Skip to content
BY 4.0 license Open Access Published by De Gruyter Mouton October 6, 2023

Exploring the potential of sentiment analysis for the study of negative empathy

  • Carmen Bonasera EMAIL logo

Abstract

In connection to literature, negative empathy is a sophisticated form of narrative empathy with fictional characters portrayed as markedly evil and seductive at the same time. Several studies on narrative engagement have explored negative empathy mainly from a theoretical perspective. Conversely, empirical approaches have rarely delved into the dynamics of the linguistic construction of the texts studied. To fill this gap, this paper employs computational techniques to investigate the language of a corpus of novels whose characters are particularly apt for the arousal of negative empathy. More specifically, this study uses Sentiment and Emotion Analysis to explore the lexical representation of emotions and to locate fluctuations in the emotional content of the texts. The ultimate aim is to assess both the potential and the vulnerabilities of Sentiment Analysis for detecting emotional shifts in a literary text and thus for revealing the intensity of its emotional content, which may facilitate the readers’ morally challenging engagement with negative characters.

1 Empathy or empathies?

Recent studies into the link between empathic processes and literary works have led to a range of theoretical and empirical research on readerly involvement with fictional characters (Burke et al. 2016; Hammond and Kim 2014). While the human ability to experience empathy differs from one individual to another, leading to various degrees of engagement with narratives and characters, the concept of “narrative empathy” has been put forth to generally describe “the sharing of feeling and perspective-taking induced by reading, viewing, hearing, or imagining narratives of another’s situation and condition” (Keen 2014: 521). Research on narrative empathy has highlighted the prosocial outcomes of the empathic connection established with fictional characters and stories (Kidd and Castano 2013; Koopman and Hakemulder 2015), focusing on both positive and negative emotional states. Indeed, investigations in social psychology and neuroscience have distinguished between positive and negative empathy, each activating different regions of the brain and triggering different behavioral patterns. In this regard, negative empathy has become a particularly debated topic in recent discussions within aesthetics and literary theory (Ercolino 2018; Ercolino and Fusillo 2022). However, these highlighted that the definitions of negative empathy provided by psychologists are quite narrow if compared to the multifaceted experiences of literary reading, which can result in forms of engagement that stretch the boundaries of what is normally understood as “empathy”. Therefore, studies have argued for broader descriptions that would address the seemingly paradoxical nature of empathizing with both negative emotions and negative characters—i.e., immoral, nasty or repulsive individuals (Ercolino 2018: 244).

An issue connected with this is the lack of empirical research on the linguistic features that may trigger negative empathy. In fact, the few empirical studies on this topic focus on cognitive experiments with readers using questionnaires (e.g., de Jonge et al. 2022; Rebora 2022), without delving deep enough into linguistic or stylistic dynamics. In this article, I address this gap in linguistic research into negative empathy from the perspective of computational literary studies. My main objective is to assess the potential of specific computational tools in examining the texts’ semantic aspect, in order to support the stylistic study of empathy in literary works. More precisely, my general hypothesis is that lexical items linked to the expression of emotions, as well as the oscillation between positive and negative sentiment in a text, might contribute to arousing complex affective experiences, since emotion presentation may well be involved in encouraging readerly empathy with characters. In order to reveal the intensity of the texts’ emotional content, I employed Sentiment Analysis and emotion extraction techniques, which allow to identify several features related to emotion representation, namely: the overall emotional content conveyed by lexical items in a text, the prevalence of certain emotions over others, and emotional shifts throughout the narrative. After introducing the concept of negative empathy among the array of forms of narrative engagement (Section 2), I describe the literary corpus assembled for this study (Section 3), followed by a critical assessment of the chosen computational methodology (Section 4). Finally, Section 5 presents the results of the analyses in comparison with close reading of selected excerpts from the corpus and Section 6 provides a tentative conclusion.

2 Negative empathy: a twisted kind of narrative engagement

As humans, we are capable of being fascinated by, as well as resonating with works of art that portray negative emotions, tragic events, suffering or even repulsive human beings. In fact, Keen suggests that “empathetic responses to fictional characters and situations occur more readily for negative emotions” (2007: xii). While this is true for narrative empathy in general, negative empathy expands on this idea in more sophisticated ways. It is described as “a form of high-level empathy” (Ercolino 2018: 252) that involves:

an aesthetic experience that consists in a cathartic identification with characters, figures in paintings, performances, objects, musical compositions, buildings and spaces, which are disturbingly portrayed as markedly negative and seductive at the same time […]. These are capable of triggering a profound empathic angst in the reader (or viewer, or listener) of the work of art, as well as of persistently prompting her to engage in a moral consideration and of pushing her to assume an ethical position […]. (Ercolino and Fusillo 2022: 70; my translation, emphasis in the original).

This implies that, when translated into aesthetics, the working definitions of negative empathy as “empathy for the negative emotions of others” (Andreychik and Migliaccio 2015: 2) fail to capture the manifold affective reactions that people have while experiencing a disturbingly problematic work of art.[1] In real life, it is highly unlikely to empathize with people who perpetrate disgusting or atrocious acts—such as murderers, psychopaths, pedophiles and rapists—because our moral sense generally impedes our ability to experience empathy for these profiles.[2] However, readers of fiction seem to be able to empathize―albeit with various degrees of intensity―with twisted characters. Keen notes that the narrative’s “protective fictionality” (2007: xiv) allows to connect with negatively portrayed characters because texts are capable of constructing “safe spaces” (2007: 131) for us to be affected by them without experiencing demanding repercussions on real-life actions. By allowing fictional characters “a much higher degree of moral disengagement” (de Jonge et al. 2022: 151), readers can suspend their own moral judgement, which indirectly fuels their empathic disposition.

In this way, fiction keeps readers at a safe distance from immoral characters, which is an essential feature of negative empathy: no matter how enticing these individuals may appear and despite the degree of moral disengagement we allow, the disturbing aspects of their actions are likely to eventually estrange us and mobilize our ethical sense. Negative empathy may thus be characterized as an utterly problematic aesthetic experience that forces readers to oscillate between fascination and identification[3] on the one hand, and detachment and repulsion on the other, especially within certain narratives featuring certain types of immoral or unreliable characters, which are crafted in a way that triggers ambivalent affects in recipients of fiction. The next section will focus on textual factors and narrative techniques employed in the corpus which are involved in shaping immoral characters and their plots.

3 The corpus: literary analysis

This study employs a corpus of literary works that include Fyodor Dostoevsky’s Notes from Underground (1864), Crime and Punishment (1866), and Demons (1872), Vladimir Nabokov’s Lolita (1955), and Jonathan Littell’s The Kindly Ones (2006). These texts—all of which are novels, except for Notes from Underground, a novella—were chosen based on theoretical and empirical investigations (de Jonge et al. 2022; Ercolino and Fusillo 2022) which have considered them as capable of arousing negative empathy due to the perceived immorality of the protagonists. A comparative glance at the corpus shows that three out of the five texts (Crime and Punishment, Demons, and The Kindly Ones) are similar in length, over 15.000 sentences long (see Table 1).[4] This allows for a prolonged immersive experience, which is a feature associated with inducing the intensely cathartic aspect of negative empathy (Ercolino 2018: 253).[5]

Table 1:

Length of the novels calculated in number of sentences.

Novel Number of sentences
Notes from Underground 2,640
Lolita 5,481
Crime and Punishment 15,933
Demons 18,461
The Kindly Ones 22,699

As Ercolino highlights, the amount and quality of time we spend with a negative character while reading is “incomparably greater than the amount of time we would presumably spend with a criminal or a deeply immoral individual” (2018: 250). Hence, the longer we remain exposed to the motives and the interior struggles of a character, the more we would be capable of resonating with their perspective—without necessarily liking them.

Furthermore, these novels are similar in terms of aspects of the emotional representation of the negative hero. As Keen claims for narrative empathy, some specific techniques of characterization, such as the “direct description of a character’s emotional state or circumstances” (2007: 95), may precipitate the readers’ identification with such character, which is in turn a constitutive feature of narrative (and negative) empathy. Sentiment and Emotion analyses help explore this by detecting the numerous descriptions of the characters’ emotional conditions, as well as revealing the intensity of the emotions conveyed (as thoroughly discussed in Section 5).

In addition, especially in a lengthy narrative, first-person narration is likely to encourage some sort of alignment[6] between the reader’s perspective and that of the character, even when narrators are unreliable and manipulative. All texts in this corpus feature ambivalent male protagonists who are either narrators of their own story in fictional autobiographies or confessions (Notes from Underground, Lolita, The Kindly Ones, and the chapter about Stavrogin’s confession in Demons), or the characters from whose point of view the story is mainly narrated (Crime and Punishment). This is in line with Keen’s claim that having access to the character’s consciousness through an internal perspective—achieved either with first-person narration, third-person narration with a focus on the main character, or with an omniscient narrator that has access to the character’s mind—“best promotes character identification and readers’ empathy” (2007: 96). More precisely, she argues that “first person fiction, in which the narrator self-narrates about his or her own experience and perceptions, is thought to invite an especially close relationship between reader and narrative voice” (2007: 97).

Moreover, these characters’ profiles and stories are fashioned and presented such that readers are estranged by their immoral thoughts and actions, but at the same time, they can be captivated by the narrator’s apologetic, sympathy-seeking appeals. A textual strategy that may prompt the enticing side of negative empathy is precisely the apologetic use of “captatio malevolentiae, the courting of rejection”, which von Koppenfels argues may be used by first-person narrators to “compromise the reader by imputing his or her complicity” (2012: 142–143) and to “present himself as a victim” (2012: 144) of the reader’s hostility. Appeals toying with the reader’s complicity abound in the first-person novels in my corpus: from the opening of The Kindly Ones (“Oh my human brothers, let me tell you how it happened. I am not your brother, you’ll retort, and I don’t want to know”, Littell 2010: 3), to the opening of Notes from Underground (“I am a sick man … I am a wicked man. An unattractive man”, Dostoevsky 1993b: 3), to Humbert addressing the reader in Lolita (“Reader! Bruder!”,[7] Nabokov 1991: 262). Nabokov’s and Littell’s protagonists are particularly eager to blame other people for their crimes: the pedophile Humbert goes as far as to declare that little Dolores actually seduced him (“I am going to tell you something very strange: it was she who seduced me. [ …] I have but followed nature. I am nature’s faithful hound”, Nabokov 1991: 132; 135) and, similarly, Nazi officer Max Aue claims to have only followed orders: “you should be able to admit to yourselves that you might also have done what I did. [ …] everyone, or nearly everyone, in a given set of circumstances, does what he is told to do; and, pardon me, but there’s not much chance that you’re the exception, anymore than I was” (Littell 2010: 20). In these two novels, we are invited into an affective short circuit between several contradictory drives: (1) our perspectival alignment with the character, usually facilitated by first-person narration; (2) our compassion for the victims; (3) our moral sense to resist feeling sympathy for the pedophile and the Nazi; (4) a disturbing and shameful involvement with the rhetoric pathos of the persuasive monsters. This short circuit is then amplified when the narrator employs a confessional tone and focuses on the negative character’s inner torment and tragic motives.

As for the other texts, while Notes from Underground presents as an autodiegetic memoir of a spiteful and cynical man, Crime and Punishment and Demons are narrated in the third person with particular focalization on the protagonists; however, they both deploy a confessional register to highlight the characters’ anguish and elicit an ambivalent affective reaction in readers. For example, when we read about Crime and Punishment’s tormented Raskolnikov while he ventures a confession to Sonya about his murders, we could dramatically lose our grip on our ethical position, which would normally condemn his crime. In Demons, the confessional register materializes in Chapter nine of Part two: a censored chapter in early editions, it contains a written confession offered by the central character, Stavrogin, about having committed different crimes, including raping a ten-year-old girl, Matryosha, and causing her suicide. When he admits being haunted by unbearable visions of the little girl, his pathetic focus on her shame and despair reveals glimpses of his own empathic ability:

I saw before me (oh, not in reality! And if only, if only it had been a real vision!), I saw Matryosha, wasted and with feverish eyes, exactly the same as when she had stood on my threshold and, shaking her head, had raised her tiny little fist at me. And nothing had ever seemed so tormenting to me! The pitiful despair of a helpless ten-year-old being with a still unformed mind, who was threatening me [ …]. (Dostoevsky 1995: 703–704)

This window on Stavrogin’s torment may compel the reader into an ambivalent empathic bond with the character. On the one hand, readers might be able to experience a range of affective reactions toward Stavrogin’s anguish—from compassion to empathy—and, on the other, they may be prompted to react by taking a moral stance against him.

There is little consensus over which narrative techniques can facilitate readerly empathy. As Keen argues, narrative strategies function in tandem with other variables related to the recepients’ own empathetic dispositions and the cultural context in which a text is conceived or read (2007: xii). In addition, Fernandez-Quintanilla urges for caution when suggesting relations between textual factors and empathic reactions, “since extra-textual factors are an essential part of the picture and should not be ignored”, concluding that “readers’ engagement with characters results from a complex interplay between textual cues and readerly factors” (2018: 108). Nevertheless, among such textual cues, she lists the representation of the characters’ emotions, thus hinting at the possibility that the expression of emotions influences empathic arousal: “[t]he ways in which characters’ emotions are presented in any given narrative are crucial to the story’s effect on its recipients and the latter’s potential empathetic engagement” (Fernandez-Quintanilla 2018: 137). This assertion might be extended to the uneasy engagement with problematic characters, who inherently convey a heterogeneous range of emotions. The following section will assess Sentiment Analysis as one of the methods that computational literary studies have developed to analyze the texts’ emotional content.

4 Sentiment analysis: potential and limits of the “Syuzhet” package

Within natural language processing, Sentiment Analysis (SA) has become an “umbrella term for the determination of valence, emotions, and other affectual states from text or speech automatically using computer algorithms” (Mohammad 2021: 324). The term “sentiment” is understood as “the underlying positive or negative feeling implied by an opinion” (Liu 2020: 2). Therefore, SA has been employed in a variety of contexts, from customer reviews to blog data, political speeches, and social media posts for sociological or marketing research. In computational literary studies, SA centers on “affective computing applied to textual analysis” (Kim and Klinger 2022: 1), which has also grown to encompass the detection of emotions, so that “some now refer to the field more broadly as emotion analysis” (Mohammad 2021: 324). However, sentiment and emotion analyses entail different tasks and targets: while SA only classifies affective content in polarized terms (negative or positive), emotion analysis recognizes the specific emotions underlying a text, which results in a much more nuanced output. Nevertheless, both sentiment and emotions are the fulcrum of “a growing interest in emotion-oriented text analysis among digital humanities scholars” (Klinger et al. 2020: 238), which has particularly focused on literary texts due to the significance of emotions in narratives, both in text construction and in reader processing.

The present study employs a sentence-level SA approach in R,[8] which draws upon a collection of precompiled sentiment lexicons to label individual lexical items with sentiment scores. SA for literature has received little attention so far (Elkins 2022), perhaps because lexicon-based approaches indeed reveal some limits when dealing with complex syntaxes and semantics, such as those of literary works. The “Syuzhet” package for R developed by Matthew L. Jockers was first released in 2015,[9] with the aim of providing a proper tool to perform SA on literary texts, since it “attempts to reveal the latent structure of narrative by means of sentiment analysis” (Jockers 2020a); by extracting sentiment-based plot arcs from texts, it claims to reveal “the emotional shifts that serve as proxies for the narrative movement between conflict and conflict resolution” (Jockers 2020a). Shortly after its release, it was surrounded by controversy regarding its reliability, as scholars pointed out several substantial weaknesses in the design of Syuzhet’s algorithms (Swafford 2015) and noted the package’s difficulties in detecting some syntactical and semantic information, such as negators (Kim 2022: para. 25; Naldi 2019: 5). This resulted, on the one hand, in a debate over the limits of using SA in literary studies and, on the other, in an effort to improve SA approaches. As a consequence, the package was continuously upgraded over the course of five years and eventually validated by Misuraca et al. (2020). The latest version to date (Syuzhet 1.0.6) was released in November 2020. Despite critiques and some acknowledged limitations, Syuzhet has been constantly downloaded since its first release, becoming the most popular R package to perform SA on literary texts.[10]

The analysis conducted in this study is intended as a tentative appraisal of the potential of SA and emotion extraction techniques—and of the Syuzhet package in particular—in detecting dramatic fluctuations in the sentiment plot arc and in recording scores and variations in the expression of emotions in specific narrative moments, which could hypothetically concur in arousing ambivalent affects. As specified in Section 1, the intention of these analyses is to substantiate the study of negative empathy from a computational perspective,[11] by carrying out a textual investigation at the semantic level. The following section accounts for a critical examination of the results of the analyses.

5 Detecting sentiment and emotions in literary texts

5.1 Sentiment analysis

For the purposes of a homogeneous analysis, I employed the English translations of the novels originally written in Russian or French.[12] To encourage matchmaking between terms in the corpus and those present in sentiment and emotion lexicons, the corpus underwent a process of lowercasing. Subsequently, sentences were parsed using the default Syuzhet’s function (get_sentences). As for emotion extraction, each text was lemmatized and punctuation and stopwords were eliminated. The sentiment polarity of each novel was then computed with several functions available in the Syuzhet package: these functions calculate the sentiment score for each sentence—i.e., the sum of positive and negative values of all words found in the sentence that match with the lexicon—and thus converts the textual information of each sentence into numerical sentiment vectors that depict the novels’ emotional valence. The get_sentiment function was used to visualize the degrees of the sentiment polarity score with respect to the overall narrative time, expressed in number of sentences. The resulting graphs (Figures 1 5, left), accounting for both positive (e.g., joyous) and negative (e.g., sad) sentiment, are too dense to interpret for a macroanalysis of the sentiment plot, since they prevent us from seeing clearly the general emotional arc; as Jockers suggests (2020b), it is preferable to remove the noise and visualize simpler shapes. Therefore, the get_dct_transform function was used, which aims at grasping the whole emotional flow of plots by applying discrete cosine transform (DCT),[13] thus simplifying the emotional valence with a low-pass smoothing filter—i.e., a filter that can mediate between different frequencies. The resulting graphs for each novel (Figures 1 5, right) display polarity in a much simpler way (on a numeric scale from −1.0 to +1.0) and reveal considerable fluctuations in the sentiment trend of each narrative, spanning from the negative (y < 0) to the positive (y > 0) feelings embedded in the texts’ semantic information.

Figure 1: 
Raw (left) and transformed (right) sentiment plot trajectory of “Stavrogin’s Confession” (in Demons).
Figure 1:

Raw (left) and transformed (right) sentiment plot trajectory of “Stavrogin’s Confession” (in Demons).

Figure 2: 
Raw (left) and transformed (right) sentiment plot trajectory of Notes from Underground.
Figure 2:

Raw (left) and transformed (right) sentiment plot trajectory of Notes from Underground.

Figure 3: 
Raw (left) and transformed (right) sentiment plot trajectory of Lolita.
Figure 3:

Raw (left) and transformed (right) sentiment plot trajectory of Lolita.

Figure 4: 
Raw (left) and transformed (right) sentiment plot trajectory of Crime and Punishment.
Figure 4:

Raw (left) and transformed (right) sentiment plot trajectory of Crime and Punishment.

Figure 5: 
Raw (left) and transformed (right) sentiment plot trajectory of The Kindly Ones.
Figure 5:

Raw (left) and transformed (right) sentiment plot trajectory of The Kindly Ones.

To have a comprehensive view of the sentiment trajectories, the simple_plot function was used on the default Syuzhet sentiment vector to apply three smoothing methods (a rolling mean, LOESS, and DCT).[14] As Jockers (2020b) explains, this function produces two plots: the first one (Figures 6 10, top) shows all three smoothing methods on the same graph; the second one (Figures 6 10, bottom) shows only the simplified shape of the plot on a normalized time axis.

Figure 6: 
Sentiment plot trajectory in “Stavrogin’s confession” (in Demons).
Figure 6:

Sentiment plot trajectory in “Stavrogin’s confession” (in Demons).

Figure 7: 
Sentiment plot trajectory in Notes from Underground.
Figure 7:

Sentiment plot trajectory in Notes from Underground.

Figure 8: 
Sentiment plot trajectory in Lolita.
Figure 8:

Sentiment plot trajectory in Lolita.

Figure 9: 
Sentiment plot trajectory in Crime and Punishment.
Figure 9:

Sentiment plot trajectory in Crime and Punishment.

Figure 10: 
Sentiment plot trajectory in The Kindly Ones.
Figure 10:

Sentiment plot trajectory in The Kindly Ones.

At a cursory glance, all graphs display fluctuations in the polarity score, oscillating between positive and negative feelings. More specifically, the plot of “Stavrogin’s Confession”[15] in Demons (Figure 6) shows a quite regular fluctuation between positive and negative sentiment in the first half of the narrative, hitting the lowest peak at around x ≈ 400 (score −7.2). This corresponds to moments after Stavrogin confesses having raped little Matryosha (Dostoevsky 1995: 697), as shown in Figure 11.

Figure 11: 
Sentence 413 of “Stavrogin’s Confession” with sentiment score.
Figure 11:

Sentence 413 of “Stavrogin’s Confession” with sentiment score.

The sentiment score in Notes from Underground (Figure 7) reaches its highest peak at the beginning of the narrative. The polarity scores progressively decrease towards the lowest point, which occurs just before x ≈ 2,500, and can be pinpointed at x = 2,416, which scored −4.1 (“The chief martyr, of course, was myself, because I was fully conscious of all the loathsome baseness of my spiteful stupidity, and at the same time I simply could not restrain myself”, Dostoevsky 1993b: 120). However, other sentences obtained lower scores in the SA and were not displayed properly in the resulting plot, bringing to question the accuracy of the smoothing filter. A clear example is the sentence in Figure 12 (x = 158), which scored −6.45 and which was, in fact, one of the sentences that were hypothesized as being extremely negative during a preliminary manual annotation of the novel (Dostoevsky 1993b: 12).

Figure 12: 
Sentence 158 of Notes from Underground with sentiment score.
Figure 12:

Sentence 158 of Notes from Underground with sentiment score.

The plot of Lolita (Figure 8) displays a more regular fluctuation in the first half of the novel, reaching a higher peak at the beginning of the second half (x ≈ 3,000); then the plot steadily decreases towards the end. However, if compared to its raw equivalent (Figure 3, left), Figure 8 does not properly render the negative point at x ≈ 2,900 and the positive spike at x ≈ 3,900, corresponding to the following sentences: Figure 13 shows the highly negative sentence, which occurs shortly after Dolores’s abduction and rape (Nabokov 1991: 169), while Figure 14 shows the highly positive sentence, which is one of the many descriptions of her body by Humbert (Nabokov 1991: 231–232).

Figure 13: 
Sentence 2832 in Lolita with sentiment score.
Figure 13:

Sentence 2832 in Lolita with sentiment score.

Figure 14: 
Sentence 3907 in Lolita with sentiment score.
Figure 14:

Sentence 3907 in Lolita with sentiment score.

When compared to its raw equivalent (Figure 4, left), the plot of Crime and Punishment (Figure 9) ignores a negative peak around x ≈ 7,000, failing to render the drastic sentiment changes that occur in the first chapters, where Raskolnikov murders the pawnbroker and her sister. Similarly, the sentence in Figure 15, which describes Raskolnikov’s anguish while carrying out the murders (Dostoevsky 1993a: 79–80), is not adequately represented in the final graph, although it scored -5.15.

Figure 15: 
Sentence 1950 in Crime and Punishment with sentiment score.
Figure 15:

Sentence 1950 in Crime and Punishment with sentiment score.

Nevertheless, the SA correctly classifies Raskolnikov’s final confession to Sonya as negatively polarized, as the literary analysis had foreseen (see Section 3); the sentence in Figure 16 (Dostoevsky 1993a: 410) scored −4.5.

Figure 16: 
Sentence 11716 in Crime and Punishment with sentiment score.
Figure 16:

Sentence 11716 in Crime and Punishment with sentiment score.

Finally, Figure 10 illustrates a regular fluctuation in the first half of The Kindly Ones; then after a positive stability, it declines sharply at the end. The lowest peak is reached in the chapter “Air”, where Aue experiences visions of his past crimes: the long period in Figure 17 (Littell 2010: 912), with its articulated use of hypotactic structures and a confessional register, creates a sense of claustrophobic “entrapment” (Ercolino 2018: 254).

Figure 17: 
Sentence 21173 in The Kindly Ones with sentiment score.
Figure 17:

Sentence 21173 in The Kindly Ones with sentiment score.

Two consecutive sentences in this narrative suggest a connection between a relevant variation in sentiment and the possible arousal of negative empathy. These sentences show Aue executing a wounded woman while he is describing the mass execution that Nazi officers perpetrated against Jewish civilians at Babi Yar (Ukraine), one of the largest massacres in WWII (Littell 2010: 129–130). In Figure 18, the first period was evaluated as positive, scoring 2.9, but the following one displays a sudden fall at −3.95, therefore giving rise to a sharp contrast.

Figure 18: 
Sentences 3147 and 3148 in The Kindly Ones with sentiment scores.
Figure 18:

Sentences 3147 and 3148 in The Kindly Ones with sentiment scores.

This passage may exemplify how texts can be able to trigger the experience of negative empathy: a distressing oscillation between the character’s (and, possibly, the reader’s) empathic engagement with the victim, and his self-imposed detachment from the horror he is perpetrating. According to Christine Berberich, this oscillation “has the exact opposite effect on the reader than it has on Aue”: the reader cannot distance herself from the narrative, unless she stops reading, because “the ekphrastic, vivid description of the scene involves [her] too much” (Berberich 2020: 176).

With regards to polarity, Syuzhet has proven to be efficient in detecting the sentiment score of each sentence, with few incongruities if compared to the paragraphs that the literary analysis had previously highlighted (see Section 3). Nevertheless, in some cases, the plots generated by the smoothing algorithms were less accurate than what was expected from the SA raw results that were obtained with the get_sentiment function (Figures 1 5, left). In order to offer a more detailed picture of the actual emotional trajectory of each novel, it is necessary to carry out a further analysis on the emotions and their trend throughout the narratives.

5.2 Emotion extraction

Emotion analysis, i.e., the evaluation of the emotional information encoded in texts, was performed by deploying the NRC Word-Emotion Association Lexicon (EmoLex; Mohammad and Turney 2013), which includes lexical entries for around 14,000 English terms. EmoLex evaluates words by categorizing them based on their association to eight basic emotions: joy/sadness, anger/fear, trust/disgust, surprise/anticipation (Plutchik 1980). Therefore, the analysis allows to extract quantitative information related not only to sentiment polarity, as with the Syuzhet tool, but also to the distribution of each emotion. This section focuses on the analysis of emotions and of their development throughout each narrative, which was performed by computationally scanning the texts in comparison with the NRC lexicon,[16] with the aim of examining in depth the emotional density of the corpus at a sentence level. Figures 19 23 show the resulting graphs.

Figure 19: 
Emotions in “Stavrogin’s Confession”.
Figure 19:

Emotions in “Stavrogin’s Confession”.

Figure 20: 
Emotions in Notes from Underground.
Figure 20:

Emotions in Notes from Underground.

Figure 21: 
Emotions in Lolita.
Figure 21:

Emotions in Lolita.

Figure 22: 
Emotions in Crime and Punishment.
Figure 22:

Emotions in Crime and Punishment.

Figure 23: 
Emotions in The Kindly Ones.
Figure 23:

Emotions in The Kindly Ones.

In “Stavrogin’s Confession”, the emotional content develops quite unsteadily; Figure 19 reports higher peaks in anger and fear, corresponding to the moments in which Stavrogin confesses having raped Matryosha and having drawn her to suicide. In particular, the sentence that was recognized by SA as the lowest peak (see Figure 11), obtained the highest score in both anger and fear.

In Notes from Underground (Figure 20), anger, sadness and fear prevail, showing a regular increase toward the end of the narrative. More exactly, anger and disgust are particularly conveyed by the excerpt in Figure 24 (Dostoevsky 1993b: 8), through lexical items such as “bitterness” or “ashamed”.

Figure 24: 
Sentence 103 in Notes from Underground with emotion scores.
Figure 24:

Sentence 103 in Notes from Underground with emotion scores.

On the contrary, the analysis detected a prevalence of fear and sadness items (e.g., “torment”, “unbearable”, “humiliation”) in the sentence in Figure 25, in the second part of the novella (Dostoevsky 1993b: 52).

Figure 25: 
Sentence 798 in Notes from Underground with emotion scores.
Figure 25:

Sentence 798 in Notes from Underground with emotion scores.

In Lolita (Figure 21), a surprising high incidence of positive emotions (trust and joy) finds explanation in the many instances in which Humbert reflects on his infatuation for Dolores, as well as in the fact that, being a highly unreliable narrator, he is able to disguise his sexual obsession for his stepdaughter as a profound and consensual love bond. Figure 21 also displays a rising trend for both fear and sadness toward the end, when Humbert is desperately searching for the fugitive Dolores, while also growing aware of his abjection (see Figure 26; Nabokov 1991: 285).

Figure 26: 
Sentence 4934 in Lolita with emotion scores.
Figure 26:

Sentence 4934 in Lolita with emotion scores.

In these final portions of the novel, Humbert’s reflections show a distinct contrast between the expression of fear and sadness (“shame”, “despair”, “agonized”)—which the analysis detected as prevailing in sentence 4934 (Figure 26)—and that of joy—which reaches a high score in sentence 4937 (Figure 27) through items such as “enchanting”, “sweet”, “radiance”, “endearing”—suggesting that the text conveys diametrically opposed feelings in the span of a few sentences, allowing for a significant shift in the expression of emotions (Nabokov 1991: 285).

Furthermore, Crime and Punishment (Figure 22) features high peaks of fear and sadness. Two moments were detected as conveying a great amount of negative emotional content: the first one, expressing both fear and disgust, occurs when Raskolnikov realizes the misery of his situation, which would lead him to commit the murder (“The feeling of boundless loathing that had begun to oppress and sicken his heart while he was still only on his way to the old woman now reached such proportions and became so clearly manifest that he did not know where to flee from his anguish”, Dostoevsky 1993a: 9–10, emphasis mine); the second one, which entails a high amount of sadness items, is when Sonya realizes that he is the assassin (“After her first passionate and tormenting sympathy for the unhappy man, the horrible idea of the murder struck her again”, Dostoevsky 1993a: 412, emphasis mine), a moment that also poignantly describes the oscillation between closeness and detachment that characterizes negative empathy.

Figure 27: 
Sentence 4937 in Lolita with emotion scores.
Figure 27:

Sentence 4937 in Lolita with emotion scores.

Finally, in The Kindly Ones (Figure 23), fear and trust predominate over the long and unreliable narrative, but higher peaks of fear, sadness and disgust are found in the detailed depiction of the protagonist’s involvement in the massacres (see Section 5.1). While approaching the end, the overall emotional content declines steadily, with the sole exception of the penultimate sentence, which shows an increase in fear and sadness that are bound to accompany the reader through the novel’s closure: “I felt all at once the entire weight of the past, of the pain of life and of inalterable memory, I remained alone with the dying hippopotamus, a few ostriches, and the corpses, alone with time and grief and the sorrow of remembering, the cruelty of my existence and of my death still to come” (Littell 2010: 975, emphasis mine).

6 Discussion and conclusions

This study has analyzed the semantic expression of emotions as detected by computational tools in novels that may be capable of triggering readerly negative empathy, thus hinting at a connection between the lexical representation of emotions and the possible arousal of ambivalent affects. In line with the primary aim of this article, the analysis has shown that these computational techniques can be successfully employed to locate and visualize the emotional density of each sentence, as well as sentiment polarizations and emotional fluctuations over the course of a narrative. Whether these specific fluctuations coincide with the arousal of readerly negative empathy is still debatable and open to further experimental inquiries. Nevertheless, the detected variations from extreme high to extreme low sentiment levels, as well as the sudden shifts located in the emotional content might conceivably have some sort of impact on the readers’ affective reactions. Being sharp fluctuations in the representation of emotions, they might have a connection with the distressing oscillation between estrangement and affective arousal, which is consistent with the theory of negative empathy.

Clearly, lexicon-based Sentiment Analysis presents some weaknesses. Since it considers only the lexical level, it fails to detect other stylistic and pragmatic factors such as irony, ambiguity or foregrounding. It may also display varied results according to different translations (see Section 5.1). Moreover, the Syuzhet package revealed some incongruities in the functions aimed at grasping the emotional flow. In this regard, the get_sentiment function generated a more detailed output of the emotional valence, and thus seems more adequate for micro-SA. While the next step would be to integrate these results with reader reports and cognitive investigations, this computational assessment suggests an interesting parallel between the polarization of sentiment and the degree of emotional intensity in the novels’ lexical choices, on the one hand, and the theorized dynamics of negative empathy on the other, ultimately adding a further tile to traditional literary interpretations.


Corresponding author: Carmen Bonasera, University of Bologna, Bologna, Italy, E-mail:

Acknowledgments

A preliminary draft of this paper was presented at the 2022 PALA Conference at Aix-Marseille University. I am grateful to guest editors Carolina Fernandez-Quintanilla and Fransina Stradling for their perceptive comments, which helped me to improve the manuscript, and to Stefano Ercolino, who encouraged me to investigate the topic of negative empathy. I would also like to thank the instructors of the 2021 IGEL Training School on Sentiment Analysis for providing me with the fundamental methodologies to engage in computational literary studies, as well as Claudia Roberta Combei for her valuable feedback since the earliest phases of this study.

References

Andreychik, Michael R. & Nicole Migliaccio. 2015. Empathizing with others’ pain versus empathizing with others’ joy: Examining the separability of positive and negative empathy and their relation to different types of social behaviors and social emotions. Basic and Applied Social Psychology 37(5). 274–291. https://doi.org/10.1080/01973533.2015.1071256.Search in Google Scholar

Berberich, Christine. 2020. Ekphrasis and the holocaust: Traumatic images in Jonathan Littell’s The Kindly Ones. In Unni Langas & Charles Ivan Armstrong (eds.), Terrorizing images: Trauma and ekphrasis in contemporary literature, 167–182. Berlin & Boston: De Gruyter.10.1515/9783110693959-011Search in Google Scholar

Boase-Beier, Jean. 2014. Translation and the representation of thought: The case of Herta Müller. Language and Literature 23(3). 213–226. https://doi.org/10.1177/0963947014536503.Search in Google Scholar

Burke, Michael, Anežka Kuzmičová, Anne Mangen & Theresa Schilhab. 2016. Empathy at the confluence of neuroscience and empirical literary studies. Scientific Study of Literature 6(1). 6–41. https://doi.org/10.1075/ssol.6.1.03bur.Search in Google Scholar

Dostoevsky, Fyodor. 1993a [1866]. Crime and punishment, Translated by Richard Pevear & Larissa Volokhonsky. New York: Vintage Classics.Search in Google Scholar

Dostoevsky, Fyodor. 1993b [1864]. Notes from underground. Translated by Richard Pevear & Larissa Volokhonsky. New York: Vintage Classics.Search in Google Scholar

Dostoevsky, Fyodor. 1995 [1872]. Demons: A novel in three parts. Translated by Richard Pevear & Larissa Volokhonsky. New York: Vintage Classics.Search in Google Scholar

Eisenberg, Nancy & Natalie D. Eggum. 2009. Empathic responding: Sympathy and personal distress. In Jean Decety & William Ickes (eds.), The social neuroscience of empathy, 71–83. Cambridge, MA: MIT Press.10.7551/mitpress/9780262012973.003.0007Search in Google Scholar

Elkins, Katherine & Jon Chun. 2019. Can sentiment analysis reveal structure in a plotless novel? ArXiv 1910.01441. https://doi.org/10.48550/arXiv.1910.01441.Search in Google Scholar

Elkins, Katherine. 2022. The shapes of stories: Sentiment analysis for narrative. Cambridge: Cambridge University Press.10.1017/9781009270403Search in Google Scholar

Ercolino, Stefano. 2018. Negative empathy: History, theory, criticism. Orbis Litterarum 73(3). 243–262. https://doi.org/10.1111/oli.12175.Search in Google Scholar

Ercolino, Stefano & Massimo Fusillo. 2022. Empatia negativa. Il punto di vista del male. Milano: Bompiani.Search in Google Scholar

Felski, Rita. 2019. Identifying with characters. In Amanda Anderson, Rita Felski & Toril Moi (eds.), Character: Three inquiries in literary studies, 77–126. Chicago & London: The University of Chicago Press.Search in Google Scholar

Fernandez-Quintanilla, Carolina. 2018. Language and narrative empathy: An empirical stylistic approach to readers’ engagement with characters. Lancaster: Lancaster University Dissertation.Search in Google Scholar

Grisot, Giulia. 2021. IGEL 2021 Sentiment analysis training school. https://github.com/mountain-sentiment/mountain-sentiment.github.io/tree/master/code_and_scripts/IGEL_2021_Sentiment_Analysis_TS_Beginners_Grisot_G (accessed 5 February 2023).Search in Google Scholar

Hammond, Meghan Marie & Sue J. Kim (eds.). 2014. Rethinking empathy through literature. New York: Routledge.10.4324/9781315818603Search in Google Scholar

Hoffman, Martin L. 2000. Empathy and moral development: Implications for caring and justice. New York: Cambridge University Press.10.1017/CBO9780511805851Search in Google Scholar

Jockers, Matthew. 2020a. Syuzhet: Extracts sentiment and sentiment-derived plot Arcs from text Version 1.0.6. https://cran.r-project.org/web/packages/syuzhet/index.html (accessed 5 February 2023).Search in Google Scholar

Jockers, Matthew. 2020b. Introduction to the Syuzhet package. https://cran.r-project.org/web/packages/syuzhet/vignettes/syuzhet-vignette.html (accessed 5 February 2023).Search in Google Scholar

Jonge, Julia de, Serena Demichelis, Simone Rebora & Massimo Salgaro. 2022. Operationalizing perpetrator studies. Focusing readers’ reactions to The Kindly Ones by Jonathan Littell. Journal of Literary Semantics 51(2). 147–161. https://doi.org/10.1515/jls-2022-2057.Search in Google Scholar

Keen, Suzanne. 2007. Empathy and the novel. Oxford & New York: Oxford University Press.Search in Google Scholar

Keen, Suzanne. 2014. Narrative empathy. In Peter Hühn, Jan Christoph Meister, John Pier & Wolf Schmid (eds.), Handbook of narratology, Vol. 2, 521–530. Berlin, München & Boston: De Gruyter.10.1515/9783110316469.521Search in Google Scholar

Kidd, David Comer & Emanuele Castano. 2013. Reading literary fiction improves theory of mind. Science 342(6156). 377–380. https://doi.org/10.1126/science.1239918.Search in Google Scholar

Kim, Evgeny & Roman Klinger. 2022. A survey on sentiment and emotion analysis for computation literary studies. ArXiv 1808.03137. https://doi.org/10.48550/arXiv.1808.03137.Search in Google Scholar

Kim, Hoyeol. 2022. Sentiment analysis: Limits and progress of the Syuzhet package and its lexicons. DHQ: Digital Humanities Quarterly 16(2). http://www.digitalhumanities.org/dhq/vol/16/2/000612/000612.html (accessed 5 February 2023).Search in Google Scholar

Klinger, Roman, Evgeny Kim & Sebastian Padó. 2020. Emotion analysis for literary studies: Corpus creation and computational modelling. In Nils Reiter, Axel Pichler & Jonas Kuhn (eds.), Reflektierte algorithmische Textanalyse: Interdisziplinäre(s) Arbeiten in der CRETA-Werkstatt, 237–268. Berlin & Boston: De Gruyter.10.1515/9783110693973-011Search in Google Scholar

Koopman, Eva Maria & Frank Hakemulder. 2015. Effects of literature on empathy and self-reflection: A theoretical-empirical framework. Journal of Literary Theory 9(1). 79–111.10.1515/jlt-2015-0005Search in Google Scholar

Koppenfels, Martin von. 2012. The infamous “I”: Notes on Littell and Céline. In Aurélie Barjonet & Liran Razinsky (eds.), Writing the Holocaust today: Critical perspectives on Jonathan Littell’s “The Kindly Ones”, 133–152. Amsterdam & New York, NY: Rodopi.Search in Google Scholar

Littell, Jonathan. 2010 [2006]. The kindly ones. Translated by Charlotte Mandell. London: Vintage.Search in Google Scholar

Liu, Bing. 2020. Sentiment analysis: Mining opinions, sentiments, emotions, 2nd edn. Cambridge & New York: Cambridge University Press.10.1017/9781108639286Search in Google Scholar

Misuraca, Michelangelo, Alessia Forciniti, Germana Scepi & Maria Spano. 2020. Sentiment analysis for education with R: Packages, methods and practical applications. arXiv:2005.12840. https://doi.org/10.48550/arXiv.2005.12840.Search in Google Scholar

Mohammad, Saif & Peter Turney. 2013. Crowdsourcing a word-emotion association lexicon. Computational Intelligence 29(3). 436–465. https://doi.org/10.1111/j.1467-8640.2012.00460.x.Search in Google Scholar

Mohammad, Saif, Mohammad Salameh & Svetlana Kiritchenko. 2016. How translation alters sentiment. Journal of Artificial Intelligence Research 55. 95–130. https://doi.org/10.1613/jair.4787.Search in Google Scholar

Mohammad, Saif. 2021. Sentiment Analysis: Automatically detecting valence, emotions, and other affectual states from text. In Herbert L. Meiselman (ed.), Emotion measurement, 2nd edn., 324–381. Duxford (UK), Cambridge (MA) & Kidlington (UK): Elsevier.Search in Google Scholar

Morton, Adam. 2011. Empathy for the devil. In Amy Coplan & Peter Goldie (eds.), Empathy: Philosophical and psychological perspectives, 318–330. Oxford: Oxford University Press.10.1093/acprof:oso/9780199539956.003.0019Search in Google Scholar

Nabokov, Vladimir. 1991[1955]. The annotated Lolita. In Alfred Appel (ed.). New York: Vintage.Search in Google Scholar

Naldi, Maurizio. 2019. A review of sentiment computation methods with R packages. ArXiv:1901.08319. https://doi.org/10.48550/arXiv.1901.08319.Search in Google Scholar

Nussbaum, Martha. 1990. Love’s knowledge: Essays on philosophy and literature. New York: Oxford University Press.Search in Google Scholar

Plutchik, Robert. 1980. A general psychoevolutionary theory of emotion. In Robert Plutchik & Kellerman Henry (eds.), Emotion: Theory, research, and experience, vol. 1, 3–33. New York: Academic Press.10.1016/B978-0-12-558701-3.50007-7Search in Google Scholar

R Core Team. 2022. R: A language and environment for statistical computing. Vienna: R Foundation for Statistical Computing. https://www.R-project.org/ (accessed 5 February 2023).Search in Google Scholar

Rebora, Simone. 2022. Empatia negativa ed eroe ibrido: il contributo dei metodi empirici. In Giuseppe Carrara & Laura Neri (eds.), Con i buoni sentimenti si fanno brutti libri? Etiche, estetiche e problemi della rappresentazione, 109–121. Milano: Ledizioni.10.4000/books.ledizioni.14968Search in Google Scholar

Swafford, Annie. 2015. Problems with the Syuzhet package. In Anglophile in academia: Annie Swafford’s blog. 2 March 2015. https://annieswafford.wordpress.com/2015/03/02/syuzhet/ (accessed 5 February 2023).Search in Google Scholar

Published Online: 2023-10-06
Published in Print: 2023-10-26

© 2023 the author(s), published by De Gruyter, Berlin/Boston

This work is licensed under the Creative Commons Attribution 4.0 International License.

Downloaded on 1.3.2024 from https://www.degruyter.com/document/doi/10.1515/jls-2023-2011/html
Scroll to top button