💻 computer science

The Shape of Testimony: A Scalable Framework for Oral History Archive Comparison

This study employs a scalable computational framework to analyze over 1,600 Holocaust survivor testimonies, confirming general structural differences between the USC Shoah Foundation and Yale Fortunoff archives while revealing significant narrative overlaps that challenge the simplistic "structured versus free-form" dichotomy.

Original authors: Itamar Trainin, Renana Keydar, Amit Pinchevski

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Itamar Trainin, Renana Keydar, Amit Pinchevski

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine two massive libraries, each filled with thousands of books written by survivors of the Holocaust. For decades, historians have argued that these two libraries are built in completely different ways.

One library, Yale, is said to be like a quiet, open park where survivors are invited to sit and tell their stories however they feel like it, with interviewers acting mostly as silent, supportive listeners. The other library, USC, is said to be more like a guided tour through a museum, where interviewers follow a strict map, asking specific questions to ensure every part of the story is covered in a specific order.

This paper is like a team of digital detectives who decided to stop guessing and start measuring. They used powerful computer tools (specifically Large Language Models, or "super-smart AI") to read and analyze over 1,600 of these stories at once. Their goal was to see if the "park" and the "museum" really are as different as everyone says, or if the stories inside are more similar than we thought.

Here is what they found, using simple analogies:

1. The "Map" vs. The "Meandering Walk"

The researchers looked at the topics (the subjects) the survivors talked about.

The USC "Museum" (Structured): They found that USC interviews really do follow a strict map. The stories usually start with childhood, move to the war, then the camps, and finally life after. It's like walking down a hallway where the doors open in a predictable order. The interviewers often ask, "What happened next?" or "When did this occur?"
The Yale "Park" (Free-form): The Yale interviews were more like a meandering walk through a garden. Survivors would jump from a memory of their mother to a memory of the war and back again. The topics overlapped more, and the story felt more fluid and emotional.

The Twist: Even though the "paths" were different, both groups of survivors ended up telling the same general story. They all started with life before the war and ended with life after. The "map" and the "walk" eventually led to the same destination.

2. Who is Driving the Car? (Questioning Styles)

The team counted how many questions were asked and what kind they were.

USC Drivers: The USC interviewers asked more questions, and they were often short, factual questions like "Who?" "When?" or "What?" This kept the conversation moving in a straight line.
Yale Drivers: The Yale interviewers asked fewer questions, and they were often open-ended, like "How did that feel?" or "Tell me more." This allowed the survivors to drive the car, taking the story wherever they wanted to go.

3. The "Silence" and the "Talking"

They measured how long survivors could speak without being interrupted.

Yale: Survivors here got to speak in long, uninterrupted streams. It was like a solo performance where the musician plays a long, emotional song without stopping.
USC: The conversations were more like a tennis match, with the ball (the topic) being hit back and forth more frequently. The interviewers interrupted more often to guide the story.

4. The Big Surprise: They Meet in the Middle

The most interesting finding was that the difference between the two styles wasn't constant.

At the beginning: The two styles were very different. USC was very structured; Yale was very free.
At the end: As the interviews went on, they started to look more alike. In the final parts of the stories, both groups of interviewers and survivors settled into a similar rhythm. It seems that no matter how you start the conversation, when you are talking about deep, painful, and profound life experiences, the human connection naturally creates a similar flow.

The "Toolbox" They Built

Beyond just comparing these two libraries, the authors built a new digital toolbox.
Imagine trying to compare two huge piles of sand by hand; it would take a lifetime. This paper shows how to use a machine to sift through thousands of stories in a few hours, looking for patterns in how people tell their stories. They proved that you can use AI to study these sensitive, emotional stories without losing the human heart of the narrative.

In Summary

The paper confirms that the two archives do have different styles (one is more like a guided tour, the other like a free conversation), but it also shows that the human spirit is resilient. Whether guided by a strict map or a free-flowing path, the survivors' stories converge on the same powerful truths. The authors didn't just prove a theory; they built a scalable way to compare oral histories in the future, showing that computers can help us understand human stories better, not replace them.

Technical Summary: The Shape of Testimony: A Scalable Framework for Oral History Archive Comparison

Problem Statement
Scholarship in Holocaust studies has long posited a dichotomy between the two primary archives of survivor testimony: the Yale Fortunoff Video Archive and the USC Shoah Foundation. The prevailing claim suggests that Yale interviews follow a loose, open-ended, and survivor-led style, whereas USC interviews adhere to a strict, structured, and interviewer-guided format. While this distinction has influenced research and archive development, it has historically been supported only by institutional policy analysis, close readings of selected texts, or small-scale empirical studies. There is a lack of large-scale, empirical, computational evidence to validate or refute this binary distinction across thousands of narratives. Furthermore, applying computational tools to such ethically sensitive and interpretively complex corpora remains a significant challenge.

Methodology
The authors developed a scalable, reproducible computational pipeline to analyze 1,668 Holocaust survivor testimonies (1,000 from USC and 668 from Yale). The methodology integrates discourse segmentation, topic modeling, and Large Language Model (LLM) classification to quantify the "structuredness" of interviews.

Data Preprocessing: All testimonies, originally transcribed manually with speaker identification, were segmented into Question-Answer (Q/A) pairs.
Dual-Level Segmentation:
- Micro-level: Q/A pairs were treated as discrete topical units, with short pairs merged to maintain coherence.
- Macro-level: Each testimony was divided into $k=15$ equal-length chronological segments based on cumulative word count, normalizing for the significant length disparity between archives (USC averages ~23k words; Yale ~13k words).
LLM-Based Topic Extraction: Instead of traditional unsupervised methods (e.g., LDA, BERTopic) requiring manual labeling, the authors employed a two-stage LLM pipeline:
- Stage 1: An LLM generated concise descriptive labels for every Q/A pair.
- Stage 2: The LLM identified the top- $K$ recurring topics for each of the 15 chronological segments.
- Validation: The approach was validated against manual samples and prior semi-supervised work, showing strong structural convergence.
Metric Extraction: The pipeline extracted and analyzed:
- Topical Coherence: Tracking the sequence, persistence, and diversity of themes across segments.
- Dialogic Dynamics: Measuring answer length, question length, and standard deviation (variability) over time.
- Intervention Density: Calculating the proportion of uninterrupted survivor speech relative to total length to gauge narrative autonomy.
- Question Classification: Using an LLM to categorize interviewer questions into seven types ("why," "what," "where," "who," "when," "how," "other") to assess interviewing style.

Key Results
The computational analysis largely corroborates the traditional distinction while revealing significant nuances and overlaps:

Structural Differences: USC interviews exhibit a more predictable, segmented topical progression with clearer boundaries between phases (e.g., pre-war, ghetto, camp). They rely heavily on factual, directive questions ("what," "when," "who") in early segments and feature a recurring "Family Memories" segment (discussing photographs) largely absent in Yale.
Fluidity in Yale: Yale testimonies display greater thematic fluidity, with topics emerging earlier, overlapping more frequently, and persisting across segments. They utilize a higher proportion of open-ended "Other" questions and feature longer, more variable survivor responses in the early stages, indicating a more survivor-centered approach.
Convergence: Despite initial differences, the two archives converge in later stages. Both show a reduction in intervention density and answer length toward the end, suggesting shared testimonial conventions and closing rituals. The macro-historical arc (pre-war $\to$ persecution $\to$ war $\to$ post-war) remains consistent across both.
Question Dynamics: USC interviewers ask longer, more frequent questions early on, reflecting a guided scaffolding that gradually yields to survivor-led narration. Yale interviewers ask shorter, less predictable questions, reflecting responsive adaptation to the survivor's storytelling.

Key Contributions

Data-Driven Reassessment: The study provides empirical, large-scale evidence that challenges the rigid "structured vs. free-form" dichotomy. It demonstrates that "structuredness" is not a static institutional constant but a dynamic property emerging from the interplay of protocol, interviewer choices, and survivor agency.
Scalable Framework: The paper introduces a reproducible pipeline for comparative corpus analysis in digital humanities. It specifically addresses the challenges of oral history (contextual overflow, diffuse topics) by combining LLM-based topic extraction with a map-reduce logic (local generation followed by global synthesis).
Methodological Ontology: The work operationalizes interpretive categories (e.g., structuredness, topical coherence) into measurable features, balancing automation with the preservation of discursive and emotional complexity.

Significance and Claims
The authors claim this work offers a dual contribution: a reassessment of a foundational historiographical claim in Holocaust studies and the establishment of a framework for scalable oral history comparison. They argue that their approach demonstrates how complex, ethically sensitive narratives can be analyzed computationally without compromising their integrity.

The paper positions itself as a "proof of concept" for broader applications in digital oral history, narrative analysis, and the design of citizen-science annotation platforms. It asserts that computational methods, when used critically, can deepen interpretive work by connecting archives and traditions of memory work, moving the field from isolated case studies toward comparative infrastructures. The authors remain modest, noting that the pipeline is designed to be extensible and that future work will focus on integrating these tools into open-access platforms for cross-disciplinary use.