1. Introduction
The prominent role of shame and guilt in contemporary societies became apparent during the COVID-19 pandemic. The desire to be seen as moral personae, and the shame and guilt resulting from the failure to be perceived as such, presumably played a key role in fostering moral behaviours. This, however, is by no means a novel phenomenon. Long before COVID-19, shame and guilt have been understood as modifiers of behaviour (Bedford & Hwang, Reference Bedford and Hwang2003), markers of personal identity (Hultberg, Reference Hultberg1988), and mechanisms for social control across cultures (Creighton, Reference Creighton1990). Yet, their effects on the emoterFootnote 1 (i.e., whoever experiences the emotion; Bednarek, Reference Bednarek2008, p. 14; Glynn, Reference Glynn2014, p. 72) are controversial. In the field of psychology, it is now widely acknowledged that shame and guilt are central in fostering socially responsible behaviours or avoidance of behaviours that may lead to disapproval (Sabiston & Castonguay, Reference Sabiston, Castonguay, Eklund and Tenenbaum2014). However, some studies positively correlate shame and guilt with a variety of negative behavioural, psychological, and physical outcomes such as depression and anxiety (Cavalera, Reference Cavalera2020; Sabiston & Castonguay, Reference Sabiston, Castonguay, Eklund and Tenenbaum2014, p. 626). The degree to which these aspects of shame and guilt are shared across cultures is also controversial, as they have been categorised differently by different scholars – even within the same field, including linguistics and psychology. Unlike anger, surprise, disgust, enjoyment, fear and sadness, which have traditionally been considered to be primary emotions shared by all humans (e.g., Ekman, Reference Ekman1992),Footnote 2 secondary emotions are learnt through socialisation, and are, therefore, culture-specific (Wierzbicka, Reference Wierzbicka1999). In this study, we subscribe to the view that shame and guilt are secondary emotions. The inherently culture-dependent nature of shame and guilt, and the actual behaviours associated with them, have been examined across cultures from a psychological perspective (Arimitsu, Reference Arimitsu2001; Sheikh & Janoff-Bulman, Reference Sheikh and Janoff-Bulman2010; Smith et al., Reference Smith, Webster, Parrott and Eyre2002; Suzuki, Reference Suzuki2007; Tangney & Dearing, Reference Tangney and Dearing2002). Linguists have also tackled these phenomena, focusing on the linguistic structures that give us access to shame and guilt as emotion concepts and their variability across languages (e.g., Fabiszak & Hebda, Reference Fabiszak and Hebda2007; Krawczak, Reference Krawczak2017, Reference Krawczak2018; Kumamoto, Reference Kumamoto2019; Tissari, Reference Tissari2006). The underlying theoretical framework in these studies and ours is that language is usage-based. That is, the idea that linguistic knowledge is shaped by the context and frequency of language use, thus the language of shame and guilt is also ‘usage-based’ (e.g., Fabiszak et al., Reference Fabiszak, Hilpert and Krawczak2016; Geeraerts, Reference Geeraerts2010; Glynn, Reference Glynn2007, Reference Glynn, Glynn and Fischer2010a, Reference Glynn, Schmid and Handl2010b; Langacker, Reference Langacker1987; Vigliocco et al., Reference Vigliocco, Meteyard, Andrews and Kousta2009, p. 222), and can be better understood by exploring the different contexts in which expressions of shame and guilt are present.
This study adds to the existing literature on the topic by proposing a linguistic approach to shame and guilt in two typologically and pragmatically different languages, namely Japanese and English. It does so by asking how these two emotions are metadiscursively framed in two online forums that are relatively similar in terms of audience, aim, and structure – a factor that increases comparability. Hence, the main focus is on only one aspect of the highly complex phenomena of shame and guilt, namely how people talk about them. By using computational tools from the field of natural language processing (NLP), we extract data in novel ways and further add to existing methodologies in corpus linguistics. Our aim is to further contribute to existing theories of emotions and the language through which such emotions are verbalised by extracting new insights with novel approaches and comparing the results to previous findings. We aim to enhance academic perspectives on emotions by focusing on how people discuss them. This approach is based on the premise that overlooking people’s conceptualisations could lead to the creation of analytical artifacts. These artifacts might lack real-world relevance and fail to accurately represent the experiences of those involved in the interaction.
2. ‘Shame’ and ‘guilt’: what they are and what they do
This section presents a working definition of ‘shame’ and ‘guilt’, in single quotation marks when we want to indicate that they are not the English words shame and guilt actually used in interaction, but more general emotional labels that are shared, to some degree, across languages and cultures. It then presents the link between (moral) emotions and evaluation. Finally, it reviews some important studies that tackle the language of emotions across linguacultures.
2.1. A working definition
Drawing from Tangney and Dearing (Reference Tangney and Dearing2002, p. 25), we tentatively define ‘shame’ and ‘guilt’ as negatively valenced moral emotions typically experienced in interpersonal contexts. They are negatively valenced because they imply a negative evaluation by others of one’s conduct/identity, and moral because they guide and are guided by our sense of good and bad. Since ‘shame’ and ‘guilt’ are not predetermined genetically, but learnt, negotiated, and eventually challenged through socialisation, it is reasonable to assume that their conceptualisations may differ across linguacultures (Bedford & Hwang, Reference Bedford and Hwang2003).
2.2. Emotions and evaluation
The presumed causal link between ‘shame’ and ‘guilt’ and a higher moral order (i.e., ‘a culture-specific ideology about what counts as right or wrong’ (Culpeper & Tantucci, Reference Culpeper and Tantucci2021, p. 148) is also evident in their characterisation as moral emotions that ‘provide immediate punishment (or reinforcement) of behavior’ (Tangney & Dearing, Reference Tangney and Dearing2002, p. 133). This definition makes apparent the evaluative nature of moral emotions. On this basis, and drawing from the evaluative tradition on emotions (Scarantino, Reference Scarantino, Haviland-Jones, Lewis and Barrett2016), ‘shame’ and ‘guilt’ – and other valenced emotions – are here conceptualised as forms of evaluation, which in turn is intended in a sense described by Hunston and Thompson (Reference Hunston and Thompson2000, p. 5) as the indication that something is good or bad. In the case of ‘shame’ and ‘guilt,’ the latter applies: ‘shame’ and ‘guilt’ are emotions that originate from the producer’s awareness of having failed to be or behave in accordance with the standards recognised as proper by the group. They entail a negative judgement of actions or behaviours deemed shameful or guilt-worthy and, by extension, of the individuals responsible for these actions. Looking at (the linguistic manifestation of) emotions as a form of evaluation accommodates the inherent normative aspect of ‘shame’ and ‘guilt’. It is important to note that the methods employed in this study do not give us access to the emotions themselves. Emotions are internal phenomena and, as such, not amenable to direct empirical observation. However, we can use the linguistic manifestations of emotions as a proxy for the actual emotions (Laaksonen et al., Reference Laaksonen, Pääkkönen and Öhman2023).
2.3. Emotion talk across languages
Bednarek (Reference Bednarek2008) makes an important distinction between emotion talk and emotional talk.Footnote 3 The former indicates the language about emotions and ‘is constituted by all those expressions in the dictionary that denote affect/emotion, for example, love, hate, joy, envy, sad, mad, enjoy, dislike and so on’. The latter relates to ‘all those constituents (linguistic and non-linguistic) that conventionally express or signal affect/emotion (whether genuinely experienced or not, whether intentional or not)’ (Bednarek, Reference Bednarek2008, p. 11). This study focuses on the former, i.e., emotion talk, as we look at ‘shame’ and ‘guilt’ only insofar as the linguistic expressions that denote the internal states conventionally associated with them are present in the text. Such explicit linguistic resource pertains in appraisal theory (Martin & White, Reference Martin and White2005) to ‘Judgement’ – a sub-system of ‘Affect’ concerned with ‘resources for assessing behaviour according to various normative principles’ (Martin & White, Reference Martin and White2005, p. 35).
In the definition of Judgement given above, we have again the complex interrelationship of moral emotions, evaluations and norms: emotions are a form of evaluation in that they always encode interactants’ point of view towards something. When they are moral in nature, evaluations are based on normative standards that a given group or community sees as proper, that is, ideologies (Garfinkel, Reference Garfinkel1967; Heinrich, Reference Heinrich2012; Verschueren, Reference Verschueren2011). Just as evaluations in terms of (in)appropriateness can vary from one community to the other, and even among individuals of the same group, also the words conventionally associated with the emotions that are manifestations of such evaluations may acquire different meanings in different contexts. It follows that, even when we have a direct translation of English metalexemes, such as shame and guilt in other languages, what they conventionally index may differ across linguacultures (Kádár & Haugh, Reference Kádár and Haugh2013; Kádár & Ran, Reference Kádár, Ran, Ogiermann and Blitvich2019; Kumamoto, Reference Kumamoto2019; Soares da Silva, Reference Soares da Silva2020). This, however, is a point often overlooked in the literature, where scientific metalabels (which are almost invariably in English) are rarely problematised. Some notable exceptions are Wierzbicka and Harkins (Reference Wierzbicka, Harkins, Harkins and Wierzbicka2001) and Pavlenko (Reference Pavlenko2008), who point out that emotion concepts may not overlap completely in different languages or cultures. In the Japanese context specifically, Imada (Reference Imada1989) demonstrates that the English and the Japanese notions of ‘anxiety’, ‘fear’ and ‘depression’ differ, with fuan ‘anxiety’ being closer to yūutsu ‘depression’ than to kyōfu ‘fear’, while anxiety and fear are more similar than anxiety and depression (Imada, Reference Imada1989, p. 12).
2.4. Computational approaches to shame and guilt detection
In the field of NLP specifically, despite the increasing prevalence of sentiment analysis and emotion detection, ‘shame’ and ‘guilt’ are somewhat underexplored. The few studies we did find (e.g., Adoma et al., Reference Adoma, Henry and Chen2020; Meque et al., Reference Meque, Hussain, Sidorov and Gelbukh2023) tend to focus on ISEAR (International Survey on Emotion Antecedents and Reactions) as their main source of ‘shame’ and ‘guilt’ labels and mainly perform classification tasks on the data, that is, the focus is on the classification task where automatic detection of ‘shame’ and ‘guilt’ is attempted, not on how ‘shame’ and ‘guilt’ are expressed. In the adjacent field of corpus linguistics, there is some prior research that employs multivariate corpus methods to investigate cross-linguistic use of ‘shame’ and ‘guilt’ (see Krawczak, Reference Krawczak, Novakova, Blumenthal and Siepmann2014a, Reference Krawczak2014b, Reference Krawczak2018; Krawczak & Badio, Reference Krawczak and Badio2015). The present work builds on these studies but presents a relatively novel methodological perspective (a combination of corpus and NLP methods) and focuses on a different type of data (online written forums).
Other researchers have investigated adjacent moral emotions linked to social control, such as condolence and empathy in online communities (Zhou & Jurgens, Reference Zhou, Jurgens, Webber, Cohn, He and Liu2020) and hope and regret detection (Sidorov et al., Reference Sidorov, Balouchzahi, Butt and Gelbukh2023). The former concluded that online and in-person engagement with condolences and empathy, in general, were based on quite different social clues, suggesting that posting about struggles online is about seeking positive reinforcement rather than ‘comments that require emotional effort to engage with complex emotions’ (Zhou & Jurgens, Reference Zhou, Jurgens, Webber, Cohn, He and Liu2020, p. 617). Many studies have also tried to model or detect suicidal tendencies online, which often partly include ‘guilt’ or ‘shame’ as components or parameters in the detection model (see, e.g., Guidère, Reference Guidère2020).
Emotion detection, in general, is a highly active research field. From a purely NLP perspective, these approaches mostly aim to improve models in terms of accuracy metrics. Such models are, therefore, almost invariably based on supervised machine learning and tend to work only within a specific domain. However, it has been argued that more real-world congruent results with reusable methods can be obtained by using lexicons either independently (Teodorescu & Mohammad, Reference Teodorescu and Mohammad2022) or together with data-driven methods (Öhman, Reference Öhman, Hämäläinen, Alnajjar, Partanen and Rueter2021). Öhman and Rossi (Reference Öhman and Rossi2023) use emotion lexicons to create affective word embeddings that allow them to create domain-specific models that take semantic shifts into account when attempting to use affect as a proxy for mood in literary texts.
3. Methods and aims
Building on these earlier studies, the present work combines NLP and corpus-based collocational methods to investigate how people talk about, negotiate, and eventually challenge shame and guilt and haji and zaiakukan in two online web forums. The working hypothesis is that they may denote slightly different experiences and concepts.
3.1. Why shame and guilt and haji and zaiakukan
The selection of the English search items was quite straightforward because shame and guilt are discussed at length in the psychology literature on emotions (e.g., Ekman, Reference Ekman1992; Sabiston & Castonguay, Reference Sabiston, Castonguay, Eklund and Tenenbaum2014; Tangney & Dearing, Reference Tangney and Dearing2002). The next step was selecting their Japanese translations. The Genius English–Japanese dictionary (6th edition) proposes the following possible translationsFootnote 4:
Among these, we selected items that have received some attention in the Japanese literature on emotions (e.g., Higuchi, Reference Higuchi2002; Inaba, Reference Inaba2009; Suzuki, Reference Suzuki2007), that is, zaiakukan for guilt, and haji and shūchishin for shame. A search on the web corpus JaTenTen11 revealed that haji (96,201) is overwhelmingly more frequent than shūchishin (16,794), hence the former was preferred.
Importantly, we understand that experiences of ‘shame’ and ‘guilt’ can be verbalised in a multitude of ways that go well beyond the explicit use of emotion talk, let alone two sets of specific nouns. In Section 4.1, we attempt to incorporate as many such expressions as possible into our work by utilising word embeddings. This method allows us to find words and phrases that are closely semantically related.
3.2. Research questions
Our primary research question is (1) What are the main similarities and differences between the experiences verbally labelled as shame and guilt in English and haji and zaiakukan in Japanese? We also touch upon the more specific questions: (2) Who feels ‘ashamed’ and ‘guilty’ and for what? (3) Do people differentiate between these two experiences?
3.3. The data sources
The English data, amounting to 115,582,531 tokens, come from Reddit (https://www.reddit.com/), a predominantly North American pseudo-anonymous online discussion forum with 52 million regular users. Our data were extracted from the relationship_advice subreddit that centre on the topic of relationships.
The Japanese data are more restricted in size, amounting to 1,137,135 tokens after segmentation. They come from Hatsugen Komachi (https://komachi.yomiuri.co.jp/), which has been operated by the Yomiuri Newspaper (one of the largest newspapers in Japan) since 1999. Hatsugen Komachi, which literally means ‘little town of speech’, is a forum for dare ni mo kikenakatta onna no nayami (lit. ‘I couldn’t ask anyone’; worries of women’) originally addressed exclusively to women. According to Yahoo! Japan, it averages 2,000 posts a day and more than 100 millions monthly page views, and a much broader user base than when it was first launched. The fact that, despite its popularity, it has not received much attention from scholars to date possibly attests to resistance in Japan to approaching online forms of data in a scientifically adequate way (Miyake, Reference Miyake2022). The present study investigates online forums as scientifically adequate sources of linguistic data.
3.4. Tools for the data collection and analysis
Although we focus on shame and guilt and their Japanese counterparts, these emotions are not always expressed using these specific words. Previous research suggests that the related terms embarrassed and ashamed are also used (Krawczak, Reference Krawczak2014b). In Japanese, negative emotions are rarely expressed directly and, similarly to English, hazukashii ‘embarrassing’ is also a frequently used emotion term (Farese, Reference Farese2016). Consequently, we utilise word embeddings to computationally extract words and expressions that are used in semantically similar contexts to shame, guilt, haji, and zaiakukan.
Word embeddings, or semantic vector space models, are shallow neural networks that reconstruct the linguistic context of words as vectors by iterating over a corpus to learn associations between words and mapping semantically similar words to geometrically close embedding vectors (Mikolov et al., Reference Mikolov, Chen, Corrado and Dean2013). Cosine similarity provides the angle between two vectors and is the most used similarity measure for word similarity calculations (Sidorov et al., Reference Sidorov, Gelbukh, Gómez-Adorno and Pinto2014). Such approaches have been previously used within a usage-based linguistic framework (specifically Pankratz & Van Tiel, Reference Pankratz and Van Tiel2021). We use the collected data to build language- and context-specific vector space representations and examine which words and expressions are semantically closest to the keywords by using cosine similarity measures and word ngrams. This allows us to find many more examples of how ‘shame’ and ‘guilt’ have been expressed beyond the words themselves, including periphrastic expressions. We also employ manual evaluations of the results to ensure they are robust and not random (Antoniak & Mimno, Reference Antoniak and Mimno2018; Pierrejean & Tanguy, Reference Pierrejean, Tanguy, Nissim, Berant and Lenci2018).
The orthography of the Japanese language presents additional challenges as NLP applications rely on tokenisation and lemmatisation of words (i.e., splitting up the text into word units in their base form), a process which is fairly simplistic and achieves near-perfect accuracies in English with the tools that exist today. However, as there are no spaces separating word-like units, Japanese texts first need to be segmented into ‘words’ – a concept that is hard to define in any language but can be largely ignored in languages that have rules for where to place break spaces (Grefenstette & Tapanainen, Reference Grefenstette and Tapanainen1994; Papandropoulou & Sinclair, Reference Papandropoulou and Sinclair1974). Many different segmentation tools exist, but they all output slightly different segments based on different logic, making comparisons between lexical items a challenge. Issues with segmentation also affect the data at hand in that when zaiakukan 罪悪感 ‘guilt’ is present in a text, it is split into zaiaku 罪悪 ‘crime’ and kan 感 ‘feeling’ (cf. Figure 3 and Table 2). However, in our sample, zaiaku (n = 24) appears almost exclusively as part of zaiakukan (n = 23). This justifies the choice to focus on zaiaku for the word embeddings (considering the segmentation issues mentioned earlier), whilst the concordance analysis illustrated in Section 4.2.6 examines zaiakukan.
4. Results
4.1. Vector space representations
Figures 1 and 2 show a map of the semantic vector space of the seed words (shame, guilt, embarrassment, regret, remorse) and their most semantically similar words as measured by cosine similarity. Embarrassment, regret and remorse were added as separate keywords after looking at the most similar words of shame and guilt. This approach suggests that the semantically most similar expressions to shame are related to disgrace, dishonor and embarrassment, whereas guilt-related expressions are more closely related to fault, conscience, culpability, (sincere) remorse, but also grief and sorrow.
Words semantically similar to shame and guilt in English.
Words semantically similar to shame, guilt, embarrassment, regret and remorse in English.
Figure 3 shows a map of the semantic vector space of the target words in Japanese. Loss of face (oime), distrust (fushin), and regret (ushirometa*) are linked to zaiaku 罪悪, while embarrassment (hazukashii) and dishonour (akahaji) are more related to haji 恥.
Words semantically similar to haji and zaiaku in Japanese.
Tables 1 and 2 refer to Figures 2 and 3, respectively. Note that they were pruned to improve legibility by excluding spelling variations (e.g., embarasing/embarrasing/embarrassing/embarassing) and different verbs (kaka かか, kakasu かかす, kakukara かくから, kakisute かきすて, etc.) and adjectives (hazu 恥ず, hazukashii 恥ずかしい) forms. Some English compound terms (e.g., deeply_regrets, deeply_regretted, deeply_regret, expressed_regret), Japanese morphemes with no clear meaning on their own (sukashi スカシ, kashii かしい rabokku ラボック, i イ) and idiosyncratic uses employed, for example, as part of manga or TV shows titles (kōai 肛愛, shūchū 羞中, etc.) were also excluded from the tables but can be seen in the vector space visualisations.
The semantically closest matches to keywords in English
The semantically closest matches to keywords in Japanese
Word embeddings are complemented by the corpus-based analysis of texts, which was carried out with the tools offered by Lancsbox (Brezina et al., Reference Brezina, Weill-Tessier and McEnery2020), a recently developed software package for the analysis of language data and corpora. We used collocation analysis to access meanings that recur across different pieces of text. In the attempt to access ‘non-obvious general semantic preferences’ (Partington, Reference Partington2004, p. 164) and picture what the semantic space of ‘shame’ and ‘guilt’ may look like in the two linguacultures, the analysis is not limited to first-order collocates (i.e., the collocates of our search items), but we have further travelled ‘the collocational network’ (Marchi, Reference Marchi2023). When relevant, we accessed the extended concordance lines to zoom in on specific and often highly context-dependent linguistic constructions. The findings shed light on some frequent and meaningful patterns that collocate with shame/haji(-related) and guilt/zaiakukan(-related) expressions.
4.2. Collocational and concordance analysis
A preliminary examination of the data using corpus tools indicated a notable difference in the frequency of the search words. For instance, in the English corpus, occurrences of guilt (9,894) were more numerous than those of shame (6,363). Contrastingly, in the Japanese corpus, haji (representing ‘shame’) appeared 560 times, surpassing the occurrences of zaiakukan (representing ‘guilt’), which appeared only 23 times. However, these figures should be approached with caution. Raw frequency counts do not necessarily equate to linguistic or cultural prominence and can be influenced by various factors unrelated to the emotional salience of these terms. Additionally, our findings seem to diverge from external sources such as the COCA database, where shame appears more frequently than guilt. This discrepancy highlights the importance of considering multiple data sources and methodologies when examining linguistic phenomena. Furthermore, the relevance of these occurrences in relation to the emotional states they represent must be carefully considered. Not all instances of shame and guilt in the English corpus may pertain directly to the emotional states. Therefore, a more in-depth analysis that discerns the context of each occurrence is necessary to draw meaningful conclusions about the emotional landscape in each language.
Concerning the Japanese data, while haji appears more frequently than zaiakukan, this result alone is insufficient to conclude definitively about the emotional landscape in the Japanese linguaculture. It is a preliminary observation that suggests a potential avenue for further research rather than a conclusive statement. With these caveats in mind, the following sections investigate and compare the collocational network of our search words in the two samples.
4.2.1. Semantic space of shame and guilt
The semantic space of shame and guilt as verbalised by their collocates, and the way they relate to each other in our sample, is illustrated in Figure 4. Drawing from Sinclair, we define collocates as ‘the occurrence of two or more words within a short space of each other in a text’ (Reference Sinclair1991, p. 170). For the scope of the present paper, such short space corresponds to five words on each side of the node. The statistical measure employed for visualising the results is logDice, which indicates the tendency of two words to occur exclusively in each other’s company (Brezina, Reference Brezina2018). This measure is not affected by the size of the corpus, hence can be used to compare co-occurrence across corpora, and is particularly useful to highlight the cumulative forces of discourse representations (Brookes & Chałupnik, Reference Brookes and Chałupnik2022, p. 4). We included collocates with a logDice of at least 7, a score that returns statistically significant results (Brezina, Reference Brezina2018) and prevents over-populated graphs.
L5R5 collocates of shame and guilt in the Reddit corpus. LogDice value cut-off <7.
In the figure, the closer the collocate is to the node word, the stronger their association. The frequency is indicated by the intensity of the colour of the collocate, and the position of the collocate in the semantic space is the actual position where it appears in the texts, either at the right or the left of the word, or a mix of the two. Tables 3 and 4 report the values associated with each collocate in more detail.
L5R5 collocates of shame in the Reddit corpus (logDice ≥ 7)
L5R5 collocates of guilt in the Reddit corpus (logDice ≥ 7)
Shame and guilt are polysemous words, hence Figure 4 and Tables 3 and 4 include items that are not necessarily related to ‘shame’ and ‘guilt’ as emotions. We tracked down the contextual meanings of the collocates by zooming in on the text. In what follows, we focus on those collocates that the close reading of concordances revealed to be relevant to shame and guilt as emotion terms.
Among the five most typical collocates of guilt, we have shame and free, at the third and fifth position, respectively, while the remaining three are different forms of the lemma trip. Free is employed in the idiomatic construction guilt-free, which is not as relevant as other patterns for the scope of the paper, whilst the co-occurrence of shame and guilt is further addressed below. We shall now turn to trip* in combination with guilt. The most typical collocation is guilt trip (logDice 10.1, freq. 735) in the construction SBJ guilt trip OBJ, as in the following example (all examples are reproduced faithfully to the original, including non-standard spellings, punctuation and so on):
The second most typical collocation is guilt tripping (logDice 9.9, freq. 299), followed by guilt trips (logDice 9.1, freq. 197). These values testify to the saliency that this construction and the phenomenon it indexes have among the English-speaking users of Reddit, who conceptualise ‘guilt’ as something that can be forced on others. Notably, previous studies argued exactly the opposite, saying that while ‘shame’ can be imposed on us by others because it is based on a socially constructed identity, ‘guilt’, which originates in issues of responsibility, cannot (Bedford & Hwang, Reference Bedford and Hwang2003, p. 128). Our findings seem to suggest that there may be a gap between academic notions of ‘guilt’ and how it is conceptualised by the layperson.
As Table 3 shows, shame has a more diversified list of collocates, at least if we look at the five most typical items, which are fool, guilt, twice, slut and embarrassment. Embarrassment typically co-occurs with shame, but not with guilt. This is in line with what was observed in Section 4.1 and with the assumption that embarrassment is a ‘low level version of shame’ (Barrett, Reference Barrett2005, p. 955). In other words, although not conclusive, these patterns of co-occurrence suggest that the two emotions are qualitatively similar but differ in degrees of intensity, with ‘shame’ being perceived as more intense and destructive than ‘embarrassment’.
The use of derogative expressions such as fool and slut testify to the public nature of shame and are worthy of further examination. However, a closer look at the co-text showed that fool in collocation with shame is used almost exclusively in the idiomatic expression fool me once, shame on you, fool me twice, shame on me. It does not refer to the emotional experience under analysis, hence it is of little interest to the study. Slut revealed a more interesting usage, that is, slut shame (employed in 88 out of 93 concordances):
Although in these examples we do not have personal accountings of emotional experiences, the salience of this construction in the Reddit corpus indexes a socio-cultural phenomenon that sheds light on potential causes of ‘shame’ (here ‘shame’ arising is due to other people’s judgments of sexual behaviour) and highlights its social nature.
Going back to the list of collocates, remorse typically collocates with guilt but not with shame. This mirrors the tendencies illustrated in Figure 1, where remorse is closer, hence semantically more similar, to guilt than shame. Finally, another striking feature is the high degree of correlation between shame and guilt. As illustrated in Tables 3 and 4, with 337 instances of co-occurrence, shame is the third most typical collocate of guilt and, conversely, guilt is the second most typical collocate of shame. A closer look at the concordance lines showed that they tend to co-occur in the construction shame and guilt (or vice versa). The conjunction and suggests that what is projected as shameful is likely to arise (or is functional to the projection of something as worthy of) guilt, and vice versa. It follows that shame and guilt are indeed related in English but are by no means the same thing – or there would be no need to distinguish between them (Baker et al., Reference Baker, McEnery, Hardie, Pace-Sigge and Patterson2017, p. 47). This is in line with the vector space representations illustrated in Section 4.1, which demonstrated that, despite overlap between the two, there are components that are closer to either shame or guilt (e.g., regret), and that allow English speakers to differentiate between them.
4.2.2. Semantic space of haji and zaiakukan
The collocation analysis of haji is visualised in Figure 5. Translations and transcriptions are provided in Table 5, together with the statistical and frequency values associated with each collocate. Note that in the figure, the cut-off value is 8, and not 7 as elsewhere in the paper, otherwise the graph would be very difficult to read. For comparative purposes, however, the cut-off value considered during the data analysis is 7. The full list of collocates (not reported here for reasons of space) can be accessed through the link in the Data Availability Statement.
L5R5 collocates of haji in the Hatsugen Komachi corpus. LogDice value cut-off <8.
L5R5 collocates of haji in the Hatsugen Komachi corpus (logDice ≥ 8)
The analysis shows that haji is indeed a social emotion (Shott, Reference Shott1979), because a number of its collocates can be traced back to interpersonal relationships or society in general. For instance, there is a strong collocation (logDice 9.6) with bunka ‘culture’ in the construction haji no bunka ‘culture of shame’, showing that this second-order classification (Raz, Reference Raz2002; Sakuta, Reference Sakuta1967) is something that laypeople talk about. Of note are also rikon ‘divorce’ and danna ‘husband’ (logDice values are 8.3 and 7.7, respectively), signalling that marital relationships are a recurrent topic among members of the community and, more importantly, that they are often associated with ‘shame’. Finally, hitomae ‘in front of people’ (logDice 8.0) mirrors the English publicly (see Table 3), showing that in both samples, ‘shame’ is linked to the self in relationship with others (Tangney & Dearing, Reference Tangney and Dearing2002). These findings provide linguistic evidence for the assumption that ‘shame’ lies ‘at the intersection of subjective experience of one’s own self and inter-subjective sensitivity to the social reality and the self’s presence therein’ (Krawczak, Reference Krawczak2014b, p. 442).
As for zaiakukan, its low frequency (23) does not allow for big generalisations and no collocate with a logDice equal to 7.0 or higher was found in the corpus. Such a low frequency, however, may be taken to signal that among the users of Hatsugen Komachi ‘guilt’ may not be as salient as ‘shame’, both as a personal psychological phenomenon and as an interpersonal social one. Differently from what was observed in the English data, haji and zaiakukan are not mutual collocates, a factor suggesting that in the Japanese data ‘shame’ and ‘guilt’ do not overlap as they seem to do in English. Another possibility, however, is that zaiakukan is not the best candidate to access ‘guilt’ in texts. Future studies may use the terms listed in Table 2 to corroborate or falsify these preliminary findings.
4.2.3. Guilty
In the attempt to locate what warrants ‘guilt’ in the English data, we looked at the adjective guilty (11,602 occurrences, available through the link in the Data Availability Statement). First, we restricted the analysis to words that are used to the left of the node (span L1–L5) to identify the emoter, that is, who feels guilty. This time, we looked at raw frequencies of collocations because statistical measures tend to eclipse functional words that are nonetheless relevant to the aims of this section. The findings revealed that in our data you is the most frequent pronoun observed in the L1 position (3,100), followed by I (1,912), she (1,190) and he (806). This suggests that the emoter and the receiver of the utterance often overlap. We then asked what people feel guilty for, which prompted us to travel to the collocational network and investigate second-order collocates of guilty for. We restricted the analysis to positions R1-R2 on the assumption that what people feel guilty for is likely to appear immediately to the right of the node. The most frequent collocate found within these parameters is not (196) in the construction guilty for not, which suggests that people feel guilty for not doing/having done something. This is best illustrated with examples:
Of note, it is also the fact that users tend to adopt linguistic strategies to distance themselves from the negatively evaluated action. This was achieved, for instance, by using the indexicals it (93), that (66) and what (52) so as not to mention what triggered the emotion, as in:
Attention to the wider co-text is key to identifying what the producer was referring to. Future studies may provide a more comprehensive close reading of concordances to shed further light on what people feel guilty for.
4.2.4. Ashamed
Following the same procedure, a second collocation analysis showed that the personal pronoun most frequently used to the left of ashamed is you (813), followed by I (477), he (384) and she (368). We can thus assume that in our sample, the internal states verbalised as guilty and ashamed are more strongly associated with the receiver of the utterance. Triggers of shame were tentatively accessed by looking at the words and expressions employed immediately to the right of the construction ashamed of (R1–R2). The analysis of second-order collocates of ashamed of revealed that it (196) and you (187) are the two most frequent collocates in this position. Your (94) and yourself (87) follow in fourth and fifth position respectively. Importantly, these observations provide descriptive insights into the sample but should not be overgeneralised. The patterns observed may be specific to the corpus and not necessarily indicative of broader linguistic or cultural trends. The close reading of concordances revealed an additional reason for caution, showing that these collocational patternings tend to be preceded by a negative clause + be or feel. For instance, in 31 out of 196 occurrences of ashamed of it, the immediately preceding co-text reads don’t be, shouldn’t be or nothing to be. Similar patternings apply also to you and yourself. Clearly, then, even when we have it or you at the right of ashamed of it does not necessarily mean that someone is experiencing ‘shame’, but the utterance may have a supporting function (as in you shouldn’t be ashamed of it).
Another recurrent pattern at the right of the node and related to the collocate yourself is the construction ashamed of my-/your-/him-/her-/them-selves (271), as in
Similarly to what was previously observed for the collocation guilty for, here the specific behaviour that triggered shame is not mentioned, and the producer seems to distance themselves from it using the indexical that. However, in contrast with example (4), in what follows in the utterance, the producer expresses a negative evaluation whose object is their whole self, not what they have done.
In discussing the causes of emotions like ‘shame’, as exemplified in phrases like ashamed of myself, we recognise that such instances offer valuable insights but are not sufficient for broad generalisations. Despite this caveat, our perspective aligns with the empirical investigations in Krawczak (Reference Krawczak2017, Reference Krawczak2018), which adopt different but complementary methods within corpus linguistics and provide a more comprehensive analysis of these complex emotional constructs across three languages and cultures.
4.2.5. Haji
The collocational analysis of the Japanese data mirrors the one carried out in English. In Japanese personal pronouns are often omitted and, in line with this, there were no personal pronouns among the collocates of haji or zaiakukan, hence collocational analysis did not reveal whether the emoter is the producer of the utterance, other participants in the thread, or general third parties. As for the behaviours and events people are ashamed of, they were accessed by restricting the analysis of concordances to instances where haji is immediately followed by the copula da (informal) or desu (formal) (freq. 39), a construction that can be roughly translated as ‘it is shame[ful]’. The choice to focus on concordances instead of collocates is motivated by the relatively low frequency of this two-word collocation, which allows for the manual annotation of the data.
After manually removing instances where haji is employed in a negative clause (e.g., haji da to wa omoimasen ‘I don’t think it is shame[ful]’) and invalid examples, we annotated the remaining 30 concordance lines according to what triggers the emotion. The coded concordance lines can be accessed through the link in the Data Availability Statement. Based on semantic similarities, six groupings of triggers of haji and zaiakukan were identified, namely Relationships (11), Identity/Personality (9), (Lack of) knowledge (4), Money (3), Deception (1) and Sexuality/Body (1). In what follows, we illustrate representative examples of the first two.
This taxonomy does not aim to be exhaustive, very often the groupings are blurred and topics overlap; therefore, the chosen categories are subjective and reductive. We are also aware that 30 concordances are way too few to provide considerations that are generalisable beyond the data set at hand. Despite these limitations, however, a more qualitative approach shed light on triggers of haji, showing that in a society where cis-heterosexual marriage is recognised as normative and proper, haji can be triggered by being, or considering the idea of getting, divorced (as exemplified in (6) and corroborated by the collocational analysis illustrated in Figure 3), or by aspects related to the identity or personality of the emoter.
4.2.6. Zaiakukan
Similarly to haji da, the low frequency of zaiakukan (23, which drops to 18 after removal of invalid examples) motivates the adoption of a more qualitative approach to its examination. The data were coded according to the same categories identified in the previous section, with the addition of the category Emotions: Relationships (8), Sexuality/Body (4), Deception (4), Money (1) and Emotions (1). By way of illustration, some examples are provided below.
The low numbers preclude any kind of statistical analysis, and these preliminary findings should be viewed merely as pointers for future research. Nonetheless, they show some interesting potential differences with haji: explicit references to the emoter identity and/or personality do not trigger zaiakukan in the data, which seems to be relatively frequently warranted by sexuality-related matters, such as being in a same-sex couple or cheating on the partner.
5. Discussion and conclusions
In this section, we elucidate patterns from the varied observations reported so far and address our research questions (RQs). We started from the text and the recurrent structures observed in it, based on which we proposed a (tentative) semantic representation of shame and guilt in English and haji and zaiaku(kan) in Japanese as they emerge in our samples. Rooted in the principles of usage-based linguistics, this approach posits that key elements of meaning representation are acquired through the statistical distribution of words across various texts. Methodologically, this means that empirical observations of statistically significant lexical patterns are crucial in illuminating cognitive and psychological processes. This perspective is grounded in the understanding that language is inherently tied to its use in real-life contexts. Consistent with usage-based theories, it is assumed that individuals’ cognitive and linguistic experiences are shaped by the frequency and context of language exposure (Divjak, Reference Divjak2019), a concept known as priming (Hoey, Reference Hoey2005). This reflects the core usage-based tenet that language structure and function are deeply interconnected with how language is habitually encountered and processed by individuals. In our hypothesis, ‘shame’ and ‘guilt’ can be illustrated as contiguous and overlapping semantic spaces, where certain expressions would be closer to certain elements than others or shared across elements. Areas of overlap between semantic spaces can thus be encoded in terms of shared collocates or shared semantic sets.
Semantic vector space representations revealed that shame is semantically close to disgrace, dishonor and embarrassment, whilst guilt is more closely related to notions of fault, culpability, and (sincere) remorse. Our analysis also suggests that ‘shame’ is often portrayed as a public experience (Figure 4), and ‘guilt’ as an emotion that encompasses both private elements, akin to sadness, and public aspects, such as the motivation to openly acknowledge (Figure 1) a transgression (RQ1). This interpretation, however, requires careful consideration. Specifically, it remains unclear whether openly acknowledge is a construction unique to ‘guilt’ as an emotional state, or if it also emerges in contexts where ‘guilt’ pertains to legal situations. Further investigation is needed to disentangle these different usages. Areas of overlapping were also observed. For instance, the two emotions can co-occur (as indexed by the construction shame and guilt) and both can be imposed on others – despite previous research suggesting otherwise (Bedford & Hwang, Reference Bedford and Hwang2003, p. 128). Moreover, within our sample, you is the most frequent collocate for both guilt and shame. Similarly, both the adjectives guilty and ashamed tend to be associated with the receiver of the utterance (RQ2). Our collocational analysis, while illuminating, should be interpreted with caution, particularly regarding the implications for the types of emoters associated with ‘guilt’ and ‘shame’, due to the low frequency of data. Nonetheless, our findings suggest a nuanced interplay in the attribution of these emotions in discourse rather than a straightforward assignment of ‘guilt’ to the receiver and ‘shame’ to the producer of the utterance.
As for Japanese, zaiaku(kan) is semantically related to loss of face, trust and regret, whilst haji is closer to embarrassment and dishonour. The corpus-assisted close examination of haji and zaiakukan, whose low frequencies allowed for a more qualitative approach, also showed that they are both triggered by violations of interpersonal norms along the lines of their English counterparts. However, whilst in our data zaiakukan is often associated with sexuality-related matters, haji links more directly with the identity or personality traits of the emoter. Based on what was observed in our sample, we can then hypothesise that, in English as well as in Japanese, ‘shame’ and ‘guilt’ are correlated with different causes and that this, together with the features of meaning representations mentioned above, allows people to distinguish between the two experiences (RQ3). Clearly, however, it is difficult to arrive at firm conclusions, especially for the Japanese data where low frequency is an issue. Moreover, although the two data sources share a number of common features that increase their comparability, texts from Reddit focus on relationships, whilst Hatsugen Komachi has traditionally been focused on ‘women’s issues’. This may skew the results concerning the cause of the emotions and calls for further studies to corroborate, or falsify, the tendencies observed here.
Methodologically, this article illustrated a replicable process to access semantically similar expressions, where we first built language-specific vector space representations and then looked at their meanings in context with corpus linguistic tools. This is an innovative and effective way to further advance our knowledge of the language of emotions, showing the value of merging corpus methods and NLP. Emotions, however, have fuzzy boundaries, and their cognitive and social reality cannot be accessed by looking only at examples of explicit emotional labels. Future studies may employ the hybrid and corpus-based methodology proposed here to further explore ‘shame’- and ‘guilt’-related expressions both inter- and intra-linguistically.
Data availability statement
All data and materials can be found at https://osf.io/n8d5g/?view_only=f51077116e9846c7aa79f6ac0dd1670e, or at the first author’s OSF page.
Acknowledgements
We would like to thank our two anonymous reviewers and the editors of this special issue for providing invaluable comments.
Funding statement
This work was supported by JSPS KAKENHI under grant 22K18154.
Competing interest
The authors declare no competing interests exist.
Open access
