DHQ: Digital Humanities Quarterly
2022
Volume 16 Number 3
Volume 16 Number 3
Studying Large-Scale Behavioral Differences in Auschwitz-Birkenau with Simulation of Gendered Narratives
Tim Hempel
, Freie Universität Berlin, Department of Mathematics and Computer
Science / Department of Physics 
https://orcid.org/0000-0002-0073-9353
https://orcid.org/0000-0002-0073-9353Krishna Somandepalli
<ksoman_at_google_dot_com>, Google Research https://orcid.org/0000-0002-2845-1079
https://orcid.org/0000-0002-2845-1079Shri Narayanan
, USC Viterbi School of Engineering https://orcid.org/0000-0002-1052-6204
https://orcid.org/0000-0002-1052-6204Abstract
In Auschwitz-Birkenau men and women were detained separately; anecdotal evidence suggests that they behaved differently. However, producing evidence based insights into victims' behavior is challenging. Perpetrators frequently destroyed camp documentations; victims' perspective remains dispersed in thousands of oral history interviews with survivors. Listening to, watching, or reading these thousands of interviews is not viable, and there is no established computational approach to gather systematic evidence from a large number of interviews. In this study, by applying methods and concepts of molecular physics, we developed a conceptual framework and computational approach to study thousands of human stories and we investigated 6628 interviews by survivors of the Auschwitz-Birkenau death camp. We applied the concept of state space and the Markov State Model to model the ensemble of 6628 testimonies. The Markov State Model along with the Transition Path Theory allowed us to compare the way women and men remember their time in the camp. We found that acts of solidarity and social bonds are the most important topics in their testimonies. However, we found that women are much more likely to address these topics. We provide systematic evidence that not only were women more likely to recall solidarity and social relations in their belated testimonies but they were also more likely to perform acts of solidarity and form social bonds in Auschwitz-Birkenau. Oral history interviews with Holocaust survivors constitute an important digital cultural heritage that documents one of the darkest moments in human history; generally, oral history collections are ubiquitous sources of modern history and significant assets of libraries and archives. We anticipate that our conceptual and computational framework will contribute not only to the understanding of gender behavior but also to the exploration of oral history as a cultural heritage, as well as to the computational study of narratives. This paper presents novel synergies between history, computer science, and physics, and it aims to stimulate further collaborations between these fields.
I. Introduction
1. General background
The Auschwitz Concentration Camp Complex was liberated seventy-seven years ago on the
27th January 1945. Since then the camp where an estimated 1.1 million people were
murdered has became the symbol of the Holocaust. Despite the large number of studies and
memoirs by survivors, the question to what extent solidarity and social bonds existed
among victims is still open [Pingel 2010] [Piper 2009]. While some survivors claimed
that friendships, mutual support, and solidarity among victims thrived, others have
vehemently denied it.[1] It has been
recently argued that solidarity and social bonds were key to survive Nazi concentration
camps; no one could survive without the help of others and mutual aid was a form of
opposition among victims [Swiebocki 2000] [Bacon 2017]. Nevertheless, it has been also
noted, “conditions varied greatly across the vast Auschwitz complex” [Wachsmann 2015, 342]. Some victims were therefore more likely to experience solidarity and form social
bonds than others. Based on the close-reading of some victims’ accounts, some
researchers have for instance suggested that group cohesion and solidarity were stronger
among women, who were detained separately, than among men in Nazi death camps; however,
in the lack of evidence, others have rejected this hypothesis [Pingel 2010] [Schwarz 1998].
Even though a
comparative study of how men and women behaved could enhance our understanding of the
Holocaust from the victims’ perspective, producing evidence based insights into the
world of Nazi concentration camps is challenging. Perpetrators destroyed camp
documentation and victims’ perspective of the past is dispersed in tens of thousands of
oral history interviews they gave. By watching a handful
of interviews, we can gather only anecdotal evidence. At the moment, there is no
reliable methodology to gather systematic evidence from a large number of oral history
interviews. Traditional methods of corpus and computational linguistics cannot alone
address this challenge. To produce systematic evidence from thousands of oral history
interviews, it is important to account not only what is discussed but also how it is
discussed. In other words, it is necessary to analyze how the narratives in thousands of
oral history interviews unfold. We therefore developed a novel conceptual and
computational approach adopted from the study of molecular processes. This novel method
enabled us to study how narratives unfold in 2500 hours of oral history interviews with
Auschwitz survivors. Even more importantly, it provided us with systematic evidence that
among women solidarity and social bonds were stronger in Auschwitz.
Before presenting the results of our study, we introduce the dataset we worked with, the
challenges it raises, and the way we tackled these challenges conceptually and
computationally.
2. Our dataset and its historical context
In this study we analyzed testimonies given by 6628 Jewish victims (2033 men and 4595
women) born in 13 different countries. The 6628 victims we worked with all stayed in the
Birkenau sub-camp of the Auschwitz complex.[2] In the Auschwitz camp system, Birkenau was
the largest and the most lethal sub-camp where most Jewish victims were housed and
perished.[3] After arriving at the camp, the SS [Die
Schutzstaffel], the Nazi paramilitary organization in charge of running death
camps, selected those men and women who were able to work. Those who did not pass the
arrival selection were murdered with Cyclon-B gas and cremated. Those few people who
passed the arrival selection were deprived of their belongings, disinfected, registered
(often tattooed) and forced to live in harsh conditions; men and women were housed in
different sections of Birkenau. Victims often died due to starvation, various diseases,
physical violence, and the extreme cold. During their stay in Birkenau, most victims had
to work [Hayes 2000]. Sometimes the SS could, however, not keep pace with the influx of
new deportees, and victims were waiting to get some work. There was also a large group
of victims in Birkenau who never worked. Auschwitz-Birkenau was also a transit camp from
where victims were sent to other camps. During the time (which could be quite long) that
victims were waiting for transfer, they were, in fact, not given any work; they just
hung around in the camp.
Our study is based on a data set that consists of oral history
interviews with Holocaust victims. The USC Shoah Foundation provided us with this data
set from its collection of approximately 50.000 Holocaust video testimonies recorded
between 1981 and 2019. These video testimonies are preserved in the Visual History
Archive of the Foundation at the University of Southern California in Los Angeles.[4] Precisely, the Foundation gave us access
to all of its interviews with Auschwitz survivors, which is a collection of 12200
Interviews in 26 different languages recorded in 47 different countries. We were also
provided with basic biodata (gender and place of birth) about each survivor whose
testimony is included. Given that we limited our investigation to Jewish victims who
stayed in the Birkenau subcamp of the Auschwitz complex, we worked with 6829 interviews
from the original data pool. These interviews were conducted in 26 different languages
in 43 different countries.
Until today, only a small portion of USC Shoah Foundation’s testimonies
have been transcribed and published [Toth 2021]. In this study we could not therefore
work with interview transcripts. We worked with the topical annotation of testimonies by
the Foundation. In the last three decades, the professional staff of the Shoah
Foundation indexed the minute based segments of each testimony with topic words from a
designated thesaurus or controlled vocabularies. This thesaurus, a very large formally
defined set of topic words, has been developed by the Foundation in the 1990s; the goal
of the thesaurus is to cover all possible aspects of genocide experiences, including
acts of solidarity and memories of social bonds [Gustman 1998]. Professional annotators
(or indexers as the Foundation calls them) watched testimony videos and indexed their
content with the help of topic words from the thesaurus in a purpose-built testimony
database; visual and audio features of testimonies were not indexed.
For the purposes of this research the Foundation provided us access to the raw data
underlying their testimony database.[5] The Foundation has reproduced, specifically for
our study, all testimony segments, understood as a set of topic words (see a more
technical description in the Implementation section), in which Auschwitz survivors
discuss their stay in the camp. In our study, we relied exclusively on the
English language topic words. Hence a limitation of our study is given by the fact that
we were unable to study how topics related to solidarity and social bonds are discussed
across testimonies in different languages. Another limitation of our study is that we
were unable to work with the visual and audio aspects of the testimonies (see further
discussion of these limitations in the Conclusion section).
3. Two Computational Challenges Related to Survivor Testimonies
Since the late 1970s archives and museums have been recording oral history interviews
with Holocaust victims [Ecker 2007]. The large-scale study of these interviews could
offer a unique insight into victims’ experience [Greenspan 2012]. Nevertheless, oral
history testimonies raise two key computational and conceptual challenges.
Researchers would need to account for not only which topics are present in individual
testimonies but also how these topics are discussed across thousands of testimonies. On
the one hand, survivors’ oral history interviews consist of questions and answers. On
the other hand, a survivor testimony is a sequence of topics that together form a
narrative about the past. Survivor testimonies can be thus seen as an ensemble of
individual narrative processes. Given that interviews unfold over time, the ensemble of
narrative processes has an underlying dynamic. The first computational challenge that
survivor testimonies raise can be summarized in the following two questions:
- How can we model the ensemble of narrative processes understood as sequences of topics unfolding over time?
- How can we measure the probability and the importance of topics in the ensemble of narrative processes?
The second challenge relates to the latter question. Survivor testimonies are mediated.
Survivors gave testimonies decades after their liberation; their testimonies may have
been shaped by post-Holocaust culture, including films and novels. Furthermore,
survivors recall deeply traumatic memories, which are characterized by flashbacks and
the suppression of certain particularly painful themes [Langer 1993]. Some survivors
might overemphasize the significance of certain topics such as solidarity and social
bonds, while others might underestimate it. Finally, the existing oral history
testimonies are only the tip of the iceberg. The overwhelming majority of victims never
had the chance to speak. For all these reasons, we needed a computational model that can
take the mediated nature of survivor testimonies into account when measuring the
probability and the importance of topics in narrative processes.
II. Methodology
1. Our Conceptual Framework
We developed a conceptual framework to tackle the problems above that is centered around
three aspects: narrative system, narrative flux, and narrative pathway.
We define the narrative system as the ensemble of all possible topics that survivors
address, including transitions between these topics. We represent the narrative system
as a state space with topics as discrete states [Alpay and Gohberg 2006]. Each
interview, which is viewed as a sequence of consecutive topics, is a “walking” over the
state space. In other words, each interview is a possible realization of the narrative
system. Once all possible realizations were learned (see below), we could estimate the
probability of each topic in the ensemble of all interviews. This approach also allowed
us to compare testimonies of men and women. We viewed the interviews of each gender
group as different instances of the same narrative system; we could thus compare the
probabilities of topics that different instances yield (see the Implementation
section).
To get insights into dynamics underlying narratives of survivors, we drew on the
concepts of narrative pathway and narrative flux. In our conceptual framework, the
narrative pathway describes a given sequence of states (i.e., topics) connecting two
topics. The narrative flux quantifies the relative amount (percentage) of traffic going
through a given narrative pathway. Consider the following example.
Two key topics in testimonies are arrival at a death camp (arrival) and discussion of
social relations (social bonds). Testimonies can get from the topic arrival to the topic
social bonds in many different ways. For instance, after discussing arrival, they can
discuss the topic selection, and then social bonds. This corresponds to the narrative
pathway arrival → selections → social bonds. The narrative flux
considers all possible pathways between arrival and social bonds
in the ensemble of all interviews (or interviews of women and interviews of men); then
it tells us what fraction percentage goes through selections. In
short, through the notion of narrative pathway and narrative flux, we can get insight
into how narrative processes are likely to unfold in the ensemble of many
testimonies.
2. Our Computational Model
From molecular physics we adopt the Markov State Model (MSM) and we perform a
computational analysis of the ensemble of all interviews as a narrative system [Chodera and Noé 2014] [Husic and Pande 2018]. Specifically, we use the Transition Path Theory
(TPT) framework to model the dynamics of a narrative system [Vanden-Eijnden 2006].
In the context of our research, we interpret the ensemble of interviews as realizations
of a Markovian dynamic system; this is defined over a state space consisting of all
topics that survivors address. By drawing on the Markov State Model we obtain the
stationary probability (also known as equilibrium state or stable probability) of each
topic. In simple terms, this describes the overall probability of hearing a survivor
mentioning a topic if we keep listening to interviews for a reasonably long time (see
the Implementation section for technical details on and definition of overall
probability).
We uncovered narrative pathways and quantified narrative flux by using the TPT, a
quantitative framework based on MSMs [Metzner et al. 2009]. In the context of studying
narrations with MSMs, TPT is used as a means to organize and summarize large numbers of
individual narrations (i.e. sequences of consecutive topics or narrative pathways as
realized in thousands of interviews) and to make them accessible to our interpretation.
In practice, TPT associates a number value (called the flux) to any possible narration
(or any possible sequence of topics). This number tells us how prevalent a given
narrative pathway is in the underlying corpus of interviews, i.e., it summarizes how
likely it is that an interviewee reported about topics in a certain order. We can
therefore use TPT to sort narrations by their prevalence, which also allows us to
identify the most common narrations from the data.
This method was borrowed from dynamic systems modelling of molecular processes where TPT
studies ensembles of reactive trajectories, i.e., trajectories that make transitions
between defined starting state A and end state B.
In our work, we considered interviews as such trajectories. TPT therefore analyses
narrations from a starting to an ending topic. We have chosen topics associated with
camp detention as the starting and topics associated with liberation as the ending
points. The narrations are therefore handled consistently and our results show prevalent
narrations that are aligned to the Auschwitz-Birkenau experiences.
We used the technique of pathway decomposition, elaborated as part of TPT, to uncover
the most important sequences of states (narrative pathways in our research) and to
quantify the amount of traffic (or flux) that goes through a certain sequence of states.
As a whole, we used the TPT to decompose the ensemble of all possible narrative pathways
by ranking each of the pathways by their flux. This enabled us to identify the “busiest”
narrative pathways that men and women used between a pair of topics.
Finally, based on our TPT analysis, we also model the narrative system as a directed
graph. In this graph each topic (or a state in the state space model described above) is
a node; the move of interviews from topic A to topic B is represented as an edge, which
is weighted in terms of the probability of moving to topic B after topic A. To measure
the importance and influence of each topic within the narrative system represented as a
directed graph, we apply two different metrics. First, we measure the PageRank value of
each topic as a node in the directed graph. The PageRank value is often used to measure
the influence of humans in social networks, as well as the importance of webpages in
Google search algorithms. By the same token as the PageRank value gives insight into the
importance of webpages by modelling internet users movements from webpage to webpage, it
reveals the importance of topics in survivor testimonies by modelling how these topics
follow each other in testimonies. Actually, the PageRank values of nodes in a directed
graph are closely related to the stationary probabilities described above; if a weighted
directed graph represents a Markovian dynamic system, the PageRank values of nodes in
the graph are the dumped stationary probabilities of states in the corresponding
Markovian system [Page et al. 1999]. The major advantage of the PageRank value is that
it takes into consideration inbound and outbound links to a given node (topic in the
context of our research), as well as the importance of connected nodes; hence the
PageRank value expresses how embedded a given node is in the entire network [Lee et al. 2013]. Second, we calculate the normalized closeness centrality of each node. This
evaluates importance by summing the shortest paths from a given node; in other words, it
expresses how connected a given node (or topic in the context of our research) is
[Freeman 1978].
III. Our Results
1. Understanding the significance of topics related to solidarity and social bonds in survivor narratives
The 6628 Birkenau survivors we worked with discuss hundreds of topics when recalling
their stay in the death camp; these topics cover many different aspects of camp life
ranging from sanitary conditions to everyday interactions. As explained above, in the
last three decades the Shoah Foundations’ indexer team watched these testimonies and
annotated each testimony segment by means of topic words from the Foundation’s topical
thesaurus. In total, there are 727 different topic words attached to interview segments
related to Birkenau. However, many of these topics occur only in a few interviews. As a
first step, a human agent with domain knowledge analyzed the 727 topic words and reduced
them to 101 higher level topics (see Methodology). Next, we estimated the stationary
probability of each higher level topic (when using probability we always mean the
stationary probability of a topic). We specifically focused on two higher level topics:
social bonds and aid giving. The former includes topics discussing social relations
(friendships, camaraderie, non-intimate social relations etc.) in the camp; the latter
relates to various acts of solidarity (food sharing, clothing provision, sustenance
provision, etc.) We also calculated their closeness centrality and PageRank Value.
In the following, we ranked the topics according to stationary probability in descending
order. This allowed us to identify the most probable topics in survivor narratives (see
Figure 1).
Figure 1.
Wordcloud featuring the most probable topics in testimonies of
Auschwitz-Birkenau survivors identified with a Markov State Model. The font size of
topics is adjusted according to their stationary probability. The wordcloud demonstrates
that two key topics related to solidarity and social bonds (aid giving and social bonds
in red) are among the most probable topics. The word cloud also demonstrates that they
are as probable as tattoo and more probable than clothing (in red).Not surprisingly, one of the most probable topics that survivors discuss is selections. This is the third most common topic after living
conditions and camp procedures. Selection was a wanton and highly feared practice in
Nazi death camps. All victims arriving at camps first underwent a selection procedure.
Selection was also a daily routine in Nazi concentration camps; the SS kept eliminating
the unfit and sending them to gas chambers. Since selection was a highly probable event,
we used it as a reference point to assess the probability of topics describing
solidarity and social relations. Compared to social bonds, the
probability of selections is 3.4 times higher. Compared to aid giving, the probability of selections
is 2.28 times higher. Aid giving and social
bonds are not as probable as selections. But they are
still very probable if we take into consideration that on average selections is 92 times more probable than any other topic. The overall
importance of social relations and aid
giving are also demonstrated by the fact that they are part of the upper
quantile in the ranking of all topics in terms of their stationary probabilities. Aid giving is the 14th most probable topic and social bonds is the 18th most probable topic. The ranking of topics by PageRank
value resulted in similar results.
We also measured the normalized closeness centrality of social
bonds and aid giving in the directed graph representing
the narrative system. We found that the topic aid giving is the
4th most important element of survivor narratives and social
bonds is the 20th most important element in survivor narratives. They are both
part of the upper quantile.
Due to the limitations of the Markovian framework (see Implementation section), we could
not harness specific forms of social bonds and acts of solidarity. Our approach was
restricted to broader topics. In the next section, as part of a comparative analysis of
narratives by women and men, we will combine the Markovian framework with inferential
statistics, which will give insights into details of social relations and acts of
solidarity.
Our findings highlight that survivors discuss a great deal of social activity and
solidarity when recalling memories of Birkenau. It is important to underline that
survivors were not asked by interviewers to address solidarity and social bonds in the
camp. It was the survivors' own decision to address these topics. In light of this, we
think that what is recalled in the following testimony excerpt cannot be viewed as an
anecdotal or nostalgic memory but in fact describes a relatively common experience:
We became friends with everybody — anybody and everybody. We — we — we really cared for each other. We did everything possible to try and encourage each other. If somebody wanted to hang themselves because they couldn't take it any more; we somehow talked 'em out of it. If somebody, for instance, said, I'm giving up, I can't do it any more, we somehow talked them out of it. We maybe shared a little piece of bread. The bread was mostly made out of sawdust, but somehow we put a little piece of bread in his mouth. We did all kinds of those kinds of things. [Gross 1996, Tape 2, 00:28:34]
2. Comparative analysis of solidarity and social relations in narratives of men and women
Men and women were detained and housed in different sections of Birkenau. We therefore
examined and compared the probability of topics related to acts of solidarity and social
bonds in narratives of women and men (see the Methodology section). To prove that
differences were not due to either chance or sample bias, we used the method of
bootstrapping (see below and the Methodology section) [Efron et al. 2004]. Again, we compared the importance of topics related to acts of
solidarity and social bonds in narratives of men and women by drawing on normalized
closeness centrality and the PageRank value. We also used the Fisher’s exact test to
assess whether the difference is statistically significant (p value 0.05) [Fisher 1922]. Finally, we measured the strength of association between gender and each topic
with odds ratio analysis (Morris and Gardner, 1988). This enabled us to focus on
specific forms of social bonds and acts of solidarity (see Figure 2).
Figure 2.
Topics more likely to be discussed by women. Strength of
association of topic words for gender of Auschwitz-Birkenau survivors was obtained by
using odds ratio analysis; topic words are positively or negatively correlated for
strengths of association >1 or <1, respectively; dot line shows a degree of
association equals to one; the difference is statistically significant (p <0.05).
Figure 3.
Topics more likely to be discussed by men. Strength of
association of topic words for gender of Auschwitz-Birkenau survivors was obtained by
using odds ratio analysis; topic words are positively or negatively correlated for
strengths of association >1 or <1, respectively; dot line shows a degree of
association equals to one; the difference is statistically significant (p <0.05).We found that the topic social bonds is much more probable in
testimonies of women; in their testimonies we are significantly more likely to hear
about discussions of social relations in the camp. Our findings through bootstrapping is
particularly revealing (see Figure 3). In short, over thousands of times we randomly
sampled testimonies of women and men; each time we trained two Markov State Models over
the same state space and measured the stationary probability of every topic. We observed
that even if we randomly select testimonies of men and women, in the testimonies of
women social bonds will always be more probable. The measurement
of strength of association also confirmed this finding. Social
bonds is positively correlated with women and negatively correlated with men.
Finally, the normalized closeness centrality and the PagreRank value are also higher
when testimonies of men and women are modeled as directed graphs.
In the case of the topic aid giving the difference between men
and women is less strong. The probability of this topic is only by 7 percent higher in
testimonies of women, even though the difference between men and women is statistically
significant and remains consistent through the randomized sampling (see Figure 3).
Similarly, the normalized closeness centrality and the PageRank value are higher for
women.
Figure 4.
Bootstrap sample distributions for stationary probabilities for aid giving. The orange represents women, the blue men, and the dotted line the model estimate. Note that the distributions do not overlap. On the one hand, this
means that the difference between stationary probabilities is significant given our
model. On the other hand, it shows that if we randomly sample interviews by men and
women over thousands of times, the stationary probability of topics above will always be
higher in the women sample. In other words, the difference between men and women is
consistent and not due to sample bias.
Figure 5.
Bootstrap sample distributions for stationary probabilities for social bonds. Note that the distributions do not overlap. The orange represents women, the blue men, and the dotted line the model estimate. On the one hand, this
means that the difference between stationary probabilities is significant given our
model. On the other hand, it shows that if we randomly sample interviews by men and
women over thousands of times, the stationary probability of topics above will always be
higher in the women sample. In other words, the difference between men and women is
consistent and not due to sample bias.With the help of statistical significance testing and odds ratio analysis, we also examined specific forms of solidarity and social bonds (see Figure 2 and 3). The Shoah Foundation’s keyword
thesaurus has various terms to describe different types of solidarity. For instance, the
topic word food sharing describes moments when a survivor recalls
an episode of food sharing in the camp. We found that in testimonies of women we are
much more likely to hear about episodes of food sharing. Similarly, women are more
likely to discuss memories about another person providing them with clothes (the topic
of clothing provision). Other topics related to solidarity, such
as for instance medical care provision and sustenance provision, are also positively correlated with women, though they
lack statistical significance. As a whole, we were unable to find any topic related to
solidarity, which is positively correlated with men.
The Shoah Foundation’s keyword thesaurus is less rich in terms of details about social
bonds. Its keywords friends and friendships
describe moments when survivors discuss intimate bonds in the camp. Again, we found
a very strong and statistically significant positive correlation with women and a
negative correlation with men. Similarly, women are more likely to recall less intimate
social bonds from the camp. Other topics related to sociability and solidarity feature a
similar difference. For instance, topics describing cultural activity, or rumors are all
positively correlated with women (see Figure 2) and feature statistical significance.
Women are also more likely to discuss suicide attempts in the camp.
At the same time, we also found interesting commonalities between men and women. It is a
leitmotif in testimonies that victims warned each other. We did not find a significant
difference in terms of the probability of hearing about this topic (see Figure 7).
Another commonality is stealing (see Figure 6); victims often discuss memories about
other victims stealing from them; men and women are equally likely to recall this
experience. Finally, preferential treatment by camp functionaries is also equally likely
to be recalled by both men and women (see Figure 8).
Figure 6.
Bootstrap sample distributions for stationary probabilities for stealing. The orange represents women, the blue men, and the dotted line the model estimate. Note that the distributions overlap. This
means that the difference between stationary probabilities is not significant given our
model. If we randomly sample interviews by men and women over thousands of times, the
stationary probability of topics above will sometimes be equal in the two samples. 
Figure 7.
Bootstrap sample distributions for stationary probabilities for warning. The orange represents women, the blue men, and the dotted line the model estimate. Note that the distributions overlap. This
means that the difference between stationary probabilities is not significant given our
model. If we randomly sample interviews by men and women over thousands of times, the
stationary probability of topics above will sometimes be equal in the two samples. 
Figure 8.
Bootstrap sample distributions for stationary probabilities for preferential treatment. The orange represents women, the blue men, and the dotted line the model estimate. Note that the distributions overlap. This
means that the difference between stationary probabilities is not significant given our
model. If we randomly sample interviews by men and women over thousands of times, the
stationary probability of topics above will sometimes be equal in the two samples. We also looked at those topics that dominate men's recalling of Auschwitz-Birkenau (see
Figure 3). We found that men are more likely to recall topics related to physical
violence (killings, beatings, executions, etc). Furthermore, we found that topics
related to various clandestine activities (bribery, resistance, covert activities) are
also more correlated with men than with women.
Furthermore, we examined whether the difference between women and men can be also
observed if we compare men and women of the same nationalities. We found that the
probability of topics related to social bonds and solidarity are either higher for women
or there is no difference. We were unable to find any nationality where the probability
of topics related to solidarity and social bonds would be higher for men.
By further partitioning the data, we also surveyed women and men who did not work while
they were in Birkenau. We observed the same trend as above. In narratives of women who
did not report any forced labour activity, topics related to solidarity and social bonds
are more probable.
Finally, we also studied the testimonies of those women and men who had to work. Given
that the forced labour type and its difficulty varied a lot in Nazi death camps, we
first divided the forced labour activities into three groups based on physical effort
demands: easy work (secretarial work, science/research-related forced labor, translation
etc), medium hard work (tattooing, factory work, gas chamber cleaning, etc), very hard
physical work (mining, road and rail construction, etc). In our dataset, percentages of
women and men who did the same type of work are quite similar. The majority of men
(16.5% of all men survivors) and women survivors (14.5% of all women survivors) reported
medium hard work. Even when comparing the working population of Birkenau, we could
observe the trend that in narratives of women, topics related to social relations,
friendships, and food sharing are more probable. However, there was no difference in the
probability of the topic aid giving. This is equally probable in
narratives of men and women who worked while they stayed in Birkenau.
3. Narrative pathways leading to and from topics of solidarity and social bonds
Using TPT, we found that both men and women discuss solidarity in the context of the
everyday living environment (i.e., camp barrack) and forced labour activity, which is an
interesting commonality. However, when we examined the narrative flux that goes through
the most important narrative pathways leading to acts of solidarity in narratives of
women and men, we found an important difference: women are more likely than men to
discuss acts of solidarity in the context of camp barracks (see Figure 10 and Figure
12).
Figure 9.
Topics related to solidarity following topics related to forced labour. Assessing the narrational context of topics related to solidarity and to
social bonds by simulating how likely they are followed by topics related to forced
labour activity in testimonies
of women and men as a function of time lag. The plots show that women are more likely to
discuss social bonds and acts of solidarity following the discussion of the everyday
living environment in Auschwitz than men; whereas men are more likely to address these
topics after addressing forced labor activity. Probabilities converge after a time lag
of 3-5 minutes; hence, the plots highlight that the difference between men and women can
be also observed once the dynamical systems representing the narratives of men and women
reach equilibrium.
Figure 10.
Topics
related to solidarity following topics related to living environment. Assessing the narrational context of topics related to solidarity and to
social bonds by simulating how likely they are followed by topics related to the everyday living environment in testimonies
of women and men as a function of time lag. The plots show that women are more likely to
discuss social bonds and acts of solidarity following the discussion of the everyday
living environment in Auschwitz than men; whereas men are more likely to address these
topics after addressing forced labor activity. Probabilities converge after a time lag
of 3-5 minutes; hence, the plots highlight that the difference between men and women can
be also observed once the dynamical systems representing the narratives of men and women
reach equilibrium.Figure 11.
Topics related to social bonds following topics related to forced labour. Assessing the narrational context of topics related to solidarity and to
social bonds by simulating how likely they are followed by topics related to forced
labour activity in testimonies
of women and men as a function of time lag. The plots show that women are more likely to
discuss social bonds and acts of solidarity following the discussion of the everyday
living environment in Auschwitz than men; whereas men are more likely to address these
topics after addressing forced labor activity. Probabilities converge after a time lag
of 3-5 minutes; hence, the plots highlight that the difference between men and women can
be also observed once the dynamical systems representing the narratives of men and women
reach equilibrium.
Figure 12.
Topics
related to social bonds following topics related to living environment. Assessing the narrational context of topics related to solidarity and to
social bonds by simulating how likely they are followed by topics related to the everyday living environment in testimonies
of women and men as a function of time lag. The plots show that women are more likely to
discuss social bonds and acts of solidarity following the discussion of the everyday
living environment in Auschwitz than men; whereas men are more likely to address these
topics after addressing forced labor activity. Probabilities converge after a time lag
of 3-5 minutes; hence, the plots highlight that the difference between men and women can
be also observed once the dynamical systems representing the narratives of men and women
reach equilibrium.We examined all possible pathways leading from the topic arrival
to the topic aid giving in narratives of men and women. In
narratives of women, more traffic is found going through the pathways in which aid giving is preceded by the topic living
conditions. For instance, the following narrative pathway is present in both men
and women testimonies, but again, in testimonies of women the narrative flux (7.8%) is
higher than in testimonies of men (5.2%):
arrival → selections → intake procedures → living conditions → aid giving
The investigation of narrative pathways connecting the topic living
conditions with the topic aid giving also confirmed that
women are more likely to connect solidarity with the world of their barracks.
We also found that in testimonies of women the discussion of social relations is more
likely to be preceded by the discussion of living environments. The following narrative
pathway is present in testimonies of both men and women. However, in testimonies of
women the amount of narrative flux is much higher (14.12%) than in testimonies of men
(9.7%).
arrival → selections→ camp procedures → living conditions → social bonds
Even though both men and women often discuss social relations and solidarity following
forced labour activity, the narrative flux is much higher in the case of men (see Figure
9 and Figure 11). For instance, we examined all possible narrative pathways that connect
the topic arrival with the
Figure 13.
Network depiction of narrative pathways connecting the topic Arrival to
Auschwitz with Social bonds in testimonies of women as
extracted from Transition Path Theory pathway decomposition analysis. Circle sizes are
adjusted according to the stationary probabilities of topics and connection arrow
strengths visualize the amount of narrative flux going through a link. In testimonies of
men most of the narrative flux is going from forced labor and living conditions to
social relations, in testimonies of women, most of the narrative flux is going from food
and living conditions to social relations.
Figure 14.
Network depiction of narrative pathways connecting the topic Arrival to
Auschwitz with Social bonds in testimonies of men as
extracted from Transition Path Theory pathway decomposition analysis. Circle sizes are
adjusted according to the stationary probabilities of topics and connection arrow
strengths visualize the amount of narrative flux going through a link. In testimonies of
men most of the narrative flux is going from forced labor and living conditions to
social relations, in testimonies of women, most of the narrative flux is going from food
and living conditions to social relations.topic social relations and involve forced
labor priming social relations (see Figure 13 and Figure 14). In case of women, the following path is the most
important one (the narrative flux is 3.6%):
arrival→camp procedures→living conditions→forced labor→social
relations
In the case of men, the following path is the most important one (the narrative flux
going through them is 6.8%).
arrival→selections→camp
procedures→tattoos→forced labor
→social relations
To sum up, the way men and women discuss social relations and solidarity feature notable
similarities. They both tend to discuss them following topics related to living and
housing conditions and forced labor. Nevertheless, we also found that women are more
likely to discuss social activity and acts of solidarity in the context of the everyday
living environment; by contrast men tend to discuss solidarity and social relations in
the context of forced labour. All this suggests that women mainly socialized in their
barracks, whereas men experienced solidarity and social interactions in the working
environment.
3. Conclusion
1. Discussion of results
The availability of large survivor testimony datasets is now making it possible to
accomplish evidence based studies of victims’ experience during the Holocaust. A key
problem has been how best to study thousands of testimonies together and produce
systematic evidence. We modeled the testimonies of 6628 Auschwitz-Birkenau survivors
with MSM, which 'summarized' the narratives underlying these testimonies. This gave us
an unprecedented insight into victims’ experience.
We found that acts of solidarity and social ties among victims were not at all uncommon
in Auschwitz-Birkenau. The significance of this finding can be understood in light of
what researchers and survivors often described as dehumanization. The following words by
a Hungarian survivor, demonstrate what dehumanization meant in Auschwitz-Birkenau:
And, you know, it's very hard to explain how logic or emotion doesn't come into it. You're fighting for your life. You're fighting for a piece of bread. You don't think, now what do I think is happening? It's happening. You're glad today you weren't beaten. You're glad yesterday you had a mouthful more bread. You — you don't speculate. So this one was run onto the wire and got killed. Fine, she's. She's lucky. I'm not lucky. Because I haven't got the guts to do it. So that is the mentality. You know, that's how you live. You're not a human. You're dehumanized. You don't think. You follow instructions like an animal, not even — an animal wouldn't listen to it. So you become a piece of robot really, which if you don't where emotions and hunger meets or takes over. [Neumann 1996, Tape 4, 00:21:00]
By reducing victims into a dehumanized state, the SS deliberately thwarted victims’
otherwise natural effort to take care of each other. Furthermore, the SS deliberately
undermined social bonds between victims by frequently transferring them between forced
labour squads, camps, and barracks. As the historian, Gideon Greif, has put it, “the
Nazi camp system aimed to break down all solidarity human relations as well as the
belief in the possibility of self-determined actions” [Greif 1999, 242]. The
question whether some form of solidarity and sociability could still exist in
Auschwitz-Birkenau therefore translates into the question of how successful the SS was
in its effort to destroy the human nature of victims. Accounts such as the one above or
the testimony excerpt cited at the beginning of this paper have often convinced
researchers that solidarity and deeper human ties were actually anomalies in Auschwitz.
For instance, the sociologist, Wolfgang Sofsky, has recently described the Nazi death
camp as a place where victims lived in “a social state of nature in which the person who
prevailed was the one who was stronger” [Sofsky 2013, 162]. This view is in
contrast with the accounts of many survivors and with the findings of some historians.
Again, as one survivor cited above said, “we did everything possible to try and
encourage each other.” The hypothesis that solidarity and social bonds among victims
were just anomalies in Auschwitz is also in contrast with our findings.
In fact, the 6628 survivors in our dataset do speak about a great deal of solidarity and
social interactions. Acts of solidarity and memories of social life are highly important
elements of their memories. Their importance and high probability within the narrative
system indicates that they were also significant and relatively frequent elements of
camp life. Even if statistical frequencies and probabilities of topics measured by
modelling thousands of narratives cannot be equated with the broader and non-statistical
concept of importance, they can still be treated as reliable indicators giving
information about the world that these narratives recall. As a matter of fact, we will
never know how victims lived in Nazi death camps. But by analyzing thousands of survivor
narratives, we can attempt to construct estimates about basic aspects of life in Nazi
death camps.
Our computational approach also allowed us to compare the narratives of women and men
and to shed light on a key difference: in testimonies of women solidarity and social
relations are more probable and more important topics than in testimonies of men. Based
on the close reading of a handful of testimonies, this gender difference has been
already noticed but lacked the backing of wide-scale quantitative evidence. In this
study, we provide systematic evidence that in narratives of women survivors of
Auschwitz-Birkenau solidarity and social relations are much more important topics.
Nevertheless, this finding leads to a key problem that any study of belated memories
about a past event involves. Memory sometimes alters the way humans see their past;
therefore, we need to ask how can we produce evidence about the past by drawing on
belated memories? Taking this question into the context of this study, to what extent
does the fact that in belated memories of women solidarity and social bonds are more
probable, indicate that they were also more prevalent in the women sections of Birkenau?
We tackled this problem with the help of bootstrapping, i.e., with the help of
randomized trials. It is possible that some women overestimate the importance of
solidarity and social relations in the past and some men underestimate their
significance; it is possible that memory and post-Holocaust experience (including
watching specific films and reading Holocaust novels) have "sweetened" the experiences
made in the past for some women and triggered bitterness in some men. Nevertheless, with
the help of randomized samples we pointed out that the difference between men and women
is consistent. If our initial finding that solidarity and social relations are more
probable in narratives of women would have been due to the sweetening effect of
nostalgic memory or due to sampling bias, the difference between men and women would
have faded over thousands of randomized trials.
Our findings suggest that women generally tend to show more solidarity in extreme
circumstances such as detainment. However, it is important to highlight a historical
factor as a result of which social adaptation for Jewish men in Birkenau must have been
particularly difficult. By the time that mass deportations of European Jewry to
Auschwitz-Birkenau began, many of the young and middle aged Jewish men had already been
in forced labour service [Gruner 2006]. Most of the men who were deported to
Auschwitz-Birkenau were older men; they had therefore very slim chances to survive the
selection process at the arrival. As a result, most probably there were more women of
the same nationalities than men of the same nationalities. Being surrounded by women of
the same nationality, a woman could find solidarity and friends much easier than a man.
To corroborate this explanation, we would however need to know more about whom women
formed friendships with and who offered solidarity and help to them.[6]
2. Limitations of our methodology
This points out the limitation of this study. We could study only those topics that the
Shoah Foundation’s index system covers. Unfortunately, the index system does not offer
any information about whom victims were friends with. Neither does it contain keywords
that signal conflicts between victims. We could not therefore study whether men or women
are more likely to recall animosity. Similarly, the use of generalized keywords to
describe human behavior has its own limitations since they cannot describe the richness
and subtleties of human behavior. For instance, friendship and aid giving are complex
concepts and behavioral patterns with many different possible meanings and nuances that
a general keyword cannot describe. This limitation can in part explain the difference
between men and women. When discussing social relations in the camp, women and men might
express their ideas and memories very differently. Where women use friendship, men might
for instance use other words that suggest less intimate social bonds. Unfortunately, the
unified keyword system cannot account for such differences and the close-listening of
2500 hours of survivor testimonies is not viable.
Another limitation of our study is that we could not work with
interview transcripts in the original languages. Interviewees of different languages
might have expressed similar experiences in different ways, which in turn might have
produced unevenness in the topical annotation. For instance, victims of different
linguistic backgrounds may use very different linguistic strategies to account
experiences of social bonds and aid giving.
Finally, topical annotation cannot address a key feature of traumatic narration. This is
silence. Traumatic memories often remain suppressed and speakers lack the linguistic
ability to articulate them. Hence, the bulk of traumatic memory can hardly be described
with unified topic words.
3. Implications for future research
Despite these limitations, the use of the Markov State Model has helped to tackle an old
problem that among others, the Italian survivor of Auschwitz and the author of many
Holocaust novels, Primo Levi (1917 - 1987), has pinpointed. This is the fallacy of the
individual memory of a survivor. As Levi said, the human memory is a marvelous
instrument but it makes mistakes [Levi 1986, 14]. The fallacy of individual memories
can be counterbalanced with the study of the memory of many survivors. But the study of
many individual memories poses its own challenges: testimonies are mediated and
unbalanced in terms of gender for instance; they miss the voice of those who did not
survive or did not give a testimony. To face these challenges, in this paper we applied
the Markov State Model combined with the Transition Path Theory. This enabled us to
approach thousands of individual testimonies as expressions of the collective
experience. Most importantly, it enabled us to compare narratives of women and men.
Future work can use the methods we present in this paper to study victims’ experience in
other Nazi death camps.
Future work can also rely on our methodological and conceptual approach to study a great
variety of big data involving sequences of states in the humanities. Such datasets are
ubiquitous in history and in other humanities fields. Datasets of national biographies,
for instance the Oxford Dictionary of National Biography, could be studied by
representing historical actors’ life trajectories as sequences of states. Similarly, the
circulation of artistic objects among different owners (so-called “object biographies”)
can be represented as sequences of states. The concept of state space and the Markov
State Model can be applied to model and analyze these sequences of states. Most
importantly, a methodology and conceptual framework applied to understand processes both
in nature and in human civilizations can offer new bridges between science and
humanities in the future.
IV. Implementation
1. Introduction
In this paper we discuss our
methodology and its implementation in two different sections. The Methodology section
above aimed to offer a general overview of our methodology. Here we describe
step-by-step how we implemented our methodology. In addition to the step-by-step
description, Figure 15 and 16 also help understand the key steps in the implementation
process. At the same time, for a complete understanding of the implementation, readers
might need some background in probabilistic modeling, which we cannot provide here.
Figure 15.
Methodological Overview 1: From individual interviews to the
modelling of 6628 interviews as a Markovian dynamic system. A, a single interview is
processed into 1 minute segments that are assigned to topic terms each (color-coded).
Interviews are therefore represented as sequences of consecutive topics (or narrative
pathways). B The whole interview corpus consists of 6628 interviews that are converted
into 6628 series of topic terms. B, Transition matrix that summarises the probabilities
of hearing about a given topic following another topic as learned from all available
interviews. The transition matrix can also be represented as a directed graph; nodes are
the topics and the edge weights are the probabilities that an interviewee move from one
topic to another, it therefore stores all possible narrative trajectories that a
recalling of Auschwitz can follow in the corpus of all interviews. D. The “busiest” (or
the most prevalent) narrative pathways obtained by means of Transition path theory
(TPT). TPT extracts prevalent narrations from the MSM transition matrix; a narration is
here depicted as a (possibly non-linear) sequence of topics that goes from left to
right. Please note that this figure is a conceptual overview and does not represent real
interview data.
Figure 16.
Methodological Overview 2: Key steps to train and analyse a Markov State Model
representing narratives of Auschwitz survivors.2. Preprocessing
In this research project we worked with the topic words that the trained expert human
annotators of the Shoah Foundation had assigned to minute based segments of interviews
in our data set. As a first step, we applied a feature engineering procedure; we grouped
topic words and formed higher level topics. In total, we identified 101 higher level
topics. Our feature selection procedure followed the keyword thesaurus of the Shoah
Foundation. In this thesaurus each topic is clearly defined and formally not
overlapping.
The interview dataset we worked with is a set of matrices. Each matrix represents an
interview; the rows correspond to minute-based interview segments and the columns
represent the topic words associated (higher level topics as described above) with
segments as a “one-hot-encoding”. This segment-topic word matrix is similar to
document-term matrices used in bag-of-words models of document collections in text
mining. We can also view each segment-topic word matrix as a sequence of topics that
form a discrete narrative pathway; hence we first represented the interviews as sets of
consecutive topics.
Interview_1 = {topic 3, topic 1, topic 2 ………. topic n}
Interview_2 = {topic 4, topic 3, topic 2 ………. topic n}
Sometimes an interview segment can feature multiple topics. Consider the following
example:
Interview_3 = {topic 3 -> topic 1, topic 4 -> topic 2}
Given that multiple topics may be assigned to certain segments, there is a fuzziness in
the interview topic assignment. We address this fuzziness by randomly sampling from the
topics of multi-topic segments. For instance, in case two topics are assigned to a
segment, we treat it such that they could have occurred with either topic.
Interview_3_A = {topic 3 -> topic 1-> topic 2}
Interview_3_B = {topic 3 -> topic 4-> topic 2}
As we do not alter the sequence of topics, the dynamics is conserved. In other words, we
interpret an interview as a set of pathways that are compressed into a single pathway
with multi-topic segments, and therefore disentangle it into single pathways.
3. The construction of a state space and its Markov modelling
Next, we construct a state space from the sets of discrete narrative pathways. Each
topic term is assigned a state number to allow further numerical processing; the topics
are interpreted as discrete states of the state space. We treat each narrative pathway
(i.e., interview) as walking over a state space that is spanned by the topic terms. The
transition probabilities between discrete states are stored in a transition
(probability) matrix. This is an n-by-n square matrix with n corresponding to the number
of discrete states. The ij-th element of our transition matrix
describes the probability that an interview discusses topic i
after having addressed topic j. In molecular kinetics, this is
called a conditional probability [Prinz et al. 2011]. In the context of our research,
the transition (probability) matrix models the narrative system consisting of discrete
states that are the topic terms; it also represents the way interviews are likely to
move from one topic to another topic. We also use the transition matrix as an adjacency
matrix; from this we construct the directed graph representing the narrative network of
topics.
The transition matrix is the central object of the MSM. However, as the transition
matrix describes the probability of hearing topic i in the
future, conditioned on topic j in the present, it does not
tell the (unconditional) probability of topic i irrespectively of
the previous topic. We can however estimate the unconditional probability, also referred
to as stationary and equilibrium state probability earlier, by noting that it is the
stable state that the system converges to in the limit of infinite time. It is often
denoted by π and follows from applying the dynamics n → ∞ times, i.e. π = lim n → ∞ p0.
It can be computed either by calculating the limit explicitly, or as the left
eigenvector of the transition matrix that corresponds to eigenvalue 1. We applied this
latter technique, also known as spectral decomposition of the transition matrix [Prinz et al. 2011]; specifically, we used the pyemma software package [Scherer et al. 2015].
The resulting MSM stationary distribution can be approached as an estimator of the
unconditional probability of a dynamic system in equilibrium. In the context of our
research, the unconditional probability of a topic gives the overall probability that we
will hear about this topic irrespectively of the preceding topic.
4. Application of TPT for the study of narrative pathways
To shed light upon the full network of narrations, we compute the TPT flux network
between start topic A and end topic B [Scherer et al. 2015]. In brief, TPT defines the
committor q+ as the probability that on a pathway from A to B, at
a given intermediate step of that pathway, we reach B before going back to A. In the
context of this manuscript, the committor thus describes the probability that during a
specific narration (defined by its starting topic A and ending topic B), the narrator
speaking about an intermediate topic first reaches the end topic B before talking about
the start topic A again. The backward committor analogously describes the probability to
go back to A before visiting B. From these two definitions, the current (usually
referred to as the reactive flux) between two topics i and j can be computed by multiplying the backward committor of i with the stationary probability of that topic and with the transition probability to change from that topic i to another topic j, multiplied by the
committor of j.
To span the full flux network, this expression is evaluated between all topic terms for
a given narration. In order to simplify the presented networks, we choose a cutoff of
0.25 below which we do not display a given link. The graph layout is conducted using a
few-step general graph-layout algorithm (Fruchtermann-Reingold) started from points
distributed according to their committor probability on the x-axis [Fruchterman and Reingold 1991].
5. Comparative analysis of narratives by women and men
To assess different kinds of sub-samples (e.g. women or men), the initial sets of
narrative pathways are split according to biographical information known about the
interviewee. Independent MSM estimates are established for each of these data sets. We
compare unconditional probabilities per topic between different datasets by computing
the ratio between them. This is possible as the full state spaces of the compared MSMs
are spanned by the same set of topics. Furthermore, converged MSM estimates of
unconditional probabilities do not depend on sample size. The estimates are thus not
biased by different sample sizes as long as they are converged. Topics that are not in
the active set of either MSM are not used for comparison. We estimate the error of
probability estimates by bootstrapping of trajectories, i.e., we have re-evaluated the
investigated models by drawing the same number of interviews from our database at random
(with replacement). This is repeated 1000 times to estimate a posterior distribution of
the MSM stationary probabilities.
6. Validation of results
To validate our findings, on the one hand we use the implied timescales test [Wehmeyer et al. 2019] [Bowman and Noé 2014]. This tests for the convergence of the model with
the chosen lag time, which is a necessary condition for the data’s Markovianity. On the
other hand, to validate our results with a different method, we also drew on the direct
counting of how many times a given topic occurs in testimonies of men and women. We
compared the proportion of women who mention a given topic to the proportion of men who
also mention that topic. We divided the data sample into men and women on the one hand,
to those that mention and do not mention a topic on the other. For each topic this
produced a 2x2 contingency table: number of women mentioning a topic, number of women
not mentioning a topic; number of men mentioning a topic, number of men not mentioning a
topic. To assess whether the difference between men and women is significant, we applied
the Fisher exact test on the contingency tables representing each topic (p-value 0.05).
This revealed whether the difference between men and women is due to chance
(null-hypothesis) or whether there is a correlation between gender and the fact that a
topic word is mentioned. Finally, we measured the strength of correlation between gender
and a topic word with the help of odds ratio analysis [Morris and Gardner 1988].
Acknowledgements
We are thankful to the USC Shoah Foundation for giving access to the dataset; we are
particularly thankful to Sam Gustman, the Chief IT Officer, and to staff members of the
Foundation for providing consultancy about the dataset. We are also grateful to prof.
Wolf Gruener, director of the Center for Advanced Genocide Research (CAGR) at the USC
Shoah Foundation, for reviewing and commenting this paper.
Data Availability
Original interviews can be retrieved from the designated website of the Visual History
Archive of the Shoah Foundation: https://vhaonline.usc.edu/. The raw data
used for this research can be downloaded from Harvard Dataverse:
Toth, Gabor Mihaly, 2021, "Replication Data for the VHA Auschwitz Report Project",
https://doi.org/10.7910/DVN/JFH2BJ, Harvard Dataverse, V1
Code Availability
All research code (including data analysis in Python Jupyter Notebooks) to reproduce
this project is available in the following repository: https://github.com/toth12/vha_auschwitz_report.
Author Contributions
Gábor Mihály Tóth conceived of, designed, and led this project. He processed the
dataset, developed the computational model, and accomplished the data analysis. With
assistance of the co-authors he wrote this paper.
Tim Hempel contributed to research design, data processing, and computational modeling.
He revised the manuscript, in particular the Methods section, and created Figure 15.
Krishna Somandapelli contributed to the research design and edited the paper.
Shri Narayan contributed to the research design, edited, and revised this paper. He
secured funding for the project.
Notes
[1] The following account by a German Jewish survivor of
Auschwitz illustrates the sharp contrast between the two opinions: “And I should
preface this by saying, in general, you know, I've read a lot — a lot of claims about
the solidarity of prisoners under those conditions. And I must tell you I've never
seen it. I've heard about solidarity in the block occupied by Russian officers, Red
Army officers. But I'm pretty sure that was the only — if — if at all, that was the
only group that showed any solidarity whatsoever. The atmosphere in the camp was such
that everybody was out for their own survival. And if it was a question of — of
listening in to what your neighbors in the bed said, it would immediately be handed
on so that — for personal profit. You couldn't trust anybody. And nobody trusted
anybody.” [Chrambach 1996, Tape 3,00:26:51]. See a very different view in the memoir
of the Polish sociologist [Pawełczyńska 1973, 115–116].
[2] The Auschwitz complex came to
existence in the Spring of 1940 in Southern Poland (that time part of Nazi Germany)
about 70 km from Cracow. Over the subsequent years, the complex continued to grow
incorporating various sub-camps.
[3] The construction of Auschwitz-Birkenau (renamed Auschwitz II in
November, 1943) began in the Autumn of 1941 and started to receive deportees from the
Spring of 1942. In addition to an extremely large Jewish population, Birkenau also
housed Polish and Roma victims. It was both a killing center, equipped with gas
chambers and crematoria, and a work camp where victims were forced to slave labour.
Victims were transported by trains to Birkenau from various European
countries.
[4]
For more details see the website of Shoah Foundation,https://sfi.usc.edu/.
[5] Upon registration, the dataset we worked on
is accessible through the purpose built webinterface of the Visual History Archive,
see https://vhaonline.usc.edu.
[6] For instance,
the mass deportations of the Hungarian Jewry began in the Spring of 1944. Hungarian
Jewish men had been taken to forced labour from 1941. The vast majority of younger
Hungarian Jewish men were not deported to Auschwitz [Braham 1977] [Michelbacher 2020].
Works Cited
Alpay and Gohberg 2006 Alpay, D., Gohberg, I. The State Space Method: Generalizations
and Applications. Birkhäuser Verlag, Basel (2006).
Bacon 2017 Bacon, E. K. Saving Lives in Auschwitz: The Prisoners Hospital in
Buna-Monowitz. Purdue University Press, West Lafayette (2017).
Bowman and Noé 2014 ] Bowman, G. R., Noé, F. “Software for Building Markov State
Models” In Advances in Experimental Medicine and Biology. Springer Netherlands,
Dordrecht (2014), pp. 139–139.
Braham 1977 Braham, R. L. “The Hungarian Labor Service System, 1939-1945.” East European
Quarterly, Boulder (Co) (1977).
Chodera and Noé 2014 Chodera, J.D., Noé, F. “Markov state models of biomolecular
conformational dynamics,” Current Opinion in Structural Biology, 25 (2014): 135–144.
Chrambach 1996 Chrambach, A. “Oral History Interview.” USC, Shoah Foundation, Visual
History Archive, 23428, Los Angeles, (1996).
Ecker 2007 Ecker, M. “Verbalizing the Holocaust: Oral / Audivisual Testimonies of
Holocaust Survivors.” In M. L. Davies and C. W. Szejnmann (eds), How the Holocaust Looks
Now: International Perspectives, Palgrave Macmillan, New York (2007), pp. 41–49.
Efron et al. 2004 Efron, B. Rogosa, D. and Tibshirani, R. “
Resampling methods of estimation.” In J. D. Wright (ed.), International Encyclopedia of
the Social & Behavioral Sciences, Elsevier, Amsterdam (2004), pp. 492 - 495.
Fisher 1922 Fisher, R. A. “On the Interpretation of χ 2 from Contingency Tables, and
the Calculation of P”, Journal of the Royal Statistical Society 85 (1922): 87-94.
Freeman 1978 Freeman, L. C. “Centrality in social networks conceptual clarification”,
Social Networks 1 (1978): 215–239.
Fruchterman and Reingold 1991 Fruchterman, T. M. J. and Reingold, E. M. 1991. “Graph
drawing by force-directed placement”, Software: Practice and Experience 21 (1991):
1129–1164.
Greenspan 2012 Greenspan, H. “Survivors’ Accounts.” In P. Hayes and J. K. Roth (eds),
The Oxford Handbook of Holocaust Studies, Oxford University Press, Oxford (2012),
10.1093/oxfordhb/9780199211869.003.0028.
Greif 1999 Greif, G. “Solidarity and Animosity among Sonderkommando Prisoners in
Auschwitz-Birkenau”, The Journal of Holocaust Research, 33:4 (1999): 239-253.
Gross 1996 Gross, A. “Oral History Interview,” USC, Shoah Foundation, Visual History
Archive, 11272, Los Angeles (1996).
Gruner 2006 Gruner, W. Jewish Forced Labor Under the Nazis: Economic Needs and Racial
Aims, 1938-1944. Cambridge University Press, Cambridge (2006).
Gustman 1998 Gustman, S. “Method and Apparatus for Cataloguing Multimedia Data,” U.S.
Patent 5,832,495190 (Nov. 3, 1998).
Hayes 2000 Hayes, P. Industry and Ideology: I. G. Farben in the Nazi Era. Cambridge
University Press, Cambridge (2000).
Husic and Pande 2018 Husic, B. E. and Pande, V. S. “Markov State Models: From an Art
to a Science”, Journal of the American Chemical Society, 140 (2018): 2386–2396.
Langer 1993 Langer, L. L. Holocaust Testimonies: The Ruins of Memory. Yale University
Press, New Haven (1993).
Lee et al. 2013 Lee, C., Ozdaglar, A. and Shah. D. “Computing the Stationary
Distribution Locally.” In Advances in Neural Information Processing Systems (2013):
1376-1384.
Levi 1986 Levi, P. I sommersi e i salvati. Einaudi, Torino (1986).
Metzner et al. 2009 Metzner, P., Schütte, C., Vanden-Eijnden, E. “Transition Path
Theory for Markov Jump Processes”, Multiscale Modeling & Simulation, 7 (2009):
1192–1219.
Michelbacher 2020 Michelbacher, D. Jewish Forced Labor in Romania, 1940–1944. Indiana
University Press, Bloomington (2020).
Morris and Gardner 1988 Morris, J. A. and Gardner, M. J. “Statistics in Medicine:
Calculating confidence intervals for relative risks (odds ratios) and standardised
ratios and rates” BMJ, 296 (1988): 1313–1316.
Neumann 1996 Neumann, E. “Oral History Interview,” USC, Shoah Foundation, Visual History
Archive, 15938, Los Angeles, (1996).
Page et al. 1999 Page, L., Brin, S. and Winograd, T. “The PageRank Citation Ranking:
Bringing Order to the Web”, Stanford InfoLab Publication Server (1999):
http://ilpubs.stanford.edu:8090/422/
Pawełczyńska 1973 Pawełczyńska, A. Wartości a przemoc: Zarys socjologicznej
problematyki Oświe̜cimia. PWN, Warschau (1973).
Pingel 2010 Pingel, F. “Social life in an unsocial environment: the inmates’ struggle
for survival.” In Wachsmann N. and Caplan, J. (eds) Concentration Camps in Nazi Germany:
The New Histories, Routledge, London (2010): pp. 58–82.
Piper 2009 Piper, F. “Auschwitz.” In Megargee, G. P. (ed.) The United States Holocaust
Memorial Museum Encyclopedia of Camps and Ghettos, 1933-1945. Indiana UP in Association
with the United States Holocaust Memorial Museum, Bloomington (2009): pp. 203–214.
Prinz et al. 2011 Prinz, J. H., Wu, H., Sarich, M., Keller, B., Senne, M., Held, M.,
Chodera, J.D., Schütte, C., Noé, F. “Markov models of molecular kinetics: Generation and
validation,” The Journal of Chemical Physics 134, 174105 (2011):
https://doi.org/10.1063/1.3565032.
Scherer et al. 2015 Scherer, M. K., Trendelkamp-Schroer, B., Paul, F.,
Pérez-Hernández, G., Hoffmann, M., Plattner, N., Wehmeyer, C., Prinz, J.-H., Noé, F.,
“PyEMMA 2: A Software Package for Estimation, Validation, and Analysis of Markov
Models,” Journal of Chemical Theory and Computation, 11 (2015): 5525–5542.
Schwarz 1998 Schwarz, G. “Frauen in Konzentrationslager – Täterinnen und
Zuschauerinnen.” In Herbert, U., Orth, K., Dieckmann, C. (eds) Die
Nationalsozialistischen Konzentrationslager: Entwicklung Und Struktur, Wallstein Verlag,
Göttingen (1998): pp. 800–821.
Sofsky 2013 Sofsky, W. The Order of Terror: The Concentration Camp. Princeton
University Press, Princeton (2013).
Swiebocki 2000 Swiebocki, H. Auschwitz: 1940-1945 ; central issues in the history of the camp Volume IV. Auschwitz-Birkenau State Museum, Oswiecim (2000).
Toth 2021 Toth. G. M. In Search of the Drowned: Testimonies and Testimonial Fragments
of the Holocaust. Yale Fortunoff Archive, New Haven (2021),
lts.fortunoff.library.yale.edu, lGabor Toth2022-03-04T12:51:22, last accessed, 1 March, 2022.
USC Shoah Foundation USC Shoah Foundation, https://sfi.usc.edu/, last accessed, 1 March, 2022.
Vanden-Eijnden 2006 Vanden-Eijnden, E. “Transition Path Theory.” In Ferrario, M.,
Ciccotti, G., Binder K. (eds), Computer Simulations in Condensed Matter Systems: From
Materials to Chemical Biology Volume 1. Lecture Notes in Physics. Springer, Berlin
(2006): pp. 453 - 493.
Visual History Archive Visual History Archive, USC Shoah Foundation,
https://vhaonline.usc.edu/about/archive, last accessed, 1 March,
2022.
Wachsmann 2015 Wachsmann, N. KL: A History of the Nazi Concentration Camps. Macmillan, New York, (2015).
Wehmeyer et al. 2019 Wehmeyer, C., Scherer, M. K., Hempel, T., Husic, B. E., Olsson,
S., Noé, F. “Introduction to Markov state modeling with the PyEMMA software”, Living
Journal of Computational Molecular Science 1 (2019): https://doi.org/10.33011/livecoms.1.1.5965
URL: http://www.digitalhumanities.org/dhq/vol/16/3/000622/000622.html
Comments: dhqinfo@digitalhumanities.org
Published by: The Alliance of Digital Humanities Organizations and The Association for Computers and the Humanities
Affiliated with: Digital Scholarship in the Humanities
DHQ has been made possible in part by the National Endowment for the Humanities.
Copyright © 2005 -
Unless otherwise noted, the DHQ web site and all DHQ published content are published under a Creative Commons Attribution-NoDerivatives 4.0 International License. Individual articles may carry a more permissive license, as described in the footer for the individual article, and in the article’s metadata.
Comments: dhqinfo@digitalhumanities.org
Published by: The Alliance of Digital Humanities Organizations and The Association for Computers and the Humanities
Affiliated with: Digital Scholarship in the Humanities
DHQ has been made possible in part by the National Endowment for the Humanities.
Copyright © 2005 -
Unless otherwise noted, the DHQ web site and all DHQ published content are published under a Creative Commons Attribution-NoDerivatives 4.0 International License. Individual articles may carry a more permissive license, as described in the footer for the individual article, and in the article’s metadata.
