Daejeon, July 27–31
Testimonies from Holocaust survivors are essential sources for Holocaust research, as they document events and perspectives that official documents and perpetrator reports either omit or distort. They are also key to overcoming the perception of passive victims and exploring decision-making and agency. Furthermore, they are a central component of remembrance culture and help keep memories alive for future generations. This is particularly critical considering the much-discussed end of contemporary witness testimony and the decline of knowledge about the Holocaust and active participation in remembrance culture(Claims Conference 2025; Stiftung Erinnerung, Verantwortung und Zukunft 2024/2025). At the same time, research shows that Gen Z, while interested, has limited knowledge of the Nazi era and that personal relevance and emotional engagement with the past are crucial (Arolsen Archives 2022).
Emotion-based and emotion-historical approaches are increasingly debated in historical education and research. While computational methods such as sentiment analysis are established in digital history, their application to personal historical accounts remains the exception – primarily because there is a lack of training corpora enriched with sentiment information and because even for lexicon-based approaches, resources for German-language texts are scarce compared to English (Flüh 2019).Problems with non-English words in manual labeling have also been noted (Woods et al.). While testimonies from Holocaust survivors have already been examined using machine learning methods in a handful of studies, this has so far been done primarily based on oral history interviews (Blanke et al. 2020). However, these interviews were created after the events and thus have a very different relationship to the memories than the more immediate experiences captured in diaries. There is also a lack of critical methodological comparison between lexicon-based and machine-based methods, particularly in this domain.Bernal et al. (2024) also mention the potential for qualitative and quantitative approaches to complement each other. While there are some comparative studies, these have a primarily technical perspective with a focus on performance, do not deal with historical texts, and in some cases have only been published as preprints or in predatory journals. A comparison of sentiment analysis approaches has been presented at DH2022 (Grisot et al. 2022).
The aim of this contribution is to move beyond the status quo of broad sentiment classifications and to achieve a deeper understanding of personal accounts. This is done through an in-depth qualitative and quantitative sentiment and emotion analysis of Holocaust testimonies recorded both during the Holocaust and in retrospect. Besides critically reflecting on computer-aided and machine learning methods for evaluating and interpreting personal accounts of the Holocaust, it aims to contribute to digital tool and source criticism, as well as to overarching discussions on the use of AI in remembrance culture.
This is explored through a case study on the diaries written by Martha Glass, a Holocaust survivor from Hamburg,from 1943 to 1945 (Glass 1943–1945). The diaries provide rare insight into experiences, observations, and feelings during her imprisonment in Theresienstadt, as they were written during the events. The detailed descriptions in her diaries testify to her attempts to maintain normality under inhumane living conditions and present writing as a survival strategy. Later, in 1947 and 1953, Martha Glass wrote travel diaries during the crossing from Bremerhaven to New York and during her three-month stay in Germany and Austria in the summer of 1953 (Glass 1947; Glass 1953). A first test study, during which we conducted sentiment analysis of the Theresienstadt diaries with SentiWS (Remus et al. 2010), classified them as positive, as many terms were interpreted as positive and phrasing, which requires contextual knowledge to classify, was not recognized (Geibel 2025).Similarly, Khulusi et al. (2022) report unexpectedly positive results in their study. This leads to the underlying hypothesis that sentiment analysis of these texts can systematically yield misleadingly positive results and to this ongoing research.This study is funded by the Incubator Fund of the German national research infrastructure for history, NFDI4Memory, https://4memory.de/aktivitaeten/incubator-funds. The main research questions are: To which extent can the automated evaluation and analysis of personal accounts of the Holocaustlead to false positive perceptions? Which methodological approach can be used to examine and understand possible misrepresentations?
The methodological approach we want to use is: 1) Close reading of the diaries with the aim of identifying references to emotions and delimiting semantic units for sentiment analysis. 2) Topic modeling for an overview of the subjects addressed in the diaries. 3) Evaluation of sentiment lexica and models for the analysis. 4) Execution and evaluation of the lexicon-based and the machine-learning approach.
As a first result of the close reading (Method. Step 1, Fig. 1), terminology and named emotions have been identified and categorized (e.g., hope, gratitude, shock, fear). A key distinction was made between emotions explicitly expressed by Martha Glass and those only implicitly present in the text.
To explore the general thematic structure of the diary (Method. Step 2), a BERTopic pipeline was applied to the TEI-encoded text, segmenting it into overlapping two-sentence windows and using the paraphrase-multilingual-mpnet-base-v2 model for sentence embeddings. The resulting topics (Fig. 2) cluster predominantly around elementary concerns of daily survival (food and ration allocation, hygiene, illness, and weather), alongside recurring clusters relating to holidays, social gatherings, and religious practice. This distribution reflects both the realities of life under conditions of extreme confinement and deprivation on the one hand, and the deliberate maintenance of social and cultural routines on the other.
For computer-driven processing of the emotions (Method. Step 3), it is essential to find a suitable annotation approach. Given the complexity, a multidimensional model is needed to adequately capture the range and nuance of these expressions. Therefore, the VAD model (Mehrabian et al. 1974) was chosen, as it captures emotions along three dimensions (Valence, Arousal, Dominance) rather than simply classifying them as positive or negative.
Drawing on close reading, topic modeling, and lexicon- and machine-learning-based sentiment analysis, the presentation will discuss both the methodological approach and preliminary results of its implementation. It will also address how such case studies can be used in teaching to help students develop critical skills in working with computer- and AI-supported text analysis.