2025 Volume 8 Issue 1 Pages 9-25
This article examines gendered emotional patterns in Japanese Atomic Bomb Literature through sentiment analysis (SA), challenging the assumption that women’s writing is inherently caregiving. Using the Oseti SA model, the study introduces sentiment protection, a measure of how authors shield their characters from trauma. Findings reveal that male authors exhibit stronger sentiment protection than female writers, reversing traditional gender expectations. Analyzing hibakusha and non-hibakusha authors, the study suggests that sentiment protection functions as a genre-defining feature, with male writers adhering to narrative conventions while female authors demonstrate greater variability. Bridging feminist literary criticism, digital humanities, and reception theory, this study formalizes gendered writing assumptions into computationally testable hypotheses. By integrating algorithmic criticism, this research provides a data-driven perspective on gender, trauma, and genre in Japanese war literature, offering new insights into the emotional structures of narratives depicting nuclear catastrophe.
Critical exploration of literary texts is always based on critical presuppositions. In the analysis of gendered writing styles, one such presupposition is that women’s social roles as caregivers naturally influence their literary expression. However, the examination of emotional patterns through sentiment analysis (SA)—a method that quantifies and formalizes the emotional structure of texts—reveals a more complex reality in Japanese Atomic Bomb Literature. This body of literature responds to one of the most tragic events in Japanese history—the mass annihilation of Hiroshima and Nagasaki—and captures the deepest emotional responses to this catastrophe.
This article applies SA to measure emotions in dialogue lines (characters’ speech) and authorial speech in Atomic Bomb fiction, revealing unexpected gender dynamics. Contrary to the stereotype that women’s writing is inherently caregiving and protective, the analysis demonstrates that male authors—who are conventionally perceived as emotionally detached—exhibit stronger sentiment protection of their characters. This characteristic, defined by how authors shield their characters from trauma, is typically associated with feminine writing but is shown to be more prevalent in male-authored texts in the current study.
This study introduces the concept of sentiment protection and quantifies it as the difference between dialogue and author’s speech sentiment scores. Using the Oseti model for Japanese-language SA, the study identifies a gender interplay in Atomic Bomb Literature, whereby traditional gender expectations are reversed: Women do not necessarily act as caregivers, breaking the social enforcement of such emotional expressions, and male authors frequently adopt protective emotional strategies.
The findings reveal a nuanced gendered interplay in Atomic Bomb Literature. While female authors, particularly those central to the first wave, often subverted caregiving roles, their writings displayed greater emotional variability and complexity. In contrast, male authors adhered to consistent patterns of sentiment protection, reflecting genre conventions and canonical expectations. By bridging feminist literary criticism, computational literary studies, and reception theory, this article provides a novel perspective on how gender, trauma, and genre intersect in one of the most morally significant literary traditions in modern history.
Japanese Atomic Bomb Literature (genbaku bungaku) is a body of works that captures the harrowing experiences associated with the Hiroshima and Nagasaki bombings. It spans various genres, including memoirs, fiction, poetry, and essays, all of which grapple with themes of trauma, survival, memory, and the ethical dimensions of nuclear warfare. It emerged as an immediate response to the tragedy, beginning in 1945 when Ota Yoko (1945), a survivor, published the short story “Like Undersea Light” in The Asahi Shimbun. This literary tradition continued to evolve as survivors, witnesses, and later generations reflected on the impact of the bombings, and it continues to the present day.
Many literary scholars (e.g., Kuroko 1993; Treat 1995) have studied this genre in Japan and abroad, and it is difficult to introduce Atomic Bomb Literature in a brief summary. However, one crucial aspect is its gendered demarcation. Researchers have defined three distinct waves of Atomic Bomb Literature, the first of which was dominated by female authors.
This first wave, which initiated the major literary movement of genbaku bungaku, was shaped primarily by female writers who survived the tragedy and shared their pain with society. Authors such as Hayashi Kyoko, Ota Yoko, and Takenishi Hiroko played a foundational role in establishing the genre. Their works led scholars to argue that female authors focus on more intimate, personal experiences, engaging with their characters empathetically and portraying individual suffering overtly1. In contrast, male authors tend to observe the tragedy from a more detached perspective, analyzing its broader social impact.
This framing aligns with literary stereotypes (understood here not as negative biases, but as explanatory models) proposed in literary theory. While women’s writing has often been associated with nurturing and emotional depth, and men’s writing with rational analysis, this study, through SA, examines how these assumptions hold up in practice within Atomic Bomb Literature.
Feminist literary criticism has proposed various rationales for distinguishing women’s and men’s experiences, drawing on social roles and performed identities (Gilbert and Gubar 1985; Showalter 1977), bodily experiences and differences (Cixous 1976), and early psychological development (Kristeva 1982, 1984). At the intersection of these approaches, a recurring emphasis emerges on women’s caregiving nature and their maternity roles. Some thinkers, such as Chodorow (1979), Ruddick (1995), Rich (1995), and Gilligan (1993), have highlighted maternity as a central aspect of women’s identities, influencing their actions and fostering a tendency toward caregiving behaviors.
Extrapolating ideas of caregiving as women’s identity to the process of women’s creativity, this caregiving disposition extends into literary texts, which inherently possess an agent-object structure compatible with the dynamics of care. Just as women, acting as agents, demonstrate parental care to those around them, female writers transfer this caregiving impulse into their works. They imbue their narratives with care, extending it to the objects of the fictional microcosm—the characters. Consequently, women’s fiction often exhibits a higher degree of caregiving than male-authored texts.
The traditional literary criticism of Atomic Bomb Literature provides observations that align with these gendered stereotypes. Treat (1995) describes the writings of Hayashi Kyoko and Ota Yoko and concludes that they delved into experiences of women and children in their fiction, focusing on their suffering and resilience and the physical and psychological scars borne by female survivors. Meanwhile, he notes that male authors’ prose was directed from this intimacy to a more global comprehension of this trauma by Japanese society. In particular, Hara Tamiki (1983) in his Summer Flowers reflects on the fragility of human existence and the search for meaning in the face of overwhelming devastation, and Oe Kenzaburo (1983), another author but not a witness, reflected on Atomic Bombing and the ethical and societal implications of nuclear warfare, often through a more detached and analytical lens.
The nature of the experiences depicted in Atomic Bomb Literature inherently directs analysis toward the emotional perspective of these works. In mainstream Western academic literary criticism, emotional response and effect have traditionally been excluded as distracting elements. The concept of the “affective fallacy,” introduced by Wimsatt and Beardsley (1949), became a foundational principle of New Criticism. According to this view, evaluating literary texts based on the emotional effects they produce in readers is a mistake. New Criticism rejected the role of emotional response for several reasons: It was believed to undermine objectivity owing to the subjective reactions of individual readers, it shifted analytical focus away from the text and toward an abstract reader, it detracted from the formal qualities of the text, and it often led to shallow or indulgent readings, fostering impressionistic rather than argumentative criticism (Wimsatt and Beardsley 1949). Although 20th-century reader-response theories—such as the “aesthetics of reception” developed by Wolfgang Iser (1980) and Hans Robert Jauss (1970, 1982), or the works of Stanley Fish—began to incorporate emotional response into literary analysis, emotion remained marginalized in literary studies until the 1990s and early 21st century.2
The empowerment of emotions in literary studies and the humanities more broadly was marked by what Patricia Ticineto Clough and Jean Halley (2007) have termed the “affective turn”—a shift that brought emotions, affects, and intensities into the center of theoretical inquiry, challenging the long-standing division between feeling and critique. Within this emerging field, Suzanne Keen (2006) contributed significantly by theorizing narrative empathy, which she defines as “a vicarious, spontaneous sharing of affect” that may be provoked by “witnessing another’s emotional state, by hearing about another’s condition, or even by reading” (208). Her work emphasizes how narrative form and focalization can guide readers’ emotional responses, allowing fiction to elicit empathy through structured representations of characters, narrators, and situations.
Earlier strands of Literary Affect Theory, as developed by scholars such as Eve Kosofsky Sedgwick and others, also contributed to this revaluation of emotion. Rather than treating feeling as an interior psychological state, these theorists conceive of affect as a relational and atmospheric phenomenon, shaped by context, repetition, and surface texture. In particular, Sedgwick (2003) emphasizes the periperformative and textural nature of affect, suggesting that emotion may not lie beneath the text but instead inhabit its surface, expressed through tone, rhythm, and narrative stance.
In trauma studies, emotion is both a content and a structural challenge. Shoshana Felman and Dori Laub (1992) argue that trauma resists straightforward narration and that testimony is not simply the communication of knowledge, but a relational act of giving witness to an experience that exceeds conventional representation. Testimony thus involves not only narration but ethical modulation—a balancing of exposure and protection that preserves the integrity of the traumatic voice.
Finally, Stephanie Shields has shown that emotion is not biologically fixed but is socially constructed and performed, particularly along gendered lines. She argues that “‘performance’ of gender creates gendered emotion; ‘practice’ of emotion creates gender and reinforces the notion of gender as difference” (Shields 2002, 100). Within this framework, the expression, suppression, or displacement of emotion becomes a strategy of doing gender—enacting and reinforcing cultural norms through affective behavior.
Japanese Atomic Bomb Literature, which centers on extreme, traumatic experiences, is a case in point. These texts inevitably convey intense negative emotions—fear, despair, grief—arising from the catastrophic events they depict. Fictional texts, however, operate on two interconnected narrative levels: the characters’ dialogic world, where emotional responses unfold through speech and interaction, and the narrative “superlanguage” of the author-narrator, who shapes the fictional world and modulates its emotional atmosphere. It is within this narrative architecture that authors may engage in sentiment protection, absorbing emotional intensity into the narrator’s voice to shield their characters from the direct expression of trauma.
In Atomic Bomb Literature, characters express their reactions to the trauma and its aftermath primarily through dialogue, often laden with negative emotions. Meanwhile, the author-narrator either experiences or observes these traumas and channels their own affective responses—such as fear, desperation, and sorrow—into the narrative. This raises the question: Where does the strategy of care manifest in such narratives? If care is understood as the act of giving something valuable without taking it, then in this context, care is expressed through the handling of emotions. Authors have the power to manage both layers of the text: They can either let characters directly voice the burden of negative emotions or absorb and redirect these emotions into their own narrative voice. By choosing the latter, authors relieve their characters of the weight of expressing inevitable trauma-related negativity. This intentional redirection of negative emotions into the author’s speech, sparing the characters, is defined in this article as sentiment protection.
At this point, the concept of sentiment protection aligns closely with the broader framework of the affective turn. From the perspective of Keen’s (2006) theory of narrative empathy, sentiment protection functions as an act of managing and redirecting empathy within the narrative. Within the domain of literary affect theory, sentiment protection—especially when quantified through SA—helps to clarify the distribution of affect across textual layers, revealing the text’s affective architecture.
As Felman and Laub (1992) has argued in trauma studies, the act of testimony entails both narrating and protecting—that is, speaking trauma while mediating its exposure. Similarly, sentiment protection can be seen as a narrative strategy of emotional mediation.
All of these approaches share an affinity with reader-response theory, in that they treat emotion not as a static textual property but as something co-constructed by the interaction between author, text, and reader. The proposed concept of sentiment protection, then, can be readily integrated into analyses grounded in the affective turn. At the same time, it introduces an additional layer: treating emotion as an intertextual property by analyzing the distribution of affect between the character-level narrative world and the author’s “metaworld” that frames and governs it.
SA algorithms help to formalize this concept by offering a method for quantifying emotional patterns in narrative texts. In doing so, they also offer a way to partially overcome the limitations of the “affective fallacy,” insofar as their formalization lends epistemological structure to the study of literary emotion.
SA broadly refers to tools and methodologies designed to analyze “people’s opinions, sentiments, appraisals, attitudes, and emotions towards entities and their attributes” (Liu 2020, 1). It evaluates sentiments, either the effect of a text on a reader or the sentiment conveyed by the writer. This evaluation is typically expressed in polar or categorical terms (e.g., negative, positive, neutral; upset, happy, frustrated) or numerical values (e.g., from −1 to +1, where −1 represents the lowest sentiment score and +1 the highest).
Initially developed primarily for business analytics and marketing purposes, SA has found significant applications in computational literary studies. Observing the trends in SA, Öhman et al. (2024) state that previous research has demonstrated the potential for SA to uncover insights into reading experiences and interpretations (Drobot 2013; Cambria et al. 2017; Kim and Klinger 2019; Brooke et al. 2015; Jockers 2023; Reagan et al. 2016). Emotional arcs—graphical representations of emotional shifts within a text—have been employed in analyses of literary genres (Kim et al. 2017), plot archetypes (Reagan et al. 2016), dynamic properties of narratives (Hu et al. 2021), narrative mood (Öhman and Rossi 2023), and correlations with reader preferences and perceived literary quality (Bizzoni et al. 2022; Öhman et al. 2024).
However, most SA tools and methodologies have been developed for use in English and other European languages, and resources for analyzing Japanese texts are limited. To address this gap, researchers have developed customized algorithms for Japanese SA, including textual vector representation models (Zhou et al. 2021), transformer-based approaches (Ikeagami et al. 2022), and dictionary-based solutions (Ikegami, n.d.). While transformer- and dictionary-based methods provide efficient solutions for quantifying emotional representations, they each have significant limitations.
SA approaches can be broadly divided into two categories: dictionary-based (rule-based) and machine learning–based ones.3 Dictionary-based approaches segment an input text into tokens, lemmatize them using morphological analyzers (for languages such as Japanese that have a complex morphology), and match them with sentiment dictionaries containing words tagged as positive or negative. The simplest models compute sentiment scores by summing the counts of positive and negative tokens. More advanced rule-based models incorporate syntactic and semantic structures, accounting for sentiment modifiers like negations (e.g., “That was a happy story” vs. “That was not a happy story”) and intensity shifts (e.g., “That story was not so happy”). These enhancements enable dictionary-based approaches to handle more complex emotional expressions.
While dictionary-based approaches are transparent and interpretable, their performance heavily depends on the quality of the sentiment lexicon and the rules governing sentiment calculation. Scaling such models requires significant intellectual resources, and the models often struggle with nuanced language or context-dependent emotional expressions.4
Transformer-based models, a class of deep learning architectures, operate differently by using mechanisms such as attention to weigh the importance of different text segments. These models are trained on large datasets of labeled text, with each chunk being associated with a sentiment score or category. Through iterative parameter adjustments, transformers “learn” to infer sentiment based on patterns in the data, much like a human annotator relying on prior experience and context (Mishev et al. 2020).
Transformers excel in capturing complex relationships and nuances in text, offering higher accuracy and adaptability. However, their black-box nature makes it difficult to trace the decision-making process or understand specific steps in the analysis. Additionally, their performance is highly dependent on the quality and relevance of the training data. Owing to both epistemological and quality-related limitations, existing transformer-based SA models for Japanese are of limited applicability to the analysis of the Atomic Bomb Literature corpus.
While transformers can be transformative for languages with limited natural language processing resources, such as Japanese, they are often unsuitable for literary analysis due to the nature of their training data. On Hugging Face, a repository for machine learning models, only 21 transformer-based SA models are available for Japanese, compared with significantly more for English. Among the Japanese models, only four provide any documentation about training data or model quality (Table 1). These models were often fine-tuned on relatively small datasets, typically drawn from nonliterary sources such as product reviews or business reports. Moreover, two models are, in practice, unapplicable for literary analysis. Specifically, the Japanese Stock Comment Sentiment Model does not output conventional “positive-negative-neutral” sentiment labels and the weights of model’s assurance in them but rather “bullish” and “bearish,” characterizing the emotions of trading patterns. In addition, Finance-sentiment-ja-base was discovered to output predominantly neutral sentiment scores, which is abnormal behavior for the dataset.5
The sentiment of literary texts, especially those dealing with nuanced topics like trauma, is inherently more complex than straightforward emotional expressions found in commercial texts. As such, the black-box behavior of transformer models poses challenges when applied to literary corpora such as Atomic Bomb Literature.
Among dictionary-based tools, Oseti is a notable model for Japanese SA. Developed by Ikegami Yukino (n.d.) in 2021, Oseti is based on the Dictionary of Japanese Language Evaluation (nihongo hyouka kyokusei jisho) (Inui-Okazaki Laboratory, 2008; Higashiyama et al. 2008; Kobayashi et al. 2008). The model processes Japanese text by splitting it into sentences using the Bunkai library (Hayashibe and Mitsuzawa 2020) and then tokenizing it into lemmas via the MeCab morphological analyzer (Kudo n.d.). Sentiment scores are calculated by summing positively and negatively tagged words and normalizing the results to a scale between −1 and +1. Oseti also accounts for some syntactic features, such as negations and parallel constructions.
Oseti is a minimalist tool for SA that operates using a sentiment dictionary and a small set of basic rules. While it is undoubtedly limited in its ability to capture the complex organization of sentiment in natural language, its transparency and traceability—even at the level of the Python code in which it is implemented—position it in direct contrast to black-box transformer-based models. In comparative testing against the well-established and academically recognized VADER, as well as two transformer models (excluding the Japanese Stock Comment Sentiment Model and Finance-sentiment-ja-base owing to their unconventional sentiment categorization and abnormal performance), Oseti performed comparably to transformer models.6 In particular, Oseti’s precision was on par with that of bert-finetuned-japanese-sentiment, and it significantly outperformed japanese-sentiment-analysis (Table 2). Moreover, Oseti’s recall—which reflects the proportion of true sentiment values captured in comparison with the annotated dataset—was 46% closer to the ground truth than the output of bert-finetuned-japanese-sentiment. Interestingly, VADER also outperformed the Japanese transformer models, despite operating on machine-translated Japanese text.
Although Oseti is far from perfect, its explicit and transparent mechanisms make it a reliable compromise in cases in which transformer models lack adequate performance or interpretability. While primarily used in business and marketing studies, Oseti provides a clear framework for analyzing sentiment in literary texts, albeit with limitations.
For this study, the Oseti model was modified to replace the MeCab tokenizer with the neologd tokenizer (Sato et. al. 2017), which offers a more extensive dictionary and improved performance with modern lexis. Images were converted to text (OCRed), using ABBYY FineReader 15 from the 15-Volume Japanese Atomic Bomb Literature Anthology (excluding journalism and non-fiction works in Volumes 14 and 15) (Signatories of the Statement … 1983). The OCRed texts were then manually reviewed to correct errors, followed by automated corrections using a script.7
Dialogue lines representing direct speech were extracted using regular expressions, while the remaining text was classified as authorial speech.8 Sentiment scores were calculated for entire texts, direct speech, and authorial speech, with overall scores derived as the mean of all sentence-level sentiment scores.
SA reveals that the shielding of characters from trauma—expressed via negative emotions—and the burden of reflecting and comprehending this trauma in the author’s speech challenge stereotypical expectations of women authors as caregivers.
General Sentiment Trends in Atomic Bomb LiteratureThe overall sentiment of the texts, reflecting their general tone and mood, underscores the pervasive negativity in Atomic Bomb depictions. Given the horror of the bombings and the lasting scars of these events, it is unsurprising that both male and female authors predominantly expressed negative emotions in their writings. The average and median sentiment scores for all Atomic Bomb fiction texts (N = 106) were −0.06 and −0.08, respectively.
However, gendered differences are evident. Female authors, often focusing on personalized experiences of women and children—groups traditionally portrayed as more vulnerable to catastrophic events—exhibited a bleaker perspective in their writings. The average and median sentiment scores for women’s texts were −0.1 and −0.07, respectively, compared with −0.06 and −0.06 for men’s works. While women’s writings generally presented a more somber and dramatic view of Atomic Bomb experiences, they did not entirely exclude the possibility of creating a caregiving environment for their characters.
Sentiment Protection in the WildThe concept of sentiment protection—or sentiment difference, if we describe this stylistic feature as the difference between sentiment values in direct speech versus the authorial narration—appears clear in terms of operationalization, but remains elusive when we consider how emotional organization is experienced while reading. Before analyzing general tendencies in sentiment protection through distant or quantitative means, let us first approach it through a closer, more experiential reading of selected texts.
Four texts by canonical Atomic Bomb Literature authors, Sata Ineko, Hayashi Kyoko, Hironaka Toshio, Kokubo Hitoshi, can be used to illustrate how sentiment protection can be interpreted through live, close reading. These examples also demonstrate that statistical patterns are not definitive, and that no gender group of authors can be said to fully determine the emotional mode of writing.
Sata Ineko’s “Today’s Story” explores memory, trauma, and the long-term psychological toll of the atomic bombing. The narrative follows a young couple, Murai and Fujiko. Initially, Murai appears calm and affectionate, and the couple’s relationship seems idyllic. However, Murai gradually becomes distant and emotionally detached, and Fujiko is left confused and hurt, unable to understand the shift. Only later does she learn that Murai has been affected by radiation and cannot have children. The story ends not in confrontation, but in quiet sorrow as love succumbs to trauma.
Although the text only subtly hints at the negative tone of dialogue, Murai’s lines become increasingly curt, signaling his emotional withdrawal. Sata’s narrative descriptions, which are comparatively light or cheerful, heighten this emotional dissonance. Fujiko’s cheerfulness, conveyed more through narrative remarks than direct speech, is in contrast to Murai’s silence, emphasizing her solitude. This results in a negative sentiment difference that reflects a protective withdrawal from direct emotional confrontation.
Hayashi Kyoko’s “The Road” is a first-person narrative about a woman returning to Nagasaki in 1975, thirty years after the bombing. She seeks to uncover the fates of three female teachers (N, T, and K) who oversaw her during wartime labor mobilization. She visits graves, interviews people, reads lists of the dead, and relives the trauma through memory and places.
The narration is laden with trauma and memory, yet the boundary between narration and dialogue is porous. Direct speech does not shy away from grief or emotional pain. For example, in a conversation with Professor Tanaka, the narrator recalls cremation scenes using oil and other vivid details. Yet, despite the intensity of these exchanges, the sentiment difference remains modest. Even as the narrative confronts past suffering, Hayashi avoids turning the narrator into a vessel for prolonged introspection. This restraint limits the emotional layering typical of sentiment protection.
Hironaka Toshio’s The Day of Fire represents a pattern of sentiment protection characteristic of a male author. The story is narrated from the perspective of Takashi, a direct witness of the Hiroshima bombing. He sees the flash of the explosion—“as if a large amount of magnesium had been suddenly ignited”—describes the ruined buildings, suffers injuries, witnesses the death of his aunt, encounters people with charred skin and corpses, and fails to save a girl burning in agony (Hironaka 1983, 63). Even though the narrative unfolds through a first-person perspective, blurring the line between internal monologue and authorial voice, the protagonist takes on the emotional burden. His inner voice meticulously records the horror, while his outward speech, directed at fellow victims, appears more composed—almost cheerful—as if conforming to an expected emotional posture of the hibakusha survivor.
Kokubo Hitoshi’s “The Mark of Summer” is another example of I-narration, but one that exhibits negative sentiment protection. The story recalls “The Road” in plot structure: The narrator, a schoolteacher based in Tokyo, returns to his hometown of Hiroshima upon learning of the death of his classmate, Shirai. As he reconnects with other old friends—Inagaki and Maki—they reminisce about their school days and the aftermath of the atomic bombing. Through these conversations and personal reflections, the narrator reconstructs Shirai’s way of life—a life shaped by humility, restraint, and a deliberate refusal to seek comfort or material success. Despite offers of help, Shirai insisted on living minimally. His death—and in particular the smile on his face at the moment of passing—leaves a deep and unsettling impression on those who knew him. The story closes with the narrator reflecting on a possible return to Hiroshima, along with the unresolved moral burden of having survived the bomb.
The narrator’s remarks are saturated with wartime memory and historical reflection, but unlike “The Road,” these reflections spill into explicit dialogues that often override the internal monologues. The narrator’s encounters with others who share his experiences do not merely function as an act of testimony or bearing witness. Rather, they become a mode of collaboratively reconstructing memory and actively debating what it morally means to survive. In contrast to “The Road,” in which direct speech tends to merge into a collective tone, the dialogues in “The Mark of Summer” remain individualized, posing ethical questions and often attempting to answer them. Here, explicit dialogues drive the narrative, overshadowing inner reflections. As a result, the author has less structural ground for sentiment protection: Traumatic experiences are not compartmentalized into authorial descriptions or monologues but are instead negotiated in conversation. Characters such as Inagaki and Maki speak openly about death, expressing a desire not to die as a hibakusha, which brings a negative emotional register into the dialogues. Still, even in these moments, SA reveals that the contrast between emotional valence in narrative versus speech is lower than the average across the corpus.
Sentiment protection—or its absence—may manifest in various narrative forms: as emotional distraction or passive aggression in “Today’s Story”; as interiorized trauma wrapped in the narrator’s monologue in “The Road”; or, as in The Day of Fire, through a split between an emotionally burdened inner voice and a composed outward one. In “The Mark of Summer,” dialogue breaks through the inner walls of trauma and drives the story forward. In each case, emotional organization shapes the text in distinct ways—but taken together, these patterns expose boundaries between subgenres within Atomic Bomb Literature.
Gender and Sentiment ProtectionThe analysis suggests that the traditionally feminine trait of protectiveness and caregiving, often associated with female authorship, is not a defining characteristic of women’s writings in Atomic Bomb Literature (Table 4). Although more than half (55.6%) of women’s works demonstrated sentiment protection—shielding characters from the most traumatic emotions—male authors displayed this trait more frequently, with 72.9% of their texts exhibiting sentiment protection.
Moreover, the intensity of sentiment protection, measured as the contrast between the nonnegativity of emotions in dialogue and the negativity in the author’s speech, was higher in male-authored texts. This finding indicates a more intense expression of care for characters in the fictional worlds of male writers, despite the deep embedding of these narratives within the harsh realities of the survivors’ tragedies.
These findings reveal a gender interplay in Atomic Bomb Literature: Male authors demonstrated a writing pattern considered traditionally feminine, while female authors often rejected such caregiving roles. This counter-feminine, unprotective attitude among female writers might be understood as a response to the extreme trauma of the atomic bombings—an experience so severe that it potentially depleted moral resources for gender-prescribed maternal care. Instead, female narrators seemed to seek support and protection themselves, with their characters becoming vehicles for the authors’ pain. However, the experiences of hibakusha authors challenge this interpretation.
Hibakusha Writers’ Sentiment PatternsTreat (1995, 8) aptly observed that the effects of the bomb were not limited to the residents of Hiroshima and Nagasaki but extended to the entire society. Among the authors in this study, hibakusha writers—those who experienced the explosions firsthand and endured radiation sickness—are particularly significant. If anyone bore the ultimate trauma, it was this group, and one might expect their trauma to blur gendered differences in sentiment protection.
Among the 106 texts in the corpus, 45 were authored by hibakusha, including Hayashi Kyoko, Ota Yoko, Takenishi Hiroko, Hara Tamiki, and Inoue Mitsuhara. Of these, 25 texts were written by the three female authors and 20 by the two male authors. Interestingly, hibakusha women demonstrated less sentiment protection than female writers overall, possibly reflecting the depth of their trauma (Table 5). In contrast, male hibakusha authors were more protective and caring than their non-hibakusha counterparts.
This suggests that gendered sentiment protection in Atomic Bomb Literature may not be dictated solely by the extremity of experiences but rather by systemic and canonical factors within the genre.
Sentiment and CanonThe pronounced use of sentiment protection by male authors may reflect not a collective psychology but rather a stylistic feature of Atomic Bomb Literature. This view aligns with Hans R. Jauss’s “horizon of expectation,” wherein a text predisposes its readers to specific types of reception through familiar textual strategies and genre conventions:
A literary work, even if it seems new, does not appear as something absolutely new in an informational vacuum, but predisposes its readers to a very definite type of reception by textual strategies, overt and covert signals, familiar characteristics or implicit allusions. It awakens memories of the familiar, stirs particular emotions in the reader, and with its “beginning” arouses expectations for the “middle and end,” which can then be continued intact, changed, re-oriented or even ironically fulfilled in the course of reading according to certain rules of the genre or type of text (Jauss 1982, 12).
Atomic Bomb Literature was not written for immediate success but as an act of confession, aiming to narrate and share trauma. Yet, like any genre, its canon was shaped by critics and editors who defined its major authors and works. Despite the genre’s origins being shaped significantly by women, the selection of canonical texts in the anthology—70 works by men versus 36 by women—reveals a male-dominated tradition (Signatories of the Statement … 1983).
The consistency of sentiment protection among male authors (Figure 1) supports the idea that it is a genre-constituting feature. Conversely, the variability in sentiment protection among female authors (Figure 2) indicates experimentation and a search for the “proper” mode of representing the complexities of trauma.
Figure 1 demonstrates the distribution of sentiment difference in works by male writers. As evident from the figure, male writers not only tend to display greater sentiment protection of their characters compared with female writers, but their sentiment difference scores also exhibit a more normal distribution (skewness measure = 0.26). This finding suggests a consistent stylistic approach and adherence to genre conventions, as though male writers are following an implicit pattern regarding the extent to which characters should be isolated from bomb-related trauma, expressed through their emotional responses.
In contrast, female writers, who still contributed a significant part of Atomic Bomb Literature, demonstrate greater variability in their use of sentiment protection to care for characters. While male prose tends to follow an unspoken pattern with lower variability, gravitating toward the mean axis, the distribution of sentiment difference in women’s texts is asymmetrical (skewness = −0.85), with notable outliers in positive sentiment protection and significant instances in which a caring strategy toward characters is absent, as illustrated in Figure 2. This variability suggests that female writers did not adhere to the conventional core of the canon, which favors slightly positive sentiment protection with limited variation. Instead, they experimented or searched for a “proper” mode of expressing the complex emotional dynamics of Atomic Bomb trauma across the planes of authorial and character expression. This finding could also indicate a counter-canon approach in their writing: a deliberate focus on expression rather than adherence to genre expectations, as reflected in the diverse variations of sentiment protection.
The concept of sentiment protection can be interpreted as reflecting an author’s attitude toward trauma management within gendered expectations. It functions adequately when two prerequisites hold: (1) Sentiment scores can be meaningfully mapped onto textual phenomena, and (2) SA can reliably extract those scores.
However, as famously summarized by Box and Draper (1987, 427) for statistical models, “all models are wrong, but some are useful.” As was previously mentioned in SA methodological critics, SA does not completely satisfy the criterion on the textual reference; that is, it measures some abstractions but not the text’s reality (Bowers and Dombrowski 2011), and it does not demonstrate the consistency in sentiment score outcomes shown in the current study and discussed by Kim (2022). Oseti is not the perfect model, moreover, which reproduces reduced and simplified textual reality. Nevertheless, being wrong but its design as well as transformer model, it resulted to be the most proper instrument for sentiment analysis in Atomic Bomb Literature.
Sentiment Difference and Statistical SignificanceThe numerical results obtained via SA led to interpretive claims that challenge gender-essentialist assumptions about emotional modes of writing in Atomic Bomb Literature. These claims must, however, be viewed through the lens of statistical significance, which poses challenges in this context.9
Formally, the research question was reduced to a null hypothesis (no gender difference in sentiment distribution) and an alternative hypothesis (female authors display sentiment protection more frequently than male authors). As Table 6 demonstrates, none of the tests return p-values below the conventional threshold of 0.05 (or 0.01 in high-accuracy domains like medicine).10
The corpus, consisting of 116 works, is compact and cannot be significantly expanded. This limitation hampers statistical robustness, especially since most inferential tests are designed for larger and more representative datasets (assuming it is correct to reduce literature to just a dataset). While continuous sentiment scores could not yield statistically significant distinctions (see Notebook 4), a binary classification approach (presence/absence of sentiment protection) yielded a difference at approximately the 10% level. Although not statistically robust in a conventional sense, such a signal is nonetheless meaningful within the interpretive practices of literary studies, in which style is not fixed in discrete categories.
I-Narration and DialogueBeyond technical limitations, Atomic Bomb Literature presents a narrative peculiarity: the frequent use of inner dialogue as a literary device. While SA frameworks rely on the clear typographic distinction between dialogue and authorial text, inner dialogue complicates this binary. Formally marked as author’s speech, inner dialogue can function in two ways: as an extension of the author’s voice or as a projection of the character’s interiority. In the latter case, the boundary between a narrator and a character becomes blurred.
This ambiguity renders annotation difficult. Whether to distinguish between true authorial voice and inner dialogue often depends on critical interpretation. The ideal solution would be to manually annotate the corpus according to Text Encoding Initiative standards, clearly distinguishing these narrative layers. Alternatively, this task might be delegated to large language models (LLMs) trained on genre-specific textual distinctions. Yet the first option is resource-intensive, and the second remains, for now, unreliable with general-purpose LLMs. However, with the rapid development of LLMs, claiming that this task cannot be delegated to properly prompted or fine-tuned models sounds shallow and calls for experiments to determine whether LLMs can, in fact, be entrusted with this task.
In his exploration of algorithmic quantitative practices for literary criticism, Steven Ramsay (2011, 16–17) noted that “readings of texts can be arrived at algorithmically, but simply that algorithmic transformation can provide the alternative visions that give rise to such readings,” elevating the researcher to “a different scale and with expanded powers of observation.” This study does not seek to overshadow the rich body of criticism on Atomic Bomb Literature—an absolutely unique genre in its moral depth and its capacity to engage readers—but rather aims to offer an alternative vision. Through this lens, it challenges the gendered stereotypes often associated with this literature, revealing that female authors do not pursue the expected feminine caregiving patterns any more than male authors do. By employing SA as a tool that enables Ramsay’s “alternative visions,” this work uncovers nuanced patterns of sentiment protection, thereby challenging traditional assumptions about gendered writing.
The SA perspective demonstrated that female authors, traditionally associated with caregiving and emotional protection, adopted a more complex and variable approach to sentiment protection. While over half of their works showed sentiment protection, this variability suggests a departure from strict gender-prescribed maternal care, favoring experimentation with emotional expression. Male authors, by contrast, exhibited consistent sentiment protection in their works, with a higher percentage of texts shielding characters from the traumatic emotions of the bombings. This consistency aligns with canonical expectations and genre conventions, demonstrating a pattern often considered feminine.
Hibakusha writers—those who narrated first-hand experiences of the atomic tragedy and who are often seen as having ultimate, gender-independent experiences—exhibited the same pattern of gender interplay in sentiment protection, and in some cases, intensified it. Female hibakusha authors demonstrated lower levels of sentiment protection compared with non-hibakusha women, likely reflecting the profound depth of their trauma. In contrast, male hibakusha authors were even more protective and caring in their writings than non-hibakusha men, further complicating the assumed relationship between trauma and emotional shielding.
Sentiment protection, as a textual feature, operated de facto as a genre-constructing element. Male authors adhered to the unwritten conventions of the genre, maintaining a slightly positive sentiment protection with limited deviation to either extreme—neither overshielding their characters from the aftermath of the atomic bombings nor being entirely indifferent to their trauma. Female authors, with greater variability, appeared to experiment with emotional representation, possibly forging a counter-canon focused on self-expression rather than strictly conforming to reader expectations, a pattern more consistent in male-authored works.
More broadly, this study offers a methodological framework for analyzing gender narratives, combining computational literary studies with feminist literary criticism. It provides a means to test a widely held assumption about women’s writing—that caregiving, derived from the intersection of women’s bodily nature and social roles, is a distinguishing feature—by formalizing these theoretical ideas into computationally testable hypotheses.
By bridging feminist literary criticism, genre studies, and computational methods, this article offers a fresh perspective on how gender, trauma, and genre intersect in literary narratives, paving the way for new interdisciplinary research avenues.
The repository with the script for SA (including the modified for neologd morphological analyzer Oseti library) and results of SA can be found by the link https://github.com/susurofu/Genbaku-Bungaku-SA.
The scripts for data analysis and all resulted data are publicly available via the repository. The textual data is not available due to copyright restrictions. ( link here ）
However, this statement does not reveal explicitly and refers not to generalized conclusions but to the critical observations of women’s and men’s authors. See Treat’s (1995, 29) interpretation of Goto Minako’s self-analysis of writing intentions and compare Minear’s (1990) critical introductions to Hara Tamiki’s and Ota Yoko’s works.
Stanley Fish indirectly incorporates emotion into his theory of affective stylistics, which emphasizes the “structure of experience,” including both cognitive and emotional processes (Fish 1980, 92). In Hans Robert Jauss’s aesthetics of reception (Rezeptionsästhetik), emotion is a central part of the reader’s aesthetic response to the text and, like the response itself, is historically conditioned (Jauss 1970, 12).
This categorization is less technically precise but sufficiently simple. Other works propose more complex classifications. For example, Kim and Klinger (2019) offer a three-fold classification: dictionary-based, feature-based machine learning, and representation-learning/deep learning methods. Rebora (2023) suggests a typology-constructor that considers criteria based on the emotion theories embedded in sentiment analysis, the emotion resources used, and the computational methods applied.
There are, however, tools that work with word embeddings, such as SentiArt (Jacobs 2019). Unlike traditional approaches based on predefined rules or dictionaries, SentiArt utilizes word2vec embeddings—vector representations of text trained via a two-layer neural network. Rather than calculating absolute sentiment values, SentiArt measures the semantic distance between word vectors and emotional anchor words (e.g., “happy,” “fearful,” or “neutral”). These distances can then be operationalized as proxies for sentiment.
As Rebora (2023) notes, one of SentiArt’s key advantages is that it covers nearly 100% of text tokens, since almost all words in the embedding space have vector representations—compared with just 10%–20% coverage in dictionary-based models. However, word2vec embeddings are corpus dependent (Mikolov et al. 2013, 6), meaning their effectiveness can vary based on the training data used.
Although SentiArt found a place within cognitive poetics, its impact was limited by the rapid development of transformer models, which offer more sophisticated contextual embeddings. As of April 2025, Google Scholar shows that SentiArt (Jacobs 2019) has received 35 citations, compared with 120 for Syuzhet (Jockers 2023), 319 for VADER (Hutto and Gilbert 2014), and 332 for one of the transformer-based models (Naseem et al. 2020).
See Notebook 3 in the repository.
The code for the calculations can be found in Notebook 3. For the sample evaluation dataset, 50 direct speech sentences and 50 author’s speech sentences were randomly selected and manually annotated with three sentiment labels: positive (1), neutral (0), and negative (−1). Although some studies perform numerical annotation of sentiment (e.g., Bizzoni and Feldkamp 2024), owing to the limited number of annotators ( N = 1) and the differences in calculating sentiment scores between rule-based and transformer models, only the quality of predicting the sentiment category was measured, as this is the minimal necessary condition for characterizing the quality of the models.
The script for replacing common OCR errors can be found in Notebook 1.
Japanese has clear typographic conventions that allow for effective distinction between dialogues (「」) and proper names or titles (『』). For more details, see the dialogue extraction script and the discussion of regular expression-based dialogue extraction in Notebook 2.
For a more detailed discussion of this section, refer to Notebook 4 in the repository.
The data from the Higgs boson experiments had a significance level of 5 sigma, corresponding to a p-value of approximately 0.00000057.
