Emotions in the parliament: Lexical emotion analysis of parliamentarian speech transcriptions Egon Werlen1 Christof Imhof1 Per Bergamin1 egon.werlen@ffhs.ch christof.imhof@ffhs.ch per.bergamin@ffhs.ch 1 Swiss Distance University of Applied Sciences (FFHS) Abstract of videos is that a large part of the observable verbal and non-verbal signals of emotional states Politics is emotional. So far, relatively in political speeches (e.g. posture, gestures, few studies investigated the emotional facial expressions, phonology, speaking style) can content in parliamentary speeches. In this be traced. In transcribed form, these signals study, we analysed emotional valence and are no longer represented to the same extent. arousal of German and French speeches However, some emotional characteristics remain. of a Swiss cantonal parliament and These are primarily the emotional potential of whether we can use them to predict the words and other linguistic features like phonemes, membership of parliamentarians to one of accent, number of syllables and letters, and word two groups: those who won more of the frequency. This is where we come in with the votings than others. The emotional text present exploratory study. We want to find out analysis showed that these speeches are whether it is possible to estimate emotional states indeed emotional. However, the results on the two dimensions valence and arousal in regarding the predictions were mixed. literal transcripts of parliamentary sessions with Arousal and language showed no effects a rather simple lexical method. Further, we and valence was only partially successful intend to find out whether we can predict the as a predictor. parliamentary groups that lose more votings than the average of lost votings in the parliament by the emotional content of the speeches. We use 1 Introduction this subdivision into ’vote winners’ and ’vote Ever since it came into existence, politics has ex- losers’ as an analogy to the more common contrast erted influence on the daily life of humans all over between ruling party and opposition found in the world. For a long time, the idea prevailed that many other countries. This differentiation is found politics has to be rational rather than emotional. in many studies on parliamentarian speeches. In However, it is not surprising that many political Switzerland however, there is no classical division issues are emotional at their core. This leads to into governing and opposing party. Instead, the debates about very emotional topics, which are parliaments of Switzerland and its 26 cantons are not always handled as rationally as one might built on consensus, which is why another approach assume. Audible and visible evidence is provided was needed to differentiate between parliamentary by debates on the Internet and on television: groups. parliamentarians cheer, yell, throw things, and even have fistfights on rare occasions. Moreover, political campaigns often aim at emotionally 2 Theoretical background relevant aspects of political topics rather than the actual ramifications of the topic at hand (e.g. 2.1 Emotions Widmann, 2021; Erisen and Villalobos, 2014). Roughly classified, there are three basic paradigms Thus, politics is very emotional. Today, in modern in emotion research (Holodynski and Friedlmeier, parliaments (e.g. Switzerland, Germany, France, 2012). The first one is the structural emotion the UK, or the European parliament) verbatim paradigm (Izard, 1991; Panksepp, 1998; Ekman, protocols as well as videos are recorded and 1999) in which emotions are defined as specific used for tracking and archiving. The advantage mental states. In the second one, the functional Copyright © 2021 for this paper by its authors. Use paradigm, emotions are viewed as a set of spe- permitted under Creative Commons License Attribu- cific mental functions, defined as changes in the tion 4.0 Interna-tional (CC BY 4.0) disposition to act and help the individuals to ad- just their motives and intentions to the changes cesses, debates and written texts (Freeden, 2013). (for example: Frijda, 1986; Lazarus, 1991; Scherer, Political science often looks at things from the 1999). Under the third paradigm, the contextual functional paradigm perspective. Barbalet (1998) paradigm, emotions are defined as socially and cul- and Freeden (2013) assume that emotions are com- turally constructed psychological functions result- mon everyday processes. They influence political ing from interpersonal interactions (for example: thinking through three syntactic functions in that Lutz and White, 1986; Matsumoto et al., 2008). they (1) emphasise concepts by reinforcing mor- In general, it can be observed that many political phological structuring, (2) relativise meanings by studies follow the functional paradigm (e.g. Lara classifying importance, or (3) reduce or reinforce et al., 2016). connections. In their qualitative research, Lara From a different perspective, according to which et al. (2016) form functional categories in parlia- emotional feelings are sometimes expressed as emo- mentary discourses by assuming that emotions are tional colouring, further fundamental distinctions used to ”emphasise the speaker’s argumentation”, of theoretical approaches can be found. There are ”attack the opponent”, ”express proximity and cre- theories that start from different distinct emotions ate a distinctive ’identity’ with respect to the rest (e.g. joy, fear, anger, surprise). Well-known ap- of the group”, and emotions are also ”used as an proaches include the basic emotion theory of Ek- argument itself” (p. 155). man (1999) or the process component theory of emotions of Scherer (2010). The latter assumes 2.3 Emotion analyses methods for texts that every emotion consists of five components In order to measure emotions in speech and text, (cognitive, physiological, motivational, motor ex- an analytical framework is first needed that helps pression, subjective feeling). Other theories as- to reduce the number of categories (Cowie and Cor- sume that emotions are based on two or three di- nelius, 2003). In the present study we have cho- mensions with high and low emotional levels. For sen to describe emotions based on the circumplex example, Bradley and Lang (1994) postulate three model of Barrett and Russell (1999) with its two- dimensions: emotional value, emotional arousal dimensional classification of emotions (valence and and emotional dominance. Another very promi- arousal). Furthermore, we use a lexical approach nent representative of this approach is Bertrand based on individual words. From a technical point Russel with his collaborators (Barrett and Russell, of view, word-based lexical analysis can be classi- 1999). In the Circumplex model two emotional fied as a semantic approach to sentimental anal- dimensions are postulated, namely the emotional ysis, but it does not necessarily implement ma- valence and the emotional arousal. Valence refers chine learning. This type of approach is histor- to the experience of one’s own positive or nega- ically based on early work by Freud (1891) and tive feelings. Arousal refers to the experience of Bühler (1934), who assumed that spoken or writ- the intensity, the activation level of one’s own feel- ten words have the potential to elicit both overt or ings. Both dimensions form the ”core affect”, as covert sensu-motoric or affective reactions. From ”the most elementary, consciously accessible affec- this point of view, words can evoke both basic and tive feelings, which do not have to be directed at induced emotions (Jacobs et al., 2015). anything” (p. 806). Lexical analysis usually relies on word lists, con- sisting of thousands of words whose values (e.g. 2.2 Emotions in politics valence, arousal, dominance etc.) were previously In formal discourses, such as parliamentary validated as the result of rating procedures. Ex- speeches, one assumes that fewer emotions are amples of such lists are the Affective Norms for expressed, compared to everyday conversations. English Words (ANEW; Bradley and Lang, 1999, Day-to-day conversations seem to offer more imme- the Warriner list of norms for valence, arousal and diacy and closeness and thus stimulate the expres- dominance for English lemmas (Warriner et al., sion of emotions (Lara et al., 2016). Historically, 2013), the NRC-VAD lexicon (National Research emotions have been part of public and political life Council Canada - Valence, Arousal, Dominance; as in the case of the Greeks, Machiavelli or Hume. Mohammad, 2018), the Berlin Affective Word List Throughout the 20th century, however, emotions (BAWL-R; Võ et al., 2009), the Semantic Lexicon were not considered important in politics and so- of Emotion (SLE; Leleu, 1987) or the French in- cial life. This changed in the 1990s, when interest terlingual metanorm for the emotional analysis of in human emotions grew in various disciplines such texts (EMONORM; Leveau et al., 2012). In many as psychology, neuroscience, sociology and philoso- cases, the emotional valence and arousal of texts phy. This led to the rediscovery of emotions in po- is calculated by averaging the values for valence litical science (Hoggett and Thompson, 2012) and and arousal of all words contained within. How- the systematic use of emotions in democratic sys- ever, values can also be derived for smaller units tems, for example, by politicians in election pro- such as sentences or paragraphs. Such a procedure has been used in the context of political studies in 2.4 Studies of emotions of transcribed the analysis of ”emotional conversations” by Lara parliamentary speeches et al. (2016) or the analyses of emotional words by The number of studies that analyse parliamentary Koschut (2020), to name two examples. speeches for their emotional content is growing but The BAWL-R is the largest German emotional still limited. Abercrombie and Batista-Navarro word list and has been utilised for the analyses of (2020) reviewed 61 studies, 28 were looking for sen- different text forms: poems (Aryani et al., 2016; timent polarity and three for emotions; 16 worked Ullrich et al., 2017), E.T.A. Hoffmann’s black- with dictionary based methods. The same goes romantic story ”The Sandman” (Lehne et al., for studies establishing a relationship between ex- 2015), passages of Harry Potter novels (Hsu et al., pressed emotions in the speeches and the role of the 2015), Shakespeare’s sonnets (Jacobs et al., 2017), parliamentary group (governing or in opposition). and short stories (Werlen et al., 2018, 2019). In Abercrombie and Batista-Navarro (2020) found 14 all these studies, the mean of the affective values studies predicting some form of party affiliation. of the individual words correlated with the whole text ratings. Studies implementing the BAWL-R One example is a study by Riabinin (2009), who to predict subjective emotional states of short texts classified politicians in the Canadian Parliament (Hsu et al., 2015) and poems (Ullrich et al., 2017) based on the dimension Liberal vs. Conservative found correlations for lexical valence with subjec- with a Support Vector Machine using the cate- tive valence of r = .53 and r = .65, and for lexi- gories of the Linguistic Inquiry and Word Count cal arousal with subjective arousal of r = .59 and (LIWC) by Pennebaker et al. (2015). The authors r = .54. The SLE was validated by Leleu (1987) used the Canadian Hansard, which includes the and was implemented in experimental studies (e.g. English and French House of Commons debates. Degner et al., 2012; Jhean-Larose et al., 2014). De- One might assume that the expression of posi- spite there not being a similar comparison between tive (empathy) or negative emotions (contempt) lexical and subjective values as in the case of the was connected to these specific political ideologies BAWL-R, the SLE is relevant to this study be- (see Freeden, 2013), which appeared to be the case cause it is the only French word list we are aware in this study, at least at face value: the authors of that includes words rated on both emotional va- found that in the speeches of the 36th Parliament, lence and arousal. the Liberals generally used positive language, while An alternative could be the NRC-VAD lexicon by the Conservatives used more negative words. How- Mohammad (2018). This lexicon contains 20,007 ever, they suppose that this difference is not due to annotated words in 103 languages. The English party affiliation, but rather the fact that the Liber- words were annotated with the help of Amazon als were the governing party and the Conservatives MTurk for valence, arousal, and dominance using were in opposition. Hirst et al. (2014) conducted the best-worst scaling method. The translation of the same analysis with the speeches of the 36th the English words into the other languages was ac- Parliament, but added the ones from the 39th Par- complished by using Google Translator. The val- liaments as well, where the roles were switched. In ues for valence, arousal and dominance were taken both cases, the respective opposition showed more from the English version on the assumption that negative emotions in its speeches than the govern- the values are stable for different languages. In ing party, which the authors concluded was due to an unpublished study, we compared the NRC-VAD a ”language of attack and defence” (p. 93). The with the BAWL-R in an emotional text analysis of differences due to political ideology or party affili- 62 short stories in German and their English trans- ation were thus negligible, confirming the assump- lation. The English version of the NRC-VAD cor- tion by Riabinin (2009). In this context, it should related with the human ratings of the English texts be noted that the authors of both studies used par- to a similar extent as the BAWL-R correlated with tially translated speeches, as the Canadian Parlia- the human ratings of the German texts. However, ment is bilingual. The French speeches were first in the German version of the NRC-VAD, the cor- translated into English before the analysis. The relation values with the human ratings of the Ger- bilingualism of the speeches and the subsequent man texts were considerably lower than with the translation may therefore have had an influence on BAWL-R. Consequently, the German translation the results. of the NRC-VAD lost some of its predictive power. Another example is a study by Rheault et al. For this reason, we decided not to use this large (2016), where the British Hansard was used, which lexicon, even though it contains both languages of includes the transcripts of all parliamentary de- interest to us. bates of the British House of Commons between 1909 and 2013. To analyse emotional polarity as a standardised measure from -1 (negative) to +1 (positive), they created a domain-specific lexicon based on the affective content of expressions to ob- tain an indicator of emotional words in the British Research questions Parliament. The mood of politicians of the British parliament was found to having become more pos- 1. Do parliamentary speeches contain emotional itive during the last decades, and the valence of information (valence, arousal)? the politicians’ speeches fluctuated in accordance 2. Are there differences in the emotional state of with economic business cycles (e.g. indicator of speeches between parliamentarian groups that recession, and indicator of labour conflicts). lost more votings compared to groups that lost To our awareness, there are no studies on emo- fewer votings? tional arousal in parliamentary speeches. Hypotheses 2.5 Research questions and hypotheses 1. Speeches by members of parliamentary groups The overall goal of this study is to replicate the re- with fewer lost votings indicate more positive sults of the studies analysing speeches of the Cana- emotional states than speeches from members dian and British parliaments and to extend them. of parliamentary groups that lost more vot- As shown in the abovementioned studies, it is ings. possible to estimate emotions in parliamentary 2. Speeches by members of parliamentary groups speeches. All of them estimated positive-negative with fewer lost votings indicate less arousal emotional states that generally correspond to the than speeches from members of parliamentary emotional valence of the circumplex model (Bar- groups that lost more votings. rett and Russell, 1999). We intend to extend these results by measuring not only emotional valence, but emotional arousal as well, the second dimen- 3 Methods sion of the circumplex model of emotions. There- fore, the first research question concerns our abil- 3.1 Samples and measurements ity to estimate emotional valence and emotional For the analyses and the testing of the hypothe- arousal in parliamentary speeches with our emo- ses, we used all the transcribed speeches from three tional text analysis approach. sessions, which each occurred within a week in the The transcribed speeches we analysed stem month of June, September, and November 2019 of from a cantonal parliament in Switzerland. The a Swiss cantonal parliament (Valais). The parlia- political systems of Switzerland and its cantons do ment includes 130 parliamentarians and 130 sub- not have a typical government - opposition struc- stitutes. The speeches of the government rep- ture. On first glance, this poses a problem for our resentatives (i.e. the five members of the can- replication in light of the results presented above: tonal council) and the president of the parlia- the prediction of party affiliation or ideology ment were not included in the analyses. The by emotions in the speeches of parliamentarians president of the parliament leads and moderates is, as the study of Hirst et al. (2014) shows, the debates but does not usually contribute to confounded with the division into government their content and the cantonal council members and opposition rather than political ideology. are not part of the parliament. The parliamen- In the parliament we analysed, there is no true tary speeches are automatically transcribed by the opposition since most parliamentary groups are company recapp IT AG (https://recapp.ch) us- represented in the government. Therefore, the ing AI algorithms. The transcripts are checked definition of an opposition cannot refer to the by the administration, corrected and formatted, parliamentary groups alone. Thus, we chose a including the insertion of the agenda items and different operationalisation approach: We exam- other notes such as information about beginning ined the proportion of lost votings during the and end of each session. The literal minutes are three session weeks that we analysed and the published in the original language on the can- groups that lost more votings were thus defined tonal website (https://parlement.vs.ch/app/ as the oppositional groups. According to Riabinin de/search/result?object_type=ParlSession). (2009), in a parliament with a real opposition, In order to categorise the parliamentary groups, one would assume that the opposition would show we first calculated the percentage of won and lost more negative emotions in their speeches. Since votings of all groups during the three sessions, con- more negative emotions are usually associated sisting of 20 half days. The parliament voted 196 with higher arousal (Kuppens et al., 2013), the times, without counting the issues that were un- opposition would also show more arousal in their controversial and did not lead to a vote. The 257 speeches. We assume that these correlations are parliamentarians - present at least at one voting - also present with our operationalisation of the cast a total of 22’963 individual votes. opposition as groups with more lost votes. Since speeches are usually given in the mother tongue of the speaker, in this case German or French, we opted for analysing the original speech to calculate the percentage of lost votings and contents with language-specific word lists. For the agreement with the parliamentarian group using German speeches, we used the revised form of the R (Core Team, 2017). We downloaded the PDF Berlin Affective Word List (BAWL-R; Võ et al., files containing the speeches from the file sever of 2009), while the Semantic Lexicon of Emotions the canton with a custom Python script, which (SLE; Leleu, 1987) served as the word list for the also served the purpose of immediately splitting French speeches. In total, we analysed the speeches the text body based on individual speeches. The of 179 parliamentarians from all nine parliamen- resulting files were subsequently further processed tary groups. Within the three sessions, the par- in R, where the speeches were first split into liamentarians held a total of 345 speeches, each chunks with regex functions. Using the R-package lasting up to five minutes. The speeches contained cldr (McCandless et al., 2013), we identified the 329’031 words and 16’630 sentences. In German, language of the text in each chunk (i.e. either 72’092 words in 6462 sentences were counted, of French or German) and split the data frame in two which 7443 words (10%) were included in the an- based on that information. We then implemented notated word list. In French, 256’939 words in spacyr (Benoit and Matsuo, 2019) separately on 10’168 sentences were counted, of which 24’535 both subsets in order to tokenise and lemmatise words (10%) were contained in the word list. On their contents, which were subsequently matched average, each speech consisted of 911 words, of with one of two data bases, again separated which an average of 89 words were represented in by language. For the German transcripts, the the annotated word lists. semantic lexical analysis was conducted with the The BAWL-R is a large German word list con- BAWL-R (Võ et al., 2009). The French transcripts taining almost 3000 words (nouns, verbs, and ad- were analysed with the SLE (Leleu, 1987). After jectives) from the CELEX database (Baayen et al., the removal of duplicate entries from the database 1996). Each word of the list was rated on va- with rules based on functions from the package lence, arousal, and imageability indicating the feel- RecordLinkage (Borg and Sariyar, 2019) and ing when reading each word. The list also includes adjusting the scales in the French database to psycholinguistic factors (e.g. number of letters, match the German ones, the subsets were reunited phonemes, word frequency, accent). It is free for and further analysed. In addition to the packages download1 . The BAWL-R enables estimations of mentioned above, we used brms (Bürkner, 2018), the emotional potential of single words but also ex- tidybayes (Kay, 2020), ggplot2 (Wickham, 2016), trapolations for sentences and whole texts. In the plotly (Sievert, 2020) and tidyverse (Wickham, BAWL-R (Võ et al., 2009), valence had been rated 2017). For each speech, we averaged the valence with the Subjective Assessment Manikin (SAM; and arousal of all the words in that speech repre- Bradley and Lang, 1994) on a 7-point scale (-3 very sented in the BAWL-R for German speeches and negative through 0 neutral to +3 very positive), the SLE for French speeches. To answer the two and arousal on a 5-point SAM-scale (1 low arousal research questions, the mean variance and mean to 5 high arousal). The split-half reliabilities of arousal of all speeches of each parliamentarian the original BAWL-R data can no longer be calcu- was calculated for each session week. Neglecting lated. According to oral communication with Jana the fact that a parliamentarian can have speeches Lüdtke (Free University of Berlin), the split-half with positive and negative emotional content, or reliability with data from a new rating of 466 words negative or positive emotional content within a resulted in a value of .97 for valence and .92 for single speech. We have not included the variation arousal. The Semantic Lexicon of Emotion (SLE; of values within the speeches of individual parlia- Leleu, 1987) is part of an unpublished master thesis mentarians in our analyses. that was integrated in the interlingual metanorm for emotional analysis of texts (EMONORM; Lev- eau et al., 2012). We used the 3000 values for va- 4 Results lence and arousal published by Leveau et al. (2012) that were transformed into the interval 1- to +1. Across all parliamentarian groups, parliamentari- ans lost 22% of the votings. We found four parlia- 3.2 Analyses mentary groups that lost about a third of the vot- ings (34%) with values from 32% to 35%. The re- After selecting the specific sessions we were maining five groups lost 14% of the votings within interested in, we downloaded the list of votings the three session weeks. Depending on the parlia- and merged them into one data frame in order mentary group, the value was between 11% and 1 18% (see table 1). The parliamentarians voted https://www.ewi-psy.fu-berlin.de/ einrichtungen/arbeitsbereiche/allgpsy/ mostly in agreement with their respective groups. Download/BAWL/index.html accessed May 2019; Only 3% of the votes were cast in disagreement To open the file a password must be requested. with the group. Figure 1: Distribution of emotional valence and emotional arousal in parliament speeches. Parl. Lost Valence Arousal types show comparable value ranges to the values group votes Mean SD Mean SD of the present study for valence (Hsu et al., 2015; Group1 11% 0.49 1.06 3.06 0.73 Jacobs et al., 2017; Jacobs and Lüdtke, 2017) and Group2 12% 0.52 1.04 3.02 0.70 arousal Jacobs and Lüdtke (2017). To be able to Group3 13% 0.62 1.11 2.71 0.60 classify this result, it is helpful to know the values Group4 18% 0.48 1.10 3.03 0.71 of emotionally neutral or non-emotional speech. Group5 18% 0.64 1.07 2.62 0.54 From a purely theoretical point of view, a neutral V W in 14% 0.55 1.08 2.88 0.68 text has a valence close to 0 and an arousal around Group6 32% 0.46 1.07 3.04 0.71 2.5. Three short stories included in the analysis of Group7 33% 0.46 1.10 2.96 0.68 Werlen et al. (2019) that were deliberately written Group8 35% 0.48 1.08 3.04 0.69 in an emotionally neutral way have valences close Group9 35% 0.67 1.09 2.64 0.50 to 0.5 and an arousal close to 2.5. In comparison, V Lose 34% 0.49 1.09 2.96 0.69 the transcribed speeches of our study have values All 22% 0.52 1.08 2.92 0.69 ranging from neutral to significantly stronger emo- Note. Parl. group=Parliamentarian group; VWin=vote tional arousal. The same is true for valence com- winners; VLose=vote losers; SD=standard deviation pared to a theoretical neutral valence. Compared to the emotionally neutral texts, the valence of the Table 1: Emotional valence and emotional arousal parliamentary speeches varies in both directions, in the speeches of the parliamentarian groups negative and positive. Figure 1 shows the distribution of emotional va- The emotional text analysis confirms the first lence (x-axis) and emotional arousal (y-axis) across research question. In the transcribed speeches, all speakers. The different colours represent the emotional states, specifically valence and arousal, nine parliamentary groups. The range of values for can be estimated with a sufficiently large variance. single words for valence is -3 to +3, for arousal 1 In the last line of table 1, the means and stan- to 5. Due to the aggregation of single words values dard deviations of emotional valence and emotional into values for each speech, the possible value span arousal of the total sample are listed. The mean got narrower. We estimate the actually possible of emotional valence is 0.52 with a standard devi- value span in the speeches for valence and arousal ation of 1.08 (absolute range: from -0.90 to 1.40). to lie within two standard deviations, i.e. between The mean of emotional arousal is 2.92 with a stan- -1.5 and 2.5 for valence, and between 2.2 and 3.6 for dard deviation of 0.69 (absolute range: from 2.25 arousal. The scaling in table 1 is adjusted accord- to 3.37). The ranges of the values for valence and ingly. Generally, the illustration shows that valence arousal are rather narrow. But they are still twice has a wider distribution than arousal. Arousal is as large as the corresponding values of an analy- divided in two sections: The section with a higher sis of 62 emotional short text with a range of 1.15 arousal contains mostly speeches in French, the points (0.02 to 1.17) for valence and the range for lower arousal section speeches in German. arousal (2.34 to 2.92; 0.58 points; Werlen et al., Table 1 shows also the percentages of lost vot- 2019). Other studies that analysed different text ings, and the means and standard deviations for emotional valence and emotional arousal of the Emodel1RE =.42,[.01,1.76]; Emodel2RE =.46,[.02,1.94], nine parliamentarian groups. The vote winners (v arousal: Emodel1RE =.62,[.03,2.09]; Emodel2RE =.67, win) have a more positive average valence, with [.02,2.69]). These results did not confirm the two a mean value of 0.55 (standard deviation: 1.08) hypotheses that emotional valence and arousal of than the vote losers (v lose) with a mean value parliamentarians’ speeches predicts the member- of 0.49 and standard deviation of 1.09. With re- ship to parliamentarian groups with different per- gard to arousal, the vote winners have a lower emo- centages of lost votings. Therefore, we have to re- tional arousal (mean value: 2.88, standard devia- ject both of them. A comparison of the models tion: 0.68) than the vote losers (mean value: 2.96, with the Bayesian ELPD LOO-criterion (theoret- standard deviation: 0.69). However, the differ- ical Expected Log Pointwise Predictive Density - ences in valence and arousal between vote winners Leave One Out) showed that model 1 with random and vote losers are very small. effects had the best fit, however the ranking is very In order to address the second research question, unreliable due to the high standard errors, which i.e. whether emotional valence and arousal are able are larger than their respective ELPD difference, to predict the membership of parliamentarians in with two exception (see table 4). one of two groups (fewer lost votings vs. more lost votings), we calculated several Bayesian regres- Predictor Estimate Est.Error l-CI u-CI sion models. Since parliamentarians spoke multi- Intercept -.48 .18 -.84 -.14 ple times across the three different sessions, result- V alence -.20 .12 -.44 .05 ing in repeated measures, we decided to calculate Arousal .22 .13 -.03 .47 multilevel models using brms (Bürkner, 2018) with N ov2019 -.10 .31 -.70 .51 session as the grouping factor. Model 0 was an Dez2019 -.07 .25 -.56 .43 intercept-only model, model 1 added the speeches’ Note. l-CI=lower lower limit credible interval; u-CI=upper valence and arousal values as predictors plus the limit credible Interval session as a categorical predictor, and model 2 added language as a fourth predictor. In order Table 2: Prediction of political affiliation (vote to reflect the nested structure of our data, mod- winners vs. vote losers); Model 1 fixed effects with els 1 and 2 were each calculated twice, once with SLE (French) and BAWL-R (German) fixed effects and once with additional random ef- fects, allowing the relation between the variables to Next, we calculated all of the models again, be moderated by the grouping factor session. In this time with the translated BAWL-R word order to inspect the role the word lists may play, we list for the French speeches. All of the mod- conducted the analysis twice, once for each of the els converged again, as indicated by the low R- two French word lists (SLE and translated BAWL- hat values. This time, a fixed effect emerged R2 ). The German word list remained constant. for valence in models 1 and 2 (Emodel1F E =- As an example of how the models were specified, .34,[-.57,-.11]; Emodel2F E =-.33,[-.63,-.03]). Arousal the design formula for model 1 is shown here: and language again showed no effects (see Ta- ble 3). As before, random effects of valence Li ∼ Binomial(1, pi ) [likelihood] and arousal emerged in both models (valence: logit(pi ) = α + βv Pi + βa Pi [linear model] Emodel1RE =.51,[.01,1.99]; Emodel2RE =.51,[.01,2.02], αi ∼ Normal(0, 10) [α prior] arousal: Emodel1RE =.69,[.02,2.41]; Emodel2RE =.77, [.03,3.08]). βv ∼ Normal(0, 10) [βv prior] βa ∼ Normal(0, 10) [βa prior] Predictor Estimate Est.Error l-CI u-CI First, we calculated the models with the SLE Intercept -.46 .18 -.81 -.12 word list for the French speeches. The R-hat di- V alence -.34 .12 -.57 -.11 agnostic with all R-hat values below 1.02 indi- Arousal .23 .12 -.01 .46 cated good convergence for all estimated param- N ov2019 -.11 .31 -.73 .50 eters in the models. However, emotional valence Dez2019 -.11 .26 -.63 .40 Note. l-CI=lower lower limit credible interval; u-CI=upper and arousal did not yield fixed effects in any of the limit credible Interval models (see Table 2), and neither did language, as the credible intervals of these predictors always in- Table 3: Prediction of political affiliation (vote cluded 0. Random effects of valence and arousal winners vs. vote losers); Model 1 fixed effects with were found in both random effects models, imply- BAWL-R for French and German ing the relationship between the predictors and the group membership depends on the session (valence: Figure 2 visualises the effects of both predictors 2 Regarding the translation of a word list, see our using the BAWL-R for German and French remarks in the discussion section speeches in model 1 (fixed effects). The figure Model Diff ELPD language speeches have a lower arousal. This ef- ELPD se LOO fect disappears when the French translation of the Leleu − BAW L BAWL-R or the NRC-VAD of Mohammad (2018) M odel1RE 0.00 0.00 -229.50 is used, which indicates a problem with the word M odel0F E -0.01 3.12 -229.50 list of Leleu (1987). M odel1F E -0.05 1.17 -229.54 Predicting the membership of speakers in par- M odel2RE -1.04 1.20 -230.53 liamentary groups with fewer or more lost vot- M odel2F E -1.16 0.26 -230.65 ings yielded ambivalent results, depending on the BAW L − BAW L French word lists. The Semantic Lexicon of Emo- M odel1RE 0.00 0.00 -227.25 tions by Leleu (1987) resulted in no effects. The M odel1F E -0.28 1.71 -227.53 alternative - a French translation of the BAWL-R M odel2F E -1.19 1.73 -228.44 - showed a weak effect for valence. Language not M odel2RE -1.19 0.23 -228.44 producing an effect was surprising, given that we M odel0F E -2.33 3.91 -229.57 found that German speeches displayed higher va- Note. FE=fixed effects; RE=random effects; ELPD=Expected lence and lower arousal compared to their French Log Pointwise Predictive Density; LOO=Leave One Out; counterparts. Despite this difference, the predictor se=standard error language was not able to predict the membership Table 4: Model fits: Model comparisons to parliamentary groups. The authors of one of the studies we intended to replicate, Hirst et al. (2014), encountered a similar issue. In compari- shows the slope (blue line) with its 95% grey- son to English transcriptions, they found a lower shaded credible interval. Arousal has a large accuracy for French transcriptions of speeches of credible interval that includes 0, indicating no the Canadian parliament. It is unclear whether effect. The effect of valence is visualised with the the discrepancies in both studies were due to the narrower credible interval. different word lists or linguistic and cultural in- fluences. We suspect that this lack of fixed ef- fects may indeed be a result of the different word 5 Discussion lists used for our analyses. The values of common The goal of this study was to find out if parlia- words in SLE and BAWL-R show correlations of mentary speeches in a Swiss canton feature emo- r=.89 for valence (457 common words) and r=.31 tional content (valence and arousal) and whether for arousal (501 common words). This suggests that content is able to predict the membership of that the SLE measures at least arousal differently the speakers in one of two groups (one with fewer than the BAWL-R does. In other studies, it was lost votings than the other, as an approximation of also found that arousal, in contrast to valence, has the more common divide between governing party a weaker correlation between different instruments and opposition). In line with our research ques- and usually has a lower inter-rater correlation (e.g. tion, we were able to estimate emotional states Kaakinen et al., prep). As mentioned in the chap- (valence, arousal) in the parliamentary speeches we ter on measuring emotions in texts, we did not em- analysed, with a rather narrow range of values for ploy the German translated NRC-VAD from (Mo- valence and even a narrower range of values for hammad, 2018) as an alternative word list due to arousal. Nonetheless, these ranges were larger as the expected loss of predictive power, as indicated the corresponding ranges in Werlen et al. (2019), by the lower correlations between human ratings where 62 emotional short stories were analysed and the valence and arousal values of the German in the same manner, or had a comparable range translation of the NRC-VAD compared to the orig- to other studies that analysed different text types inal English version. An analysis of our data with (Hsu et al., 2015; Jacobs et al., 2017; Jacobs and the German NRC-VAD confirmed this; the correla- Lüdtke, 2017). This indicates that assessing emo- tions with the values of the BAWL-R and the SLE tions by text analysis with annotated word lists were indeed very low. works well in parliamentarian speeches with a suf- The results of the random effects models indicate ficiently large variance. In Figure 1, it is noticeable that the relationships between the predictors and that the relationship between valence and arousal the outcome are influenced by the sessions them- does not have the typical u-shape often found in selves. However, we do not know why exactly ses- the literature. But as Kuppens et al. (2013) show, sions exert an influence. A plausible explanation depending on the origin of the data and the context could be the topics that were discussed within the of the study, the relationship between valence and individual sessions. Since not every topic is equally arousal may take other forms. Interestingly, in our emotional, this is likely to be reflected in the re- study, we have found two clusters that primarily spective speeches. In order to examine this, future concern the difference in arousal. The German- studies would need to quantify and categorise the Figure 2: Slopes of valence and arousal. contents of the sessions, which would also require or more votings (as an analogy to governing party more sessions and legislatures to be included in the or opposition), replicating the results of previous analysis. studies. In comparison, arousal and language were Overall, there are multiple possible reasons that far less successful. Future studies need to take could explain the weak effects we found when additional predictors into account, particularly predicting the affiliation with specific parliamen- attributes of the parliamentary sessions (e.g. the tary groups. Rheault et al. (2016) mentions that discussed topics and their affective potency) or different parliaments have their own expressions non-emotional ones. with specific meanings. Consequently, Salah et al. (2013) proposes that ”dedicated political lexicons Acknowledgments might need to be built to improve overall accu- Many thanks to the three anonymous reviewers for racy” (p. 128). Furthermore, Rheault et al. (2016) their suggestions for corrections and their valuable lists other commonly known linguistic features that comments, most of which we were able to incorpo- cannot be captured with a simple text analysis rate into the paper. based on word lists. These include sarcasm, irony, and hyperbole. In addition, there are other fac- tors besides valence and arousal that can be used to predict affiliation to parliamentary groups. For References instance, reason, logic, and culture were used in Abercrombie, G. and Batista-Navarro, R. (2020). another context, namely the effects of speeches in Sentiment and position-taking analysis of par- parliament (Freeden, 2013). liamentary debates: A systematic literature re- Finally, the strength of the prediction effects view. Journal of Computational Social Science, also was not equally consistent in the studies that pages 1–26. analysed the transcribed speeches of the Canadian parliament (Riabinin, 2009, Hirst et al., 2014). Aryani, A., Kraxenberger, M., Ullrich, S., Jacobs, Hirst et al. (2014) reported better results for the A., and Conrad, M. (2016). Measuring the ba- 36th government compared to the much less clear sic affective tone in poetry using phonological iconicity and subsyllabic salience. Psychol. Aes- results from the 39th government, indicating that thet. Creat. Arts, 10(2):191–204. inconsistent effects may be expected in this type of study. Baayen, R. H., Piepenbrock, R., and Gulikers, L. (1996). The celex lexical database (cd-rom). 6 Conclusions Barbalet, J. M. (1998). Emotion, Social theory, and social structure: A macrosociological ap- In conclusion, emotional valence and emotional proach. Cambridge University Press. arousal of parliamentary speeches can be assessed with a lexical approach of emotional text analysis. Barrett, L. F. and Russell, J. A. (1999). The struc- Depending on the word list used, the valence of ture of current affect: Controversies and emerg- parliamentary speeches is able to predict whether ing consensus. Current directions in psychologi- parliamentarians belong to groups that lost fewer cal science, 8(1):10–14. Benoit, K. and Matsuo, A. (2019). spacyr: Wrap- Holodynski, M. and Friedlmeier, W. (2012). per to the ’spaCy’ ’NLP’ Library. R package Emotionale Entwicklung: Funktion, Regulation version 1.2. und soziokultureller Kontext von Emotionen. Springer. Borg, A. and Sariyar, M. (2019). RecordLinkage: Record linkage in R. R package version 0.4-11.2. Hsu, C.-T., Jacobs, A. M., Citron, F. M., and Bradley, M. M. and Lang, P. J. (1994). Measuring Conrad, M. (2015). The emotion potential of emotion: the self-assessment manikin and the se- words and passages in reading harry potter–an mantic differential. Journal of behavior therapy fmri study. Brain and language, 142:96–114. and experimental psychiatry, 25(1):49–59. Izard, C. E. (1991). The psychology of emotions. Bradley, M. M. and Lang, P. J. (1999). Affec- Springer Science & Business Media. tive norms for english words (anew): Instruction manual and affective ratings. Technical report, Jacobs, A. M. and Lüdtke, J. (2017). Immersion The Center for Research in Psychophysiology, into narrative and poetic worlds. a neurocogni- University of Florida. tive poetics perspective. Narrative absorption, 27:69–96. Bühler, K. (1934). Sprachtheorie (language the- ory). Stuttgart: G. Fischer. Jacobs, A. M., Schuster, S., Xue, S., and Lüdtke, Bürkner, P.-C. (2018). Advanced Bayesian multi- J. (2017). What’s in the brain that ink may level modeling with the R package brms. The R character. . . .: A quantitative narrative analysis Journal, 10(1):395–411. of shakespeare’s 154 sonnets for use in (neuro-) cognitive poetics. Scientific Study of Literature, Core Team, R. (2017). R: A language and envi- 7(1):4–51. ronment for statistical computing. r foundation for statistical computing. Vienna, Austria: URL Jacobs, A. M., Võ, M. L.-H., Briesemeister, B. B., https://www. R-project. org/.[Google Scholar]. Conrad, M., Hofmann, M. J., Kuchinke, L., Lüdtke, J., and Braun, M. (2015). 10 years of Cowie, R. and Cornelius, R. R. (2003). Describ- bawling into affective and aesthetic processes in ing the emotional states that are expressed in reading: what are the echoes? Frontiers in psy- speech. Speech communication, 40(1-2):5–32. chology, 6:714. Degner, J., Doycheva, C., and Wentura, D. (2012). It matters how much you talk: On the auto- Jhean-Larose, S., Leveau, N., and Denhière, G. maticity of affective connotations of first and (2014). Influence of emotional valence and second language words. Bilingualism: Language arousal on the spread of activation in memory. and Cognition, 15(1):181–189. Cognitive processing, 15(4):515–522. Ekman, P. (1999). Basic emotions. In Dalgleish, Kaakinen, J. K., Werlen, E., Kammerer, Y., T. and Power, M. J., editors, Handbook of Cog- Acartürk, C., Aparicio, X., Baccino, T., Bal- nition and Emotion, pages 45–60. John Wiley & lenghein, U., Bergamin, P., Castells, N., Sons. Costa, A., Falé, I., Megalakaki, O., and Ruiz- Fernandez, S. (prep). Idest: International Erisen, C. and Villalobos, J. D. (2014). Explor- database of emotional short texts. Behavior Re- ing the invocation of emotion in presidential search Methods. speeches. Contemporary Politics, 20(4):469–488. Freeden, M. (2013). Editorial: Emotions, ideology Kay, M. (2020). tidybayes: Tidy data and geoms and politics. for Bayesian models. R package version 2.3.1. Freud, S. (1891). Zur Auffassung der Aphasien: Koschut, S. (2020). Emotion, discourse, and power Eine kritische Studie. F. Deuticke. in world politics. In Koschut, S., editor, Emo- tion, discourse, and power in world politics, Frijda, N. H. (1986). The emotions. Cambridge pages 3–27. Taylor & Francis. University Press. Hirst, G., Riabinin, Y., Graham, J., Boizot-Roche, Kuppens, P., Tuerlinckx, F., Russell, J. A., and M., and Morris, C. (2014). Text to ideology or Barrett, L. F. (2013). The relation between va- text to party status? In Kaal, B., Maks, I., and lence and arousal in subjective experience. Psy- van Elfrinkhof, A., editors, From text to political chological Bulletin, 139(4):917–940. positions: Text analysis across disciplines, pages 93–116. John Benjamins. Lara, E. A., Márquez, A. C., and Fuentes- Rodrı́guez, C. (2016). Emotional argumentation Hoggett, P. and Thompson, S. (2012). Introduc- in political discourse. A Gender-based Approach tion. In Hoggett, P. and Thompson, S., editors, to Parliamentary Discourse, 68:129–159. Politics and the emotions: The affective turn in contemporary political studies, pages 1–19. Con- Lazarus, R. S. (1991). Emotion and adaptation. tinuum Books. Oxford University Press. Lehne, M., Engel, P., Rohrmeier, M., Menning- Scherer, K. R. (2010). The component process haus, W., Jacobs, A. M., and Koelsch, S. (2015). model: Architecture for a comprehensive compu- Reading a suspenseful literary text activates tational model of emergent emotion. Blueprint brain areas related to social cognition and pre- for affective computing: A sourcebook, pages 47– dictive inference. PLOS One, 10(5):e0124550. 70. Leleu, S. (1987). Un atlas sémantique de concepts Sievert, C. (2020). Interactive web-based data vi- d’émotion: Normes et validation [a semantic lex- sualization with R, plotly, and shiny. Chapman icon of emotion: Norms and validation]. Unpub- and Hall/CRC. lished MA Thesis, Psychology, Catholic Univer- sity of Louvain. Ullrich, S., Aryani, A., Kraxenberger, M., Jacobs, A. M., and Conrad, M. (2017). On the rela- Leveau, N., Jhean-Larose, S., Denhière, G., and tion between the general affective meaning and Nguyen, B.-L. (2012). Validating an interlin- the basic sublexical, lexical, and inter-lexical fea- gual metanorm for emotional analysis of texts. tures of poetic texts—a case study using 57 po- Behavior Research Methods, 44(4):1007–1014. ems of hm enzensberger. Frontiers in Psychol- ogy, 7:2073. Lutz, C. and White, G. M. (1986). The anthropol- ogy of emotions. Annual review of anthropology, Võ, M. L., Conrad, M., Kuchinke, L., Urton, K., 15(1):405–436. Hofmann, M. J., and Jacobs, A. M. (2009). The berlin affective word list reloaded (bawl-r). Be- Matsumoto, D., Yoo, S. H., and Nakagawa, S. havior Research Methods, 41(2):534–538. (2008). Culture, emotion regulation, and adjust- ment. Journal of Personality and Social Psy- Warriner, A. B., Kuperman, V., and Brysbaert, chology, 94(6):925. M. (2013). Norms of valence, arousal, and dom- inance for 13,915 english lemmas. Behavior Re- McCandless, M., Sanford, M., and Firat, A. search Methods, 45(4):1191–1207. (2013). cldr: Language identifier based on CLD library. R package version 1.1.0. Werlen, E., Imhof, C., Benites, F., and Bergamin, P. B. (2019). The reader’s feeling and text-based Mohammad, S. (2018). Obtaining reliable human emotions: the relationship between subjective ratings of valence, arousal, and dominance for self-reports, lexical ratings, and sentiment anal- 20,000 english words. In Proceedings of the 56th ysis. In SwissText 2019, Winterthur, 18-19 June Annual Meeting of the Association for Compu- 2019. CEUR-WS.org. tational Linguistics (Volume 1: Long Papers), pages 174–184. Werlen, M., Moser, I., Imhof, C., and Bergamin, P. (2018). Is reading mirrored in the face? a Panksepp, J. (1998). The foundations of human comparison of linguistic parameters and emo- and animal emotions. Oxford University Press. tional facial expressions. In SwissText 2018, Winterthur, 12-13 June 2018. CEUR-WS.org. Pennebaker, J. W., Boyd, R. L., Jordan, K., and Blackburn, K. (2015). The development and Wickham, H. (2016). ggplot2: Elegant graphics for psychometric properties of liwc2015. Technical data analysis. Springer-Verlag. report, University of Texas at Austin, Austin, Texas. Wickham, H. (2017). tidyverse: Easily install and load the ’Tidyverse’. R package version 1.2.1. Rheault, L., Beelen, K., Cochrane, C., and Hirst, G. (2016). Measuring emotion in parliamentary Widmann, T. (2021). How emotional are populists debates with automated textual analysis. PlOS really? factors explaining emotional appeals in one, 11(12). the communication of political parties. Political Psychology, 42(1):163–181. Riabinin, Y. (2009). Computational identification of ideology in text: A study of canadian par- liamentary debates. MSc paper, Department of Computer Science, University of Toronto. Salah, Z., Coenen, F., and Grossi, D. (2013). Ex- tracting debate graphs from parliamentary tran- scripts: A study directed at uk house of com- mons debates. In Proceedings of the Fourteenth International Conference on Artificial Intelli- gence and Law, pages 121–130. Scherer, K. R. (1999). Appraisal theory. In Dal- gleish, T. and Power, M. J., editors, Handbook of cognition and emotion, pages 637–663. John Wiley & Sons.