just languages, we plan to incorporate cultural context into the analysis. This will involve using a native language identifier for English comments and replies combining with user profile to investigate whether and how cultural backgrounds of non-native English speakers shape reader responses. At time, we will take into account the possible impact on the readers' reception of differences between the translated edition and the original novel that may be created in order to improve cultural adaptation.