Holocaust survivors' testimonies provide a rich source of evidence about the personal experiences of survivors who witnessed and endured horrors during the Nazi genocide of Jews and other persecuted groups. The narratives contain references to the emotions experienced when describing memories of people, places, and events. Analysing the spatiality of these human emotions enables us to understand how they are connected to the places around them. We focus on fear, sadness, anger, as well as joy to examine the interplay of these emotional experiences by multiple individuals at di erent places and times and in di erent circumstances. Understanding these complex emotional landscapes, especially from very large collections of textual data requires a carefully designed technique that can e ectively and e ciently apply existing and new technologies. In this work, therefore, we explore the possibility of extracting and analysing these emotions as well as their related geographies by applying a combination of natural language processing methods including large language models.
During what became known as the Holocaust, the Nazi regime systematically persecuted and
murdered millions of Jews and other targeted groups during World War II. The survivors of
these events have shared their personal experiences and memories through various forms
of testimony, such as interviews, memoirs, and artworks. These provide valuable sources of
historical and cultural knowledge, as well as emotional and psychological insight into the
human condition under extreme circumstances. One aspect that can be explored in Holocaust
survivors’ testimonies is the spatial and temporal dimensions of the emotions expressed about
people, places, and events, otherwise known as emotional geography. As Guy Miron signals
in the case of German Jews, individuals experienced Nazi spatial control “both as a feeling
and as a physical reality” [
However, studying the emotional geography of Holocaust survivors is not an easy task, especially when dealing with very large collections of video, audio, and textual data (the collection we work with includes over 55,000 recorded interviews). Therefore, there is a need for e ective methods and tools to extract and analyse the emotions and geographies from the Holocaust survivors’ testimonies and to visualise and present the results in a meaningful and accessible way. This work aims to address this challenge by applying a combination of existing natural language processing (NLP) techniques, such as sentiment analysis, emotion detection, named entity recognition, geocoding, and geovisualisation, to a corpus of videorecorded testimonies from the USC Shoah Foundation. The key research questions that drive this work include: • How can we use NLP techniques - possibly leveraging large language models - to extract and analyse expressions of emotions in Holocaust testimonies? • How much ‘fear’ compared to other emotions - ‘sadness’, ‘anger’, and ‘joy’ - is contained in each testimony narrative? • How does the expression of ‘fear’ change across the narrative sequence of each testimony? This paper presents the outcome of exploratory work which contributes:: • An application of NLP methods to the study of the emotional geography of Holocaust survivors, and a demonstration of the potential and challenges of using these techniques for this purpose. • An analysis of the emotions and geographies expressed in the testimonies, and a discovery of new patterns and insights that can enhance our understanding of the Holocaust and its survivors.
This work draws on and contributes to two main elds of research: emotional geography and
natural language processing. Emotional geography is an interdisciplinary eld that studies
the relationship between human emotions and space, place, and environment [
Researchers have long advocated exploring the physicality of the event. Beorn et al. signalled
that the Holocaust was ‘rooted in speci c physical spaces, times, and landscape’ and also
‘characterised by a spatiality of the process - concentration, deportation, dispersal, dislocation’
[
For this work, we have focused on a small portion of the Holocaust Survivors’ Testimonies (HST) which comprises a random selection of transcripts of one thousand oral history interviews (about 21 million words) undertaken by the USC Shoah Foundation Visual History Archive1 in the 1990s.
The transcripts follow a similar format that includes a series of questions posed by the interviewer and the corresponding answers from the interviewee who is a survivor. According to the 1Information about the USC Shoah Foundation Visual History Archive can be found https://s .usc.edu/what-we-do/ collections interview guideline [12], each interview focuses on the individual’s experiences during the Holocaust which are explored in a broadly chronological order. Each interview – generally of around two hours duration – devotes approximately 20 percent of the time to pre-war life, 60 percent to wartime experiences focused on the events of the Holocaust, and 20 percent to post-war life [13]. In short, these are not full life histories, but more focused interviews asking about wartime experiences across a series of sites of incarceration or hiding. These sites serve as anchors in the narratives that describe survivors’ wartime trajectories.
In this exploratory work, we used only a small sample of 10 out of the 1000 testimonies described above as a preliminary test to assess the salience of the emotion of fear, its spatial distribution, and its relation to other important emotional states. For this experiment, we purposely restrict our sample size and the number of emotions examined to allow for the application of domain expert knowledge to check the consistency and accuracy of our methods.
This extraction pipeline used in this work was a version of the framework, illustrated in Figure 1, de ned by [11] for extracting place names and geographical feature nouns from text through named entity recognition for the Lake District corpora, and a more generalised version of the framework described in another work [10]. The key elements of the framework include processes for enhancing an o -the-shelf named-entity recogniser, Spacy [14], with lists of items (e.g. place names, geographical feature nouns, dates and time, sentiments or emotions, locative adverbs, and spatial prepositions). This was achieved by adding the Spacy’s rule-based module annotation module ‘EntityRuler‘ to the models pipeline thereby improving the ability of its named entity recogniser to identify and extract our pre-de ned spatial entities and concepts. The enhanced model is subsequently applied to similar texts to perform surface-level extraction of spatial elements and even sentiment-bearing words for basic sentiment analysis.
Using this technique, we could identify some key geospatial elements like the countries, cities, concentration camps, and even geonouns mentioned in the testimonies. Table 3 shows some of the top elements identi ed in the answers provided by the survivors in the 2 randomly selected testimonies - id=‘01’ and id=‘02’. For this work however, we explore emotion beyond the binary classi cations of positive and negative sentiments to capture more ne-grained a ect analysis but restricted to mainly ‘fear’ but ‘sadness’, ‘anger’, and ‘joy’ which were also included for richer comparative analysis. These classi cations are a selection from the 6 basic classes of emotion (‘fear’, ‘anger’, ‘disgust’, ‘joy’, ‘sadness’, ‘surprise’) identi ed by Ekman and Friesen [15] as well as the ‘neutral’ class, which are popularly used in NLP for studies in a ect classi cation and analysis.
The transcribed testimonies are available in plain text formats generally structured for conversational turn-taking mainly between the interviewer and the survivor but often with interjection from crew members or family members of the survivors. We rst pre-processed the les to separate the questions by the interviewer and the responses by the survivor. We then focused only on these responses (see Table 2 for details) given by the survivor for the emotional classication. Since some of these responses were quite long and span many paragraphs, we split them into sentences to keep the context to a manageable scope. 3.3.1. Classification Models There was no available bespoke emotion model trained on the Holocaust corpora or a dataset for training one that we were aware of during this work. Besides, this was meant to be an exploratory work with no intention of building new models at the rst stage. Hence, we proceeded with existing o -the-shelf emotion classi cation models as well as generic large language models. We therefore applied the following three models to classify the sentence contexts extracted from survivors’ testimonies and select the most ‘voted’ class out of the three.
The rst model we applied was Hartmann’s transformer model [16]. The model which is a version of the DistilRoBERTa-base ne-tuned on about 20k observations extracted from six diverse emotion datasets. The selected observations are fairly distributed across the emotion classes. It is a lightweight transformer model that is freely available and easy to use, hence the choice. Another key contribution of this model is the provision of the emotion scores as a probability distribution across the labels.
From the output of the transformer model, we retained only the contexts that were labelled with the four classes we are studying in this paper - ‘fear’, ‘sadness’, ‘anger’, and ‘joy’. Given the need for some form of evaluation, these classi ed sentences were then passed individually to a variant of the GPT-3.5 Turbo models [17] gpt-3.5-turbo-instruct as well as the GPT-4 model [18] with the same prompt basically to get another opinion. We tried several prompts to see which works best but Figure 2 shows an example of the prompt we settled for the LLM experiments.
Finally, the accepted class for each sentence is determined by selecting the most ‘voted’ class. This requires that at least two of the models will predict a particular class for it to be selected, otherwise ‘none’. While we do not expect the approach to guarantee high-quality annotations we consider it a stop-gap method that is good enough to enable us to gain some interesting insights from the Holocaust testimonies.
We performed the classi cation with the models and ‘voting’ process, and confusion matrices can be seen in Figure 6. Our nal output is a set of 5,461 sentences with emotion scores and labels for fear, sadness, anger, and joy. Table 3 shows examples of some of the sentences and the scores assigned by di erent models while Figure 4 is a plot of the distributions of the four emotions or ‘none’ as predicted by the models. The label ‘none’ is not present in the predictions from roberta because we only extracted instances of the four labels from its output. As shown in Figure 4, the emotions of ‘fear’ and ‘sadness’ tend to trend in the same pattern consistently across the model outputs.
When comparing the variations of the negative emotions fear, sadness, anger and positive emotion joy across the narrative sequence as shown in Figure 5, some interesting patterns emerge. Survivors tend to express positive emotions at the beginning and end of the interviews while negative emotions tend to follow a bimodal, increasing pattern before falling sharply towards the end of the interviews. This follows from the structure of interviews, where survivors begin by discussing pre-war childhood memories and end with re ections on more recent life events such as their own children. Across all interviews, positive and negative emotions are inversely correlated, with negative emotions of fear and sadness appearing together most often.
Finally, Table 4 shows the connections we observe in these 10 texts between named places and the emotions observed. We split the results into three groups: 1) places, 2) geographic feature nouns, and 3) camps, and for each group, we list the overall top ve in each group, plus the lists associated most with each of the four emotions. This serves to illustrate the potential of the technique for linking emotions to places of the di erent types under analysis.
In this work, we explored the idea of developing a computational framework for analysing the emotional landscapes of a textual narrative in a more nuanced form beyond the classi cation into positive and negative sentiments which has been undertaken previously. We applied our method to a sample of 10 Holocaust survivors’ testimonies speci cally focusing on only four out of Ekmans 6 emotion classes [15] - ‘fear’, ‘sadness’, ‘anger’, and ‘joy’. These were chosen for no particular reason other than that they seem like sensible themes for Holocaust-related study. The interaction of these emotions with the geography of the space they were connected to - named places, camps, and geographical features were also extracted and analysed.
Given that we had no available model or previously existing annotated dataset for our task, we applied an e cient combination of models consisting of a lightweight transformer-based model and two large language models to support our annotation process. We are aware of the limitations of this method and will focus on building a more robust, bespoke, and in-domain 5. They drove him out, out of the camp, 0.0668 0.0226 0.3528 0.0072 anger fear and they said, go back. joy sad joy none model for more accurate labelling and analysis. However, it produced su ciently good results for initial exploratory research. We will ne-tune our processes and scale up our study to analyse a larger collection of Holocaust testimonies to gain more insight into the complex interplay of emotions in a spatio-temporal context, potentially ne-tune a model on Holocaust data, to link emotions to those who experienced them, and to explore zero-shot learning methodologies in the future.
We thank the anonymous reviewers for their comments on our paper submission. The project is funded in the UK from 2022 to 2025 by ESRC, project reference: ES/W003473/1. We also acknowledge the input and advice from the other members of the project team in generating requirements for our research presented here and the UCREL Hex team for providing the compute needs for this project[19]. More details of the project can be found on the website: https://spacetimenarratives.github.io/
Places (Russia,11%); (Israel,7%); (Germany,6%); (York,5%); (Poland,5%) (Russia,23%); (Germany,9%); (Braunschweig,5%); (Palestine,5%); (Buchenwald,3%) (Russia,9%); (Germany,9%); (Israel,9%); (York,7%); (Hungary,5%) (Russia,32%); (Germany,12%); (Palestine,12%); (Poland,6%); (Boston,6%) (Poland,7%); (Vienna,7%); (Israel,7%); (York,6%); (Oswego,6%) Geographic feature nouns Overall (2950) (end,6%), (school,4%); (camp,3%); (train,2%); (house,2%) fear (118) (train,4%); (camp,3%); (rough,3%); (side,3%); (end,3%) sadness (119) (camp,5%); (end,5%); (school,4%); (shed,3%); (place,3%) anger (117) (top,4%); (shed,4%); (head,3%); (stop,3%); (door,3%); joy (121) (ice,11%); (school,8%); (well,3%); (port,3%); (house,3%); Camps Overall (2950) (Russia,21%); (Auschwitz,16%); (Germany,12%); (Poland,9%); (Czechoslovakia,6%) fear (89) (Auschwitz,27%); (Russia,26%); (Germany,10%); (Czechoslovakia,5%); (Braunschweig,5%) sadness (91) (Auschwitz,24%); (Russia,14%); (Germany,14%); (Czechoslovakia,8%); (Theresienstadt,8%) anger (82) (Russia,46%); (Germany,17%); (Poland,8%); (Czechoslovakia,4%); (Auschwitz,4%) joy (73) (Poland,22%); (Russia,13%); (Bar,11%); (Germany,11%); (Auschwitz,4%) [10] I. Ezeani, P. Rayson, I. Gregory, E. Haris, A. Cohn, J. Stell, T. Cole, J. Taylor, D. Bodenhamer, N. Devadasan, et al., Towards an extensible framework for understanding spatial narratives, in: Proceedings of the 7th ACM SIGSPATIAL International Workshop on Geospatial Humanities, 2023, pp. 1–10. [11] I. Ezeani, P. Rayson, I. N. Gregory, Extracting imprecise geographical and temporal references from journey narratives., in: Text2Story@ ECIR, 2023, pp. 113–118. [12] J. Ringelheim, A. Donahue, E. Hedlund, A. Rubin, Oral history interview guidelines, 2007. URL: https://www.ushmm.org/m/pdfs/20121003-oral-history-interview-guide.pdf, accessed: 04-01-2024. [13] N. Shenker, Reframing Holocaust Testimony, Indiana University Press, Indiana University, 2015. [14] M. Honnibal, I. Montani, S. Van Landeghem, A. Boyd, spaCy: Industrial-strength Natural
Language Processing in Python (2020). doi:10.5281/zenodo.1212303. [15] P. Ekman, W. V. Friesen, Constants across cultures in the face and emotion., Journal of personality and social psychology 17 (1971) 124. [16] J. Hartmann, Emotion English DistilRoBERTa-base, https://huggingface.co/j-hartmann/ emotion-english-distilroberta-base/, 2022. [17] T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, D. Amodei, Language models are few-shot learners, 2020. arXiv:2005.14165. [18] OpenAI, et al., Gpt-4 technical report, 2023. arXiv:2303.08774. [19] J. Vidler, P. Rayson, UCREL - Hex; A Shared, Hybrid Multiprocessor System, https://github. com/UCREL/hex, Accessed: 2024.