Our first approach of categorisation can be classified as deductive. There is a certain degree of legal-philosophical background behind the historical phenomenon investigated, but more prominently we are offered references to specific legal sources – at least for part of the case files.

Deduction is considered a most distinguished philosophical method of categorisation [Kant 1787/1904, pp. 104–130, §§14–27]. The most paradigmatic modern contribution to the conceptualisation of deduction as a philosophical method relates to law [Fichte 1966/70]. Concerning the NS Special Courts, a categorisation alongside the legal bases is particularly interesting as their qualities are remarkably diverse. First of all, we chose a legal theoretical distinction between the positive law [Kelsen 2017] and laws and legal regulations, respectively, in the sense of Carl Schmitt’s [Schmitt 1922/2021] and other NS legal thinkers’ theory of law and society. It marks the peculiarity of NS Special Courts within the contemporary judicial system of the Reich that in the same way of their institutionalisation beyond the regular courts, the NS regime began with its legislation of special penal regulations besides the regular criminal laws that had been systematically codified in the German Nation State since 1871 [RStGB 1871, 1935]. The proceedings before the Special Courts concerned first of all exclusively such “crimes”, or – to be more semantically neutral – actions that only had been criminalised by a certain kind of NS special penal regulations. A closer look on the complex process of the Nazi Seizure of Power shows that the NS regime relaunched special penal legislation even before relaunching the Weimar-invented Special Courts [Verordnung 1933], at first by means of the “Reichstag Fire Decree” [Reichstag Fire Decree 1933]. The scientific estimation of this ordinance as “the fundamental exceptional law on which the National Socialist dictatorship was primarily based until its collapse” [Bracher 1960, p. 82] shows the principal significance of this kind of repression for the establishment of NS dictatorship.

D-1: The most important and in over twelve years of NS special penal judiciary most often applied special legal regulation [Hüttenberger 1981] became the so-called “Heimtückegesetz” [Heimtückegesetz 1934]. Thereby, the term “Heimtücke”, i.e., “insidiousness”, must be understood as a typical NS dysphemism. This law criminalised every kind of verbal or written critical comment on the NS regime or on certain Nazi politicians – whereas in regular German criminal law, “insidiousness” is actually an essential characteristic of murder [RStGB 1871, 1935, §211,2]. We define the early NS special penal law regulations as our category D-1.

D-2: As another facet of their peculiar role within the NS system of persecution and repression, Special Courts were increasingly used for “regular crimes”. The longer the NS regime lasted, the more frequent such cases were brought before it. This phenomenon is well known by existing research literature (e.g., [Dörner 1998, pp. 36–39]; [Wogersien 2007]). Apart from a few exceptions (e.g., [Graczyk 2021, pp. 250–330]), concrete numbers or data on this topic have not been presented yet. The Mannheim Special Court shows, for example: In 1933 18.6 % of “regular” first instance judgements (like theft or murder) were tried before it, but in 1943 this percentage would increase up to 43.9 % [Oehler 1997, pp. 36–37]. Precise statistical analyses on the question of how often in 1933 still valid laws [RStGB 1871, 1935] were applied represent a desideratum. By the example of Special Court Munich, this can be approached through our computer-based investigation of its archive inventory from the 1970s (cf. section below). Besides the codified law, Munich Special Court even applied rather marginal regulations and decrees of regular penal law for its judgements, up to the late Weimar “Animal Protection Act” [Animal Protection Act 1930] or exclusive Bavarian state laws – what surprises in the face of anti-federal Nazi system –, like the “Slaughter Tax Act” [Slaughter Tax Act 1930].

D-3: Research so far has also thematised that the beginning of World War II meant a significant incision in the history of NS special judiciary. Already since 1938, a veritable boost of new specific NS special penal law regulations began to unfold. This “second wave” of NS special criminal law-making is correlating with the sheer increase in the number of cases brought to trial before, e.g., the Munich Special Court since 1939: Between 1933 and 1938, 3,123 proceedings were carried out, but in the years 1939–1945 they reached 6,823. This tremendously increasing repression no longer exclusively referred to the regime’s “own” German “Volksgemeinschaft” but to foreigners (mostly forced labourers and prisoners of war), too, since the annexations of 1938 and 1939 and particularly since Germany’s war of aggression. This justifies the formation of another deductive-legal-historical category, especially since these new products of NS special penal law regulation [4] – in contrast to the more generally applicable “Heimtückegesetz”– are concerned with more or less specific war-related deviance or resistance, or with the racial or ethnic affiliation of prosecuted people itself, respectively:

D-4: Finally, in the context of “total war”, the judicial practice of NS Special Courts reached out even to the legal basis of another non-regular jurisdiction instance, namely to the proprietary penal law of the German “Wehrmacht”. The Military Penal Code [MStGB 1943], originally launched by early Kaiserreich, would be the legal basis for three judgments of the Munich Special Court in 1944/45.

For an understanding of the historical phenomenon of NS Special Courts and their – since 1939 once again changed – role within the Nazi repression system, a deductive categorisation along the concrete legal bases of proceedings may provide a substantial epistemological profit. It might even go to the core of the complex relationship between National Socialism and law as well as justice, respectively, in a diachronic perspective. As far as the legal bases of the Munich Special Court during World War II are concerned, we are dealing with a simultaneous set of legal sources that arose at different times and, according to legal theory, are to be evaluated differently. This multimodality of legal bases, some of which are applied simultaneously in proceedings, is to be modelled in categories D-5, D-6, and D-7 (concrete numbers cf. below, “Comparison” and Figure 5). Especially since 1939, Munich Special Court often would apply D-2 legal bases – above all paragraphs of the RStGB [RStGB 1871, 1935] – in addition to the D-3 NS special penal law regulations whenever possible. Also, the “Heimtückegesetz” (1934) of D-1 would still be applied extensively in addition to the new special penal law regulations from 1938 onwards; even the “Reichstag Fire Decree” (1933), still kept being applied by the Munich Special Court, even in times of “total war” (17 times, remarkably in most cases against Jehovah’s Witnesses).

Ideally, deductive categorisation seems to be the most rational method of understanding the actual nature of this historical phenomenon. In our present research, however, it indeed proves to be only partially purposeful. Formally, the façade of a rationally acting judiciary was maintained even under the conditions of “total war” [Lahusen 2022, pp. 141–185], illustrated with the example of Aachen Special Court), although the Munich NS Special Court was undoubtedly part of a state of injustice. The problem lies rather in the fact that in our research we do not work primarily on the basis of the court records themselves but with the help of secondary records prepared in the 1970s (cf. above, introduction to the source material). As is the case with many digital humanities projects today, the editors of the “pre-digital” database of the BayStMUK-“Archive Inventory” changed their metadata scheme in the course of data entry. Thus, it happens that the laws/ordinances/legal paragraphs that had been applied by the NS Special Court were only recorded in the inventory records from the beginning of Part 3 (1939) onwards during the scientific retrospective of the 1970s [BayStMUK 1975–1977, 1976, p. 619; 1977, p. 2349]. In the records on the trials and discontinued proceedings of the years 1933–1938, no legal basis is given in the records of the “Archive Inventory” at all and for the phase from 1938 to 1945 are also some instances where such explicit information is missing (specifically in 473 cases). Nonetheless, even for the proceedings where the record entries offer no paragraph in the “Archive Inventory”, the offence recorded in each case can always be used to draw conclusions about an unambiguous criminal offence and the applied written law (cf. below, comparison). For the research design of our project, this means that a categorisation along the legal basis of the special court proceedings actually only makes sense for the period from the beginning of 1939 onwards, even though 6,830 of the entire 9,954 cases captured by the “Archive Inventory” took place after the beginning of 1939.

For our research and the formation of meaningful and, above all, useful categories for the quantitative historical analysis of the (secondary) source corpus, this lack of data and missing information means that we have to find a more pragmatic methodology in order to focus on the actual object of investigation – i.e., the Munich NS Special Court and its repressive practice or the “everyday” resistance and deviance of the wider population in the “Third Reich” and the “wartime society” ([Wildt 2007]; [Wildt 2019, pp. 302–325]) that manifested itself before this court.

Through a random sample of cases for close reading and the intensive study of the source material in the context of the digitisation and data modelling process, we could gain an overview of the content. This insight allowed us to think of several topics relevant to categorising the charges by applying an intuitive system. In some cases, the demarcation was relatively easy: A huge part of the charges concern negative remarks against the regime (e.g., the claim that the Reichstag was set on fire by National Socialists) and leading figures thereof (e.g., the allegation that Hitler and Göring are gay). Another very clear category deals with social interactions with prisoners of war. Non-war-related inventory records contain crimes from bodily harm, e.g., in the form of murder, to theft. Evaluating the charge content showed that the wording used to describe the individual accusations is in many cases standardised. This offered the possibility of sorting the whole corpus with a few selected keywords per category – so our hypothesis.

A first trial run of the approach – by defining eight categories, by analysing a sample of 90 randomly chosen cases and looking for similarities in their content for classification (1-bodily harm, 2-fraud, 3-economic crimes, like theft, 4-war-related economic crimes, like illicit butchering, 5-interactions with prisoners of war, 6-sophisticated regime critique, written and oral resistance, 7-conspiracy, like pamphlet distribution, 8-negative remarks) – revealed a disproportional surplus with over 40 % of cases concerning negative remarks. To cater for deeper statistical analysis in the future, differentiating within this main part of accusations became relevant in order to gain better insights into the relationship between accusations, results, and socio-political backgrounds of the accused. So, through a review and quality control loop, we decided to refine categories, what they each comprise as well as which keywords to use. Thereby, comparing the success and failure of predicted results from single keywords places induction in a pragmatic context (cf. [Holland et al. 1986]).

The query or category allocation was carried out on the basis of the normalised charge text. The number of keywords necessary per category varies. Through a review loop for each, the most efficient and concise word variations were sought and found. Working with a spreadsheet, the course of action was to test keyword variations on the column with the normalised charge text before incorporating the best version into a nested formula, i.e., the most efficient one in terms of the number of cases that could be allocated without compromising the results for other categories. In some cases, only parts of words and phrases were used in order to incorporate as many different word and verb forms as possible without compromising the results by accidentally returning cases dealing with wholly different content. This process led to the following categories with the different topics included in each.

Due to varying phrasing of the editors from the 1970s for the charges within the inventory records (e.g., for negative remarks we have “schimpfen”/ranting, “negativ, abfällig äußern”/speaking derogatorily, “kritisch äußern”/critical remarks etc.), partly due to the long process of collecting data and creating the inventory and partly due to the fact, that different persons worked on them as well as still remaining OCR-errors, it helped to keep the single category keywords as short as possible. Thus, we encapsulated varying spellings and verb forms as well as common OCR-errors like “a” instead of “ä” within one query. As the search for the keywords is operated via the normalised and shortened text, which is sometimes very individual in its wording concerning the charges, additional approximately 100 cases, which otherwise would result in being relegated to the group of uncategorised records, can be allocated to the group of “negative remarks and hearsay” by querying for “gesagt haben soll” (reported to have said) in the full text at the end. In the end, 451 records remain uncategorised (E-0) due to poor quality of the OCR-text or because an accusation related to certain allegations is mentioned only once, e.g., there is one accusation of poisoning wells, a trope used against Jews since the Middle Ages.

The resulting formula of nested if-then queries [6] is quite intricate but also streamlined concerning efficiency. Nonetheless, one has to keep in mind that the sequence within the query influences the result. If we, for example, retrieve cases by filtering for charges of the category “bodily harm and morality”, like fornication/“Unzucht” or murder, first, we would also put cases in the same category where hearsay in relation to such crimes is concerned. The assigning of the cases to single crime categories by particularly chosen keywords is subject to individual biases as some – like the unauthorised wearing of badges and uniforms (ultimately put into E-7) – could be placed in different categories depending on where we as editors choose to put them. Therefore, we also developed a more “objective” inductive-computational approach.

The inductive-computational approach aims at producing a categorisation with little manual intervention by applying algorithms from natural language processing (NLP), machine learning, and clustering. Most algorithms used must be parameterised which requires expert knowledge, structured exploration of the parameter space and experimentation.

Starting from the baseline CSV file (cf. “Workflow” above), we first trained a doc2vec model [Le and Mikolov 2014] using the gensim library for Python 3 [Rehurek and Petr 2010]. The algorithm defines an ample set of parameters, some of which have a strong influence on the quality of the document embeddings produced. Our goal was to cluster similar inventory records close together and far apart from non-similar ones. Finding the right set of parameters requires trial and error: define a parameter configuration, train a doc2vec model, obtain embeddings for all documents, and evaluate the results visually. At this point, we would like to note that we added a manual intervention: doc2vec does not solve a classification task (there is no ground truth) but, instead, generates document embeddings. Thus, no metric exists to assess the quality formally and automatically in terms of classification accuracy of the embeddings produced. Therefore, we selected a doc2vec model manually: we first computed embeddings for all documents and plotted them in 2D space separately for each model; we then used visual inspection to select a model that would generate a non-homogeneous distribution of documents with clear clusters. Figure 2 exemplifies how we explored the parameter space to select a model: each subplot represents a model trained with a distinct set of parameters and shows the document embeddings generated by the model projected into 2D space.

Even a small set of parameters result in a huge space to explore: for example, five parameters with three possible values each would require 35 = 243 models to be trained and assessed. To limit the number of parameter configurations and, thus, the number of models to be trained, we relied on our expertise and followed the advice given by Lau [Lau and Baldwin 2016] on recommendations for optimal doc2vec hyper-parameter settings. We then systematically explored the parameter space, by first defining for each parameter a range of allowed values and then iterating each possible combination of all parameters. For example, when allowing a = {1, 2, 3} and b = {4, 5, 6}, we would obtain nine possible combinations of a and b, namely {1, 4}, {1, 5}, … {3, 5}, {3, 6}. After training only a small number of doc2vec models and inspecting the results visually, we were able to identify some parameter values which would always produce insufficient results, independently of all other parameters. For example, disabling negative sampling or switching on hierarchical softmax would always produce a homogeneous distribution of documents, without distinct clusters, which is undesirable for our purpose. We excluded these parameters from the exploration by setting a fixed value and shrinking the parameter space to explore by orders of magnitude.

Besides exploring doc2vec parameters configurations, the selection of a training corpus has a significant impact on the quality of the models produced. Initially, we trained our model on our own inventory records and used the same set of documents to obtain document embeddings. Our corpus consisted of 8,532 valid inventory records with an average length of 5.3 words. We were able to improve our models to show more pronounced clusters by using a different corpus during training, containing 1,454 legal documents from between 1933 and 1945 with an average length of 444 words [7].

After training and manually selecting a doc2vec model through visual inspection [8], we were ready for the final step: clustering. First, we inferred a high-dimensional embedding for each inventory record and then projected each 300-dimensional vector into 2D space by applying t-SNE [van der Maaten and Hinton 2008]. We then transformed our data into quantiles [9] before clustering by applying DBSCAN [Ester et al. 1996]. We used pandas [McKinney 2010], numpy [Harris 2020], and sklearn [Pedregosa 2011] to implement all steps. The image below shows a clustering of document embeddings representing our inventory records.

Each dot represents a doc2vec embedding which corresponds to a single inventory record. Similar documents are located close together and form clusters. Each cluster corresponds to one category. Outliers are not part of any cluster. Note that the image shows results from a particular choice of parameters used for training the doc2vec model and for clustering document embeddings. We manually chose a set of parameters. Different parameters would produce a significantly different output.

The number of clusters/categories in this approach was much higher than in the other two approaches. This was because documents were grouped solely based on contextual information found in the texts, whereas the other approaches grouped legally related documents together, even if they were appearing in completely unrelated contexts. For example, category E-5 (prisoner of war-related, pragmatic-explorative approach) collects all accusations concerning contact with prisoners of war, whereas the three categories C-35, C-37, C-38 are produced in the inductive-computational approach, seemingly distinguishing between the nationality of the prisoners.

The following table (Figure 4) shows statistics for all 46 clusters sorted by the number of documents within each cluster. The third column shows the relative number of documents contained within each cluster. The right-most column shows the top three most frequent words occurring within the cluster. Almost 16 % of all documents belong to C-00 which collects outliers and corresponds to the “uncategorised” category in the other two approaches. The clusters are unevenly sized. The eight largest clusters (17 % of all clusters, not including C-00) contain 50 % of all categorised documents. Out of these eight clusters, five contain the word “Äußerung” (remark) within the top three cluster words which correspond to 35 % of all documents. This clearly shows that a large proportion of all accusations concerned remarks about the regime, leaders, the war etc.