Feeling like a State. The Sentiments Tide of Swiss Diplomacy through the Eye of the Algorithm

Since the second half of the 19^th century, Swiss embassies in an increasing number of world cities have to send regular reports (usually on a monthly basis) on the political and military situation in their host countries to the Federal Department of Foreign Affairs (FDFA) in Bern. These reports served to inform the government about the international situation and thus were the basis for making foreign and security policy decisions. These so-called ›political and military reports‹

SFA, fond E2300* Eidgenössisches politisches Departement: Politische und militärische Berichte der Auslandvertretungen (1848–1965) 1848–1965. For later periods see the fonds E2300-01* and E2010-02A*: Eidgenössisches politisches Departement: Politische und militärische Berichte der Auslandvertretungen (1 966–1978) 1966–1978 and (1 979–) 1979–1990.

The authors of our corpus, the diplomats, are experts.

Herren and Zala wrote about the figure of the expert in Swiss foreign policy. They concentrated on the international conferences: Madleine Herren / Sacha Zala: »›Die Experten verpflichten ihre Regierungen in keiner Weise‹. Experten im Milizsystem der schweizerischen Aussenpolitik der Zwischenkriegszeit«, in: Traverse 2001/1, pp. 96–109.

In the second half of the 20^th century, the design of diplomacy was less solemn. Negotiations, correspondence (notes, reporting), protection of interests, press and propaganda were typical tasks of an embassy.

Gerbore: Formen und Stile der Diplomatie, pp. 134–200.

Switzerland initially had more consulates than diplomatic missions. On the eve of World War II, the confederation maintained a network of 121 consular offices, some of them professional and some honorary.

Claude Altermatt: »Konsularwesen«, in: Historisches Lexikon der Schweiz, online at: http://www.hls-dhs-dss.ch/textes/d/D13809.php(1.8.2018)

Today, the FDFA recruits diplomats in a multistage admission competition, the so-called Concours, and trains them for around two years at home and abroad. For its Concours diplomatique 2018, the FDFA was seeking »young and innovative talents who use their potential of diplomacy to actively shape Swiss foreign policy«.

Diplomatischer Concours und Ausbildung, online at: https://www.eda.admin.ch/eda/de/home/das-eda/arbeiten-eda/berufserfahrene/diplomatische_laufbahn/der-diplomatischeberuf.html (1. 8. 2018).

Keller discussed the »democratization of the diplomatic corps« in the context of the Concours.

Florian Keller: Botschafterporträts – Schweizer Botschafter in den Zentren der »Macht« zwischen 1945 und 1975, Zürich 2017, p. 50.

Our algorithmic analysis of the Swiss political reports is of course subject to a series of material contingencies. The original documents of the E2300* fond are printed on paper of diverse qualities and formats, which required scanning and optical character recognition. Furthermore, the political and diplomatic reports are written in many languages, among which 46% are written in French,

According to a rough automatic detection during the scanning part of the process. Some texts detected as French at this point turned out to consist of less than 70% of French words at a later stage.

The corpus totals 15,584,725 words, of which 35% are associable with an emotion by the FEEL dictionary (cf. infra). Individual reports contain between 24 and 102,000 words. Only 130 documents exceed 10 pages length (Figure 1).

Figure 1

Figures 2 and 3 show that these reports come from all of world's and Europe's regions, respectively. This is an important precondition for testing our hypothesis according to which differences in sentiments across nations and continents can be identified. Reports from important cities, such as Moscow

There was no Swiss envoy in the USSR between 1923 and 1956, see Peter Collmer: »Russie« (27.1. 2016), online at: http://www.hls-dhs-dss.ch/textes/f/F3376.php (1.9. 2018).

Figure 2

Figure 3

We further hypothesise that sentiments show a chronological evolution, which can be correlated with general or specific political, economic or cultural trends. In such a way, TSA can contribute to a history of the administration, to a description and analysis of diplomatic and administrative practices. For this, we need our data to be sufficiently fine-grained on the historical time scale.

The date of most political reports in our possession is expressed in time spans covering several years. In effect, the dating is not based on the documents themselves but on the associated metadata in the archival database of the Swiss Federal Archives (SFA). Figure 4 illustrates the historical distribution of these dates (cf. Figure 2).

Figure 4

As Figure 4 shows, the most densely covered period across all cities, on which we therefore focus, lies between 1920 and 1960. Reports from cities such as Antwerp are few and dated to a single, wide time span (1857–1936) excluding analysis of a historical evolution. Documents from Mexico are more numerous and stem from five different time spans, but the dating precision is still insufficient. On the contrary, cities such as Paris, Washington, Tokyo, London and Stockholm provide a continuous sequence of precisely dated texts. We have selected such cities due to this availability of fine-grained chronological data. We have set the precision threshold to 3 years, meaning that we excluded reports dated from a larger time span from further analysis. This brings us down to 12,364 analysed reports. The next question is whether the language of these texts allows detection of statistically meaningful sentiments.

The European Parliament recently tested TSA of foreign documents to detect translation bias. In case of quantifiable change in mood and sentiment across translations in documents, Whelchel and Fellbaum speak of »misconstrued communications«, which the administration seeks to overcome.

John Whelchel / Christiane Fellbaum: »Sentiment Analysis of Foreign Documents to Detect Translation Bias. Misconstrued Communications in the European Parliament«, Independent Work Report Spring 2014, online at: http://www.johnwhelchel.com/static/mysite/index/IWBiases.pdf (1. 8. 2018).

If an algorithm is to attribute sentiments to data of any type, it needs a predefined discrete and finite set of sentiment categories. Whether such a set can be constructed at all is of course in dispute,

The complexity of this psychological disputatio ramifies in neuroscience, biology, linguistics, anthropology and philosophy. Darwin proposed the existence of transcultural basic emotions in an ethically motivated publication, see Charles Darwin: The expression of the emotions in man and animals, London 1872, online at: http://archive.org/details/expressionofemot1872darw (1. 8. 2018). For further discussion, see Keith Oatley / Philip N. Johnson-Laird: »Towards a cognitive theory of emotions«, in: Cognition & Emotion 1 (1987), pp. 29–50; Andrew Ortony / Terence J. Turner: »What's basic about basic emotions?«, in: Psychological Review 97/3 (1990), pp. 315–331; Paul Ekman, »An Argument for Basic Emotions«, in: Cognition and Emotion 6 (1992), pp. 169–200; Kristen A. Lindquist et al.: »Language and the perception of emotion«, in: Emotion 6/1 (2006), pp. 125–138, online at: https://doi.org/10.1037/1528-3542.6.1.125(1 8. 2018); Maria Gendron et al.: »Emotion words shape emotion percepts«, in: Emotion, 12/2 (2012), pp. 314–325, online at: https://doi.org/10.1037/a0026007 (1. 8. 2018); Paul Ekman / Dacher Keltner: »Darwin's Claim of Universals in Facial Expression Not Challenged«, 27. 3. 2014, online at: https://www.paulekman.com/uncategorized/darwins-claim-universals-facial-expression-challenged (1. 8. 2018).

We retain that natural language provides names for sentiments, which can serve as a classification basis. We also pose that language participates in the mental state of its speaker, by encouraging her to conceive her emotions in a specific way. In turn, her use of a given vocabulary to express her emotions contributes to the reification of this vocabulary and of the associated mental states of other individuals, and thus affects her social environment. Emotions »arise at junctures of social plans«.

Oatley /Johnson-Laird: »Towards a cognitive theory of emotions«.

TSA inherits its basic emotions from research in psychology and neuroscience, for example:

Further examples: [rage/terror, anxiety, joy] Jeffrey A. Gray / Neil McNaughton: The neuropsychology of anxiety. An enquiry into the function of the septo-hippocampal system, Oxford 1982; ders.: »The whole and its parts. Behaviour, the brain, cognition and emotion«, in: Bulletin of the British Psychological Society 38 (1985), pp. 99–112; [anger, interest, contempt, disgust, distress, fear, joy, shame, surprise] Silvan S. Tomkins: »Affect theory«, in: Klaus R. Scherer / Paul Ekman (eds.): Approaches to emotion, Hillsdale 1984, pp. 163–195; [expectancy, fear, rage, panic] Jaak Panksepp: Affective neuroscience. The foundations of human and animal emotions, Oxford 1998 (2004); [anger, disgust, anxiety, happiness, sadness] Oatley / Johnson-Laird: »Towards a cognitive theory of emotions«; [anger, contempt, disgust, distress, fear, guilt, interest, joy, shame, surprise] Caroll E. Izard: Human emotions, New York 1977. Note that sentiment analysis is not necessarily based on text. Its development is rather tributary to studies of facial expressions, see Paul Ekman / Wallace V. Friesen: »The repertoire of nonverbal behaviour. Categories, origins, usage, and coding«, in: Semiotica 1 (1969), pp. 49–98; idem: »Universals and Cultural Differences in Facial Expression of Emotion«, in: James K. Cole (ed.): Nebraska Symposium on Motivation, 1971, Lincoln 1972, pp. 207–283; Izard: Human emotions; Paul Ekman / Wallace V. Friesen / Phoebe Ellsworth: »What emotion categories or dimensions can observers judge from facial behavior?«, in: Paul Ekman (ed.): Emotion in the human face, New York 1982, pp. 39–55; Rachel E. Jack / Oliver G. B. Garrod / Philippe G. Schyns: »Dynamic facial expressions of emotion transmit an evolving hierarchy of signals over time«, in: Current Biology 24/2 (2014), pp. 187–192, online at: https://doi.org/10.1016/j.cub.2013.11.064 (1.8. 2018), brain activity, see Gray / McNaughton: The neuropsychology of anxiety; idem: »The whole and its parts; Tomkins: »Affect theory«; Jaak Panksepp: »Toward a general psychobiological theory of emotions«, in: The Behavioral and Brain Sciences 5 (1982), pp. 407–467; Panksepp: Affective neuroscience; or neuroendocrinology, see Panksepp: »Toward a general psychobiological theory of emotions«; Panksepp: Affective neuroscience; Judith A Toronchuk / George F.R. Ellis: »Affective Neuronal Selection. The Nature of the Primordial Emotion Systems«, in: Frontiers in Psychology 3 (2012): p. 589, online at: https://www.frontiersin.org/articles/10.3389/fpsyg.2012.00589/full (1. 8. 2018).

Most authors working with basic emotions consider these as categories subsuming a wider range. Cheerfulness, zest, contentment, pride, optimism, enthrallment and relief, for example, are all considered as variants of joy by Shaver et al.

Phillip Shaver et al: »Emotion knowledge. Further exploration of a prototype approach«, in: Journal of Personality and Social Psychology 52/6 (1987), pp. 1061 – 1086.

Figure 5

Some authors treat basic emotions as distinct from sentiment polarities: positive vs. negative,

Bo Pang / Lillian Lee: »Opinion Mining and Sentiment Analysis«, in: Information Retrieval 2/1–2 (2008), pp. 1–135.

TSA uses mainly two classes of methods to detect sentiments in texts, based on either a lexicon or machine learning.

A lexicon-based approach considers a text as a list of words. It does not analyse its sequential linguistic structure but only frequencies of isolated words (unigrams), predefined phrases or word sequences of uniform length n (n-grams). By constructing the so-called document-term matrix, one seeks to identify individual texts in which some words occur more than in others in the same corpus. Words are lemmatised beforehand to avoid redundant counts of flections (plurals, verb tenses, cases). TSA then treats lemmas as markers of specific sentiments, according to a predefined lexicon.

The most cited lexicon of this kind is the NRC-EmoLex,

Saif M. Mohammad / Peter D Turney: »Crowdsourcing a Word-Emotion Association Lexicon«, in: Computational Intelligence 29/3 (2013), pp. 436–465, online at: http://arxiv.org/abs/1308.6297 (1. 8. 2018).

The use of a lexicon consisting mainly of unigrams to detect sentiments might be surprising. Among its numerous limitations is the semantic shift

Andreas Blank/Peter Koch: Historical Semantics and Cognition, Berlin, Boston ²2013, online at: https://doi.org/10.1515/9783110804195 (1.8.2018)

Diplomatic language, rich in such wordings, poses a challenge for TSA. Proceeding by word collocation sampling, we found our most emotional lemmas, like ›heureux‹, to be mostly used in an unmodified way in our over 5 million emotionally marked lemmas of our corpus. But we did not find, yet, a systematic method for identifying the proportion of such problematic cases such as negation and irrealitis. Word sense disambiguation techniques by machine learning should offer a solution.

Machine learning approaches consist of ›training‹ a sentiment-detection algorithm. The method's advantage lies in its independence from a general-use, context-unspecific lexicon. It relies instead on the dataset itself, from which a ›training set‹ of texts is extracted and annotated by human analysts. The construction of a training set is accordingly resource demanding. Authors often reach out to pre-tagged databases of texts, such as user reviews that are both textual and ordinal. The Internet Movie Database, for instance, provides user comments associated with an ordinal rating, ranging from one to ten.

Andrew L. Maas et al.: »Learning Word Vectors for Sentiment Analysis«, in: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics. Human Language Technologies, Portland, Oregon 2011, pp. 142–150, online at: http://www.aclweb.org/anthology/P11-1015 (1. 8. 2018).

The most used supervised machine learning methods

For a recent review, see Jaspreet Singh / Gurvinder Singh / Rajinder Singh: »Optimization of sentiment analysis using machine learning classifiers«, in: Human-Centric Computing and Information Sciences 7/1 (2017): p. 32, online at: https://doi.org/10.1186/s13673-017-0116-3 (1. 8. 2018).

In this study, we focus on a lexicon-based approach with a predefined lexicon. However, we have felt a need to adjust it.

We developed our analysis algorithms using the R programming language. With regard to algorithmic efficiency, we have used the Quanteda package for R,

Kenneth Benoit et al. contributors:: »quanteda: An R package for the Quantitative Analysis of Textual Data«. GitHub. Retrieved from https://github.com/quanteda/quanteda (Original work published 2012). This citation pertains to a software: a collective open source development initiated by first author; also the one of Tyler Rinker below.

In order to use it, we have lemmatised our original corpus at the tokens' level using the LEFFF lexicon

Sagot: »The Lefff«.

Authors, political entities, and sentiment dictionary compilers: the three subjects of sentiment expression

Context adjustment is a difficult task of TSA, especially in lexicon-based approaches. Some authors propose specific lexica for domains like finance.

Loughran / McDonald: When is a Liability.

Figure 6

Unsurprisingly, words associated with the exercise of political power occur most frequently in our corpus. Surprisingly, FEEL associates such words with sentiment polarity (Figure 6). To make sense of the lexicon and to define a more solid base for its adjustment, we have decided to distinguish between three types of emotions or, more precisely, between what we shall name the three subject levels of emotional expression. The first two levels reflect what sentiment analysis also defines as subjectivity classification.

Janyce Wiebe / Rebecca Bruce / Thomas O'Hara: »Development and use of a gold standard data set for subjectivity classifications«, in: Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics (ACL-99), 1999, pp. 246–253; Stefano Baccianella / Andrea Esuli / Fabrizio Sebastiani: »SentiWordNet 3.0. An Enhanced Lexical Resource for Sentiment Analysis and Opinion Mining«, in: Proceedings of LREC, 2010.

The value judgements of FEEL's authors raise questions and concerns. Some appear as a mere result of an automatic translation. The English term ›date‹, for instance, is correctly translated to ›date‹ in French, but the many dates mentioned in our corpus hardly carry any romantic connation; its association with joy is a mere misunderstanding. Similarly, an English ›person‹ is correctly translated to ›personne‹, but the lemma then also means ›nobody‹. ›Occupation‹ has twofold meaning both in English and French and is distinguished in German by the terms Beschäftigung and Okupation. Polysemy is a great challenge to sentiments analysis. So is context. The most ›disgusted‹ report

CH-BAR#E2300#1000-716#314#5#5

Our aim being to analyse the evolution of sentiments of levels 1 and 2, we have individually examined the 100 most frequent lemmas for each polarity and for each one of the six Ekman emotions and removed all terms of level 3 that reflect judgement values of the lexicon's authors (see Annex for a complete list).

Finally, as expected, we have observed that 20^th century ambassadors make extensive use of elaborate greetings and expressions such as ›j'ai l'honneur de‹, ›veuillez agréer l'hommage de mon respects‹ ›votre excellence‹ and, most frequently, ›assurance de ma considération‹ (18,488 occurrences).

Such turns contain words charged with sentiments of level 1 (respect, honneur, considération) but actually reflect mere adherence to social conventions. It can be argued that all reflect a particular disposition of level 2: diplomatic politeness, whose degrees of emphasis may vary. Interestingly, the choice of terms evolves over time (Figure 7). A full analysis could be dedicated to this topic, since even the slightest variations in polite wordings may carry otherwise unexpressed sentiments of joy or anger. At the present stage, however, we have combined such expressions into compound n-grams without emotional value.

Figure 7

In textual statistics, simply measuring the proportion of lemmas associated with a specific sentiment is not enough. The FEEL lexicon associates more of its terms with positive feelings than, for instance, with the negative ones. This unequal distribution would induce a strong statistical bias, since counting occurrence of lemmas associated with each sentiment would reflect the relative frequencies of sentiments in the lexicon itself, not their relative frequencies in a given text. The sentiment scores (e. g. S_p = Score_positive or S_n = Score_negative) are therefore adjusted as follows:

Sp=TpT∀<!-- ∀ -->LpSn=TnT∀<!-- ∀ -->Ln \begin{align} & {{S}_{p}}=\frac{{{T}_{p}}}{{{T}_{\forall }}{{L}_{p}}} \\ & {{S}_{n}}=\frac{{{T}_{n}}}{{{T}_{\forall }}{{L}_{n}}} \\ \end{align}

where:

We have applied the same adjustment to all six Ekman emotions. A compound indicator taking into account both positive and negative polarities is defined as

compound = S_p –S_n

In the upcoming graphics, S_p shall be referred as positive and S_n as negative.

For the sentiment scores thus defined, we observe a normal distribution across texts, with varying mean values and standard deviations (Figure 8). This is encouraging for a statistical interpretation of our results. The expression of positive sentiments dominates, followed by joy and surprise. The least proportion of texts expresses negative sentiments and sentiments of disgust.

Figure 8

We present our detailed results in the section ›individual cases‹. Globally, we observe that sentiments can be organised into three groups. The first two, negative sentiments (negative, anger, fear and sadness) (Figure 12) and positive sentiments (positive, compound and joy) (Figure 20), display strong intragroup correlation. This means that an analysis in terms of positive and negative polarities is adapted for our purpose, and we shall concentrate on it in further work. In other words, in our context, the Ekman emotions mainly explicate the meaning of these polarities, without yielding more differentiated insight. Joy should be considered a heightened and probably more subjective positive emotion. Anger, fear and sadness combine into a general negative feeling. The third group, encompassing sentiments of disgust and surprise (Figure 19), shall be treated very briefly.

Before all interpretation, we need to set our results against a general historic background. The major known events between 1920 and 1960, most of which verify by text mining of our corpus, include the world economic crisis in the late 1920s (Figure 9), early Fascism, later Nazism in the 1930s (Figure 10), the Second World War and post-war economic reconstruction (Figure 11) and the beginning of the Cold War in the 1950s. For the US, the evolution of the mood was different: Fascism and Nazism did not directly affect Washington until the 1940s when the US supported the Allies and entered the war, and post-war economic development took a different course than that of war-torn Europe, whose reconstruction was funded significantly by the US under the Marshall Plan. On the other hand, the highly industrialised Japan, politically part of the Axis powers, was presumably subject to similar basic moods as Europe.

Christian W. Spang / Rolf-Harald Wippich. (ed.): Japanese-German Relations, 1895–1945. War, Diplomacy and Public Opinion, London 2006.

Figure 9

Figure 10

Figure 11

Figure 12

In psychology, the usual subjects of emotions are human individuals, observed in a precise moment. In our case, we are dealing with aggregate emotions of entire sets of political reports, written by multiple authors at different dates. The ontological gap between parts and the whole that we cross in this manner requires caution. While we do provide excerpts of individual political reports to support our detailed analysis, the reader must keep in mind that no single text can fully illustrate the aggregate phenomenon of a statistical regularity. Expecting the contrary would resemble predicting a human's behaviour from the observation of his or her individual cells.

We argue that aggregate subjects of emotions do exist. In fact, as we have discussed in earlier work, the acting subject of any action, including sentiment expression, is always a synthetic individual.

See Ourednik: L'habitant, pp. 156–195. This notion of syntheticity can be traced back, for instance, to Spinoza's conatus (Ethica 3, 6), in particular in its interpretation in terms of ›will‹ by Nietzsche, and further translation into the Freudian Ich at the interface of the Es and the exterior world. It also relates to Leibniz's notion of monad and its sociological interpretation by Gabriel Tarde, first indirectly, see Bruno Latour: Reassembling the Social, Oxford 2005, and then directly adopted by Bruno Latour in his actor-network theory, see Bruno Latour et al.: »›The whole is always smaller than its parts‹, a digital test of Gabriel Tardes' monads«, in: The British Journal of Sociology 63/4 (2012), pp. 590–615, online at: https://doi.org/10.1111/j.1468-4446.2012.01428.x (1. 8. 2018). The synthetic individual can also be conceived of as the self of an autopoietic system constructed by the system's observer, see Ourednik: L'habitant, p. 163; cf. Humberto Maturana / Francisco Varela: De Maquinas y Seres Vivos, Editorial Universitaria 1972. The synthetic subject of emotion should be conceived of as a logical place from which a sentiment about the world can be assessed and interpreted.

The advantage of aggregating sentiments of multiple documents is that we progressively leave the level of subject matters and authors to find the aggregate sentiments of institutions, epochs and places, which are only revealed in the aggregate. This is all the more significant if we recall that the reports were sent regularly to the FDFA in Berne serve to inform the government about the international situation and thus are the basis for making foreign and security policy decisions, acting not as an individual but as a nation-state. The reports permit feeling like a state.

In allusion to James C. Scott's famous Book (James C. Scott: Seeing Like a State. How certain Schemes to Improve the Human Condition Have Failed, New Haven 1998).

Throughout the history of the administration, synthetic sentiments translate into synthetic actions. This process gives a nice illustration of the old notion of Wirkungszusammenhang,

Wilhelm Dilthey: Der Aufbau der geschichtlichen Welt in den Geisteswissenschaften, Frankfurt am Main 1910 [1970], at p. 85– 196.

Thus, the earlier mentioned stance, according to which »the person shapes the activity«

Keller: Botschafterporträts, p. 348.

In our case, a fourth, easily forgotten actant intervenes:

With regard this fourth actant, we must stress to which point word–sentiment associations strongly vary with the selection of a lexicon's contributors. Our manual adjustment of the FEEL lexicon seeks to limit observation to sentiments proper to subjects of levels 1 (author and subject matter) and 2 (context and institution).

The relation between the three subject levels of sentiment expression and the synthetic subject

As suggested, there is a direct correspondence between our fourth actant and sentiment expressions of level 3 (judgement arbitral of the lexicon's authors). What we have defined as level 1 (›joie‹, ›tristesse‹, etc.) and 2 (›guerre‹, ›crise‹, etc.) of sentiment expression could allow us to distinguish the part played by single human authors in the synthesis of sentiment scores of documents or document aggregates. Methodologically, however, making this latter distinction would imply a total revision of the more than 10,000 terms of the FEEL dictionary, which we could not accomplish at this stage; thus, our sentiment scores do not account for it. We shall keep track of these levels, however, in our interpretation of individual cases.

For an interpretation of the quantitative analysis of the political reports, we have selected six countries, respectively capitals: Great Britain (London), France (Paris), Italy (Rome), Sweden (Stockholm), Japan (Tokyo) and the US (Washington). This allows us to analyse cities on three continents, while concentrating on locations from where records are available with sufficient dating precision for the whole time of observation (cf. Figure 4). We start with negative sentiments.

A first general observation shows that the range of negative feelings before and after World War II is significantly smaller than that during the war. Their frequency reaches its maxima between 1940 and 1944. Mathematically, this can be due to either a frequency decline in positively connoted lemmas or a frequency rise in the negatively connoted ones. Historically, we assume that it relates to anxiety and the negative horizons of expectation during wartimes.

Figure 13 gives better insight into rise of negative sentiments in the 1940–1944 period, by showing the obvious contribution of the negative lemma ›guerre‹ and associated terms ›troupe‹, ›armer‹ and ›militaire‹. It is striking that, according to our three subject levels of emotional expression defined earlier, the most influent sentiments cannot be attributed to the level of authors' reports (1) but to the level of political entities (2). This does not mean that the ambassadors never express personal-level sentiments, as illustrate documents found for Tokyo. TSA allowed us to detect the most negatively marked text from that Embassy (or Legation, as an Embassy was called then). It was headed by Camille Gorgé, a lawyer and the Envoyé extraordinaire et ministre plénipotentiaire of Switzerland in Japan.

Patrick Ziltener (ed.): Handbuch Schweiz – Japan. Diplomatie und Politik, Wirtschaft und Geschichte, Wissenschaft und Kultur. Texte, Dokumente und Bilder aus 400 Jahren gegenseitiger Beobachtung, Austausch und Kooperation, Zürich 2010.

Figure 13

causées par bombardements aériens et récent tremblement de terre«). But not only the American bombardments from the air test the population (»raids aériens éprouvent la population«) but also the hope for a victory of the Axis powers dwindles: The defeat seems inevitable (»défaite semblant inevitable«).

E2300#1000/716#1085*, 175 / [12], Tokio, Politische Berichte und Briefe, Militärberichte, Band 12, 1943–1945.

Nine months later, on September 1945, after the defeat of Japan, Gorgé sent a telegram to Berne, which was also very emotional; it reflects Japan's sadness. The Swiss ambassador reports on his meeting with the new foreign minister: He has not much illusion about the difficulties for Japan to win back international esteem (»II ne se fait pas beaucoup d'illusions sur les difficultés pour le Japon de reconquérir l'estime internationale«). The position of the government is extremely precarious because it depends entirely on Americans (»La position du gouvernement est d'ailleurs extrêmement précaire car elle dépend entièrement des américains«). Gorgé is amazed at the ease with which the Japanese suffer their extremely humiliating fate (»On est toujours plus étonné de la facilité avec laquelle les japonais subissent leur sort extrêmement humiliant«).

E2300#1000/716#1085*, 175 / [12], Tokio, Politische Berichte und Briefe, Militärberichte, Band 12, 1943–1945.

Relating back to the question of the subject of emotion (author or political entity),

Chapter »Global results and interpretation«.

Figure 14

These observations raise many questions: To what degree do an ambassador's personal feelings influence relations between political entities in general and between Japan and Switzerland specifically? Gorgé probably identified too strongly with his host country, which might explain why he had to leave Japan and was sent to Turkey in 1946.

Eidgenössischer Staatskalender, 1946, Politisches Departement, Türkei (Ankara), 39, online at: https://www.amtsdruckschriften.bar.admin.ch/viewOrigDoc/40001842.pdf?id=40001842 (1.8. 2018).

But let us proceed to a second example, in which the aggregate level of the political entity becomes more apparent. Surprisingly, the document with the most violent rash in terms of negativity, fear, grief and anger is a report on Sweden from the 1928 to 1932 period. Actually, it reveals the strength of the socialist movements in Sweden in the late 1920s and 1930s in comparison to other Western countries,

E2300#1000/716#1015*, 166 / [6], Stockholm, Politische Berichte und Briefe, Militärberichte, Band 6, 1928–1931.

Switzerland and Sweden had a long tradition of good relations. Since 1847, the Swiss consulate in Christiania (Oslo) was also responsible for Sweden. The first Swiss consulate in Sweden was established at Stockholm in 1887. For a longer time, the head of the Swiss diplomatic mission in Berlin was also responsible for Sweden. The diplomatic relations were upgraded in 1919 with a Swiss envoy to Stockholm. In 1957, the legation was transformed into an embassy. The two countries made similar experiences during the World War II: Both were neutral, both economically and financially dependent on relations with both the Allies and Nazi Germany. The Allies even addressed common questions in a Sweden– Swiss Committee. As Sweden and Switzerland faced new challenges during the Cold War, the two countries moved even closer. Foreign Ministries agreed on a regular mutual exchange of views in the 1950s.

Irène Lindgren / Renate Walder (eds.): Schweden, die Schweiz und der Zweite Weltkrieg, Bern 2001.

The cultural relations between Switzerland and Sweden were close as well. Schiller's William Tell was translated into Swedish in 1823. Both countries exchanged films during the war. In Switzerland, Swedish lifestyle (home furnishing, books, travel) became very popular after World War II. In Sweden, Switzerland was sometimes seen as a model for the future constitution of Europe,

Karin Naumann: Utopien von Freiheit. die Schweiz im Spiegel schwedischer Literatur, Berlin 1994.

Given this narrative of the tight and good relations between both countries, it astonishes that the document from our corpus with the second-most violent rash in terms of negativity, fear, grief and anger, after the Tokyo reports, is the aforementioned report

CH-BAR#E2300#1000-716#1015#2#9, E2300#1000/71 6#1015*, 166 / [6], Stockholm, Politische Berichte und Briefe, Militärberichte, Bd. 6, 1928–1931.

Figure 15

Figure 16

Obviously, the negativity of the reports relates to an economic crisis. But what does the Swiss embassy in Stockholm deem as radical and extreme in this context? This question can be answered by representing the semantic network of word co-occurrence in the documents from that place and period (Figures 17 and 18). The figures show not only negative and positive expressions but also their neighbouring lemmas, notably ›socialiste‹, ›ouvrier‹, ›communiste‹ and ›gauche‹. Co-occurrent terms are connected by network edges and closer to each other. Based on them, we can hypothesise that the strengthening of the labour movement in Sweden worried Swiss diplomacy. The Socialist Workers' Party (›parti‹, ›socialiste‹, ›ouvrier‹) taking over government power in Sweden (›pouvoir‹, ›gouvernement‹) seems a development that Switzerland did not follow politically and socially. This would also explain why, after World War II, Switzerland's idealised image faded and criticism became loud. Swedish authors like Arvid Brenner scourged in particular a reckless capitalism in Switzerland.

Eric Flury-Dasen: »Schweden«, in: Historisches Lexikon der Schweiz, online at: http://www.hls-dhs-dss.ch/textes/d/D3371.php (1. 8. 2018).

Figure 17

Figure 18

Figure 19

Figure 20

Disgust and surprise (Figure 19) mostly correlate with negative sentiments and provide no extra information. We see no specific added value of these sentiment categories – as implicitly defined by their associated word set in the FEEL lexicon – for the analysis of our dataset. Based on our observation, we do not recommend their use for analysis similar to ours.

One curve of evolution of the positive sentiments (Figure 20) immediately stands out: Values for Washington are falling sharply during World War II, remain at a very low level for a while and then rise again after the war (A). Lemmas such as ›croire‹ and ›foi‹ occur less often then (Figure 21), suggesting a loss of mutual confidence. Upon the background of historical knowledge,

Linus von Castelmur: Schweizerisch-Alliierte Finanzbeziehungen im Übergang vom Zweiten Weltkrieg zum kalten Krieg. Die deutschen Guthaben in der Schweiz zwischen Zwangsliquidierung und Freigabe (1945–1952), Zürich 1992; Marco Durrer: Die schweizerisch-amerikanischen Finanzbeziehungen im Zweiten Weltkrieg. Von der Blockierung der schweizerischen Guthaben in den USA über die »Safehaven«-Politik zum Washingtoner Abkommen (1941–1946), Bern 1984.

Figure 21

As much interesting are the joy scores for Stockholm that decline rapidly during the 1930s (Figure 20B), stay at a low level during the forties and decline again after the war (Figure 20C), almost in a mirrored image of the assessment for Figure 12. This points historians interested in relations between Switzerland and Sweden to possibly relevant sources for describing these relationships while socialist forces increasingly determined politics in Sweden. In addition, it raises a question similar to the case of Tokyo: To what degree do the Swiss ambassador's feelings influence relations

between Sweden and Switzerland? What role did the socialist turn play in the reports for a country with which Switzerland felt strongly connected? Did the professionalisation of diplomacy improve these relations?

Positive sentiment scores for London are consistently elevated. The same applies to Paris and Rome. In effect, literature also attests for good relations between Switzerland and its direct geographical neighbours, France and Italy.

See, for instance, the information of the FDFA to the bilateral relations to these countries. For France in particular, online at: https://www.eda.admin.ch/eda/en/home/representations-and-travel-advice/france/switzerland-france.html (1. 8. 2018).

In all cases, caution in such interpretations is important. It is, for instance, worth taking a closer look at the big joy in London (Figure 22). The terms that contribute to this result are notably good (›bien‹), create (›créer‹), success, interest and beneficial (›favorable‹). They could signal both good bilateral relations and a potentially favourable economic situation. To determine what was really meant, the individual reports should be consulted to answer more specific questions: When did the roaring 20s come to an end in the UK? How did this affect the bilateral relations with Switzerland, which in general remained rather consistently good?

Figure 22