Bag-of-words representations of key entities

We use BoW representations to capture specific entities related to various aspects of the interviews. These are listed in the metadata curated by archivists, focusing on landmarks, families, significant figures, and keywords. The keywords are curated to form a thematic representation of the interview (Sleiman and Chebaro Reference Sleiman and Chebaro2018). Additionally, we extract general named entities from the transcripts using an Arabic NER model from the CAMeL Tools library (Obeid et al. Reference Obeid, Zalmout, Khalifa, Taji, Oudah, Alhafni, Inoue, Eryani, Erdmann and Habash2020) (details in the “Methods Appendix” section of the Supplementary Material). These entities capture mentions of places or landmarks without proper names, contributing to a nuanced understanding of similarity. All BoW vectors are weighted by tf-idf, so that distinctive terms and entities contribute more to the similarity than common ones. Comparisons using BoW representations emphasize narrative similarity based on shared entities like families or significant figures. In contrast, keyword-based similarity reflects shared broad themes between interviews.

Semantic embeddings from transcripts

We measure narrative similarity through semantic embeddings of thematic summaries derived from interview transcripts. This multi-step approach addresses the challenge of comparing lengthy and diverse interviews by first distilling excerpts into focused thematic summaries and then generating embeddings from these summaries. This technique draws inspiration from the use of LLMs to decompose text for social scientific purposes in other contexts, particularly as described in (Hoyle et al. Reference Hoyle, Sarkar, Goel, Resnik, Bouamor, Pino and Bali2023; Benara et al. Reference Benara, Singh, Morris, Antonello, Stoica, Huth and Gao2024). The procedure involves three main steps: (1) thematic classification of transcript excerpts, (2) summary generation focused on interviewee experiences, and (3) embedding generation and validation. Figure 1 gives a visual representation of the process.

First, we partitioned transcripts thematically using archivist-curated TOCs, which provide descriptive headers for interview sections. We categorized TOC headers into consistent themes to enable cross-interview comparison of relevant excerpts (detailed in the “Methods Appendix” section of the Supplementary Material). For the remainder of this article, we focused on a subset of themes that were more prevalent across interviews and closely related to the Nakba: Zionist Attacks During and After the Nakba; Exile, Expulsion, and Displacement; British Mandate Colonialism and Occupation; Refugee Life and Conditions; Community Life and Social Dynamics; and Resistance and Popular Struggles. Table 2 contains the list of curated themes and the associated number of interviews that contain excerpts classified as belonging to the theme, as present in the analysis sample.

For each thematic excerpt, we used gpt-4.1 (OpenAI, 2025) to generate summaries focused specifically on the interviewee’s experience within the theme. This summarization step serves multiple purposes: it focuses on relevant narrative elements, improves manageability for validation (as comparing raw excerpts proved mentally taxing for expert evaluators), and enhances interpretability. We designed the summarization prompts to comprehensively capture experiential aspects while maintaining thematic focus. We validated summary quality using a sample of 40 summaries, with 20 percent evaluated by multiple raters. These evaluations indicated good performance with high interrater reliability. The summarization made similarity comparisons more tractable than direct excerpt comparison for complex themes, such as the “British Occupation.” Details on the summarization process and validation are provided in the “Methods Appendix” section of the Supplementary Material.

After summarization, we generated embeddings from the thematic summaries using OpenAI’s text-embedding- 3-large model (Su et al., Reference Su, Shi, Kasai, Wang, Hu, Ostendorf, Yih, Smith, Zettlemoyer, Tao, Rogers, Boyd-Graber and Okazaki2023; Vector embeddings—OpenAI API, 2025). We tested various embedding processing techniques, including normalization and all-but-the-top component removal (Mu, Bhat, and Viswanath Reference Mu and Viswanath2018; Su et al. Reference Su, Cao, Liu and Yangyiwen2021), but ultimately used the unprocessed embeddings. We hypothesize that component removal might be too aggressive, as our summaries were already thematically focused (validation results are presented in the “Methods Appendix” section of the Supplementary Material). We validated the embedding similarity measures using triplet comparisons, which are more efficient and reliable for complex data than direct similarity assessments (Hermans, Beyer, and Leibe Reference Hermans, Beyer and Leibe2017; Tamuz et al. Reference Tamuz, Liu, Belongie, Shamir and Kalai2011; Thurstone Reference Thurstone1927). Our triplet sampling ensured diversity across location, gender, and difficulty pairings. Figure 2 visualizes the resulting embeddings using UMAP (McInnes, Healy, and Melville Reference McInnes, Healy and Melville2020), showing clear separation between distinct themes.

Table 2.

Selected themes and count of interviews with at least one section classified under the theme

Table 2 Long description

Two columns - Theme and No. Interviews - with eight data rows ordered from most to least frequent. Zionist Attacks During and After the Nakba, 435. Exile, Expulsion, and Displacement, 428. Community Life and Social Dynamics, 371. Resistance and Popular Struggles, 296. Socio-Economic Life and Conditions, 241. Childhood and Family Life, 201. British Mandate Colonialism and Occupation, 192. Refugee life and conditions, 174.

Note: The themes were selected for their higher prevalence in the data and their relevance to the Nakba analysis.

Figure 2.

UMAP projection of embeddings derived from thematic summaries. Related themes appear visually indistinguishable, while distinct themes are clearly separated.

Figure 2 Long description

A dense scatter of roughly two thousand points. The horizontal axis runs from about 1 to 9 and the vertical axis from about 0 to 5; UMAP coordinates have no units and are not individually interpretable. A legend lists ten colour-coded themes: Cultural Life and Folklore; Socio-Economic Life and Conditions; Community Life and Social Dynamics; Childhood and Family Life; Zionist Attacks During and After the Nakba; Agriculture and Land; Resistance and Popular Struggles; Refugee life and conditions; British Mandate Colonialism and Occupation; Exile, Expulsion, and Displacement. Distinct themes occupy separate regions: Exile, Expulsion and Displacement at the upper right; Zionist Attacks at the lower centre-right; Resistance and Popular Struggles at the bottom centre;

Similarity as cohesion

Given representations $\mathbf {e}^A_i, \mathbf {e}^A_j$ of interviews $i, j$ , where the $e^A_i$ could be BoW representations of metadata entities or semantic embeddings of the transcripts, we measure the cohesion of the interviews via the cosine similarity $\sigma (\mathbf {e}^A_j, \mathbf {e}^A_j)$ between the representations:

(1) $$ \begin{align} \sigma(\mathbf{e}^A_j, \mathbf{e}^A_j) = \frac{\mathbf{e}^A_j \cdot \mathbf{e}^A_j}{ \left\lVert {\mathbf{e}^A_j} \right\rVert \left\lVert {\mathbf{e}^A_j} \right\rVert }. \end{align} $$

We use cosine similarity for both kinds of representation. The BoW vectors, weighted by tf-idf, capture the terms and entities that two interviews share, and cosine similarity measures this overlap in a way that is insensitive to interview length. For the dense summary embeddings, cosine similarity is the standard measure, and because these embeddings are normalized, it ranks pairs identically to Euclidean distance. Although cosine similarity can be problematic in some settings (Zhou et al. Reference Zhou, Ethayarajh, Card, Jurafsky, Muresan, Nakov and Villavicencio2022), we validate its use in the summary setting by examining low- and high-similarity pairs. This measure is used for assessing all vectorized representations in this article; however, we rely on the distinct meanings of each feature to develop a full picture of narrative cohesion.

Modeling framework for dyadic data

Pairwise similarity data present statistical challenges because each interview appears in multiple pairs, violating the independence assumption of standard regression models. Our modeling framework addresses this dependency by including fixed effects for each interview, effectively comparing how the same interview varies in similarity across different partners.

The general model specification for theme T is

(2) $$ \begin{align} z(\sigma^T_{ij}) = \beta X_{ij} + \mu_i + \eta_j + \epsilon_{ij}, \end{align} $$

where $\sigma ^T_{ij}$ represents the cosine similarity between interviews i and j for theme T, and $z(\sigma ^T_{ij})$ is the standardized similarity score (mean-centered and scaled by the standard deviation within each theme). This standardization ensures the comparability of coefficients across themes while maintaining the original interpretation of the similarity metric. We evaluated the appropriateness of the standardization by noting that the similarity distributions for each aspect (metadata/semantic themes) possess low kurtosis and are close to normality with few outliers. We provide additional analysis on the matter in the “Methods Appendix” section of the Supplementary Material.

This specification has several advantages for our research questions. First, the fixed effects $\mu _i$ and $\eta _j$ absorb all time-invariant characteristics of individual interviews, ensuring that our estimates of $\beta $ reflect within-interview variation across partners rather than between-interview differences. Second, the model naturally handles the dyadic structure by symmetrically accounting for both interviews in each pair.

We estimate models using the feols function from the fixest package and calculate dyadic-robust covariance matrices using the methods of Aronow, Samii, and Assenova (Reference Aronow, Samii and Assenova2017) and Fafchamps and Gubert (Reference Fafchamps and Gubert2007) to account for residual correlation among pairs sharing common interviews. This approach produces conservative standard errors appropriate for hypothesis testing in dyadic data.

We estimate separate models for each theme and also fit a pooled model that combines data across all themes to obtain overall estimates of pairing effects. The pooled model includes theme fixed effects to account for baseline differences in similarity levels across themes.

Gender pairing analysis

To test whether gender composition affects narrative similarity, we specify a categorical model:

(3) $$ \begin{align} z(\sigma^T_{ij}) = \sum_{X \in \{FF, FM, MM\}} \beta_{X} \, I[g_{ij} = X] + \mu_i + \eta_j + \epsilon_{ij}, \end{align} $$

where $I[g_{ij} = X]$ are indicator variables for gender pairings (Female–Female, Female–Male, and Male–Male). This categorical specification allows for nonlinear effects of gender composition on similarity.

The model is identified through variation in similarity scores for the same interview across partners of different genders. Note that when we estimate this model with fixed effects, one gender pairing category is absorbed into the interview fixed effects, so the regression coefficients $\beta _X$ represent differences between pairing types rather than absolute levels.

To obtain interpretable estimates of similarity levels for each gender pairing type, we use estimated marginal means (EMMs) computed via the emmeans package. We then calculate pairwise contrasts between these EMMs, using Female–Male pairs as the reference category. This approach allows us to test specific hypotheses about gender homophily (e.g., whether same-gender pairs show greater similarity than mixed-gender pairs) while controlling for multiple comparisons using the Holm adjustment. All EMMs and contrasts incorporate the dyadic-robust covariance matrix to ensure valid inference.

We apply inverse probability weighting based on the empirical distribution of gender pairings to address potential imbalance in pairing prevalence and ensure that our estimates are not unduly influenced by the most common pairing types.

Community pairing analysis

For location-based analysis, we use a categorical specification that allows for full flexibility in how different location-sharing patterns affect similarity:

(4) $$ \begin{align} z(\sigma^T_{ij}) = \sum_{L \in \{N, O, R, B\}} \beta_{L} \, I[\ell_{ij} = L] + \mu_i + \eta_j + \epsilon_{ij}, \end{align} $$

where $I[\ell _{ij} = L]$ are indicators for location pairing types: None (N), Origin only (O), Residence only (R), or Both origin and residence (B). This categorical approach is more flexible than a factorial specification with main effects and interactions, as it does not impose additivity constraints on the effects of shared origin and residence.

We employ inverse probability weighting based on the empirical distribution of pairing types to address substantial imbalance in pairing prevalence, particularly the rarity of pairs sharing both origin and residence.

As with the gender model, one location pairing category is absorbed into the interview fixed effects during estimation. We therefore compute EMMs for each location pairing type to obtain interpretable estimates and calculate contrasts using pairs that share no locations as the reference category. Our contrast analysis tests several specific hypotheses: (1) whether any shared location (origin only, residence only, or both) increases similarity compared to no shared location and (2) whether the combination of shared origin and residence produces greater similarity than would be expected from their additive effects alone. We test this synergy hypothesis using a custom contrast that compares the EMM for the “Both” category against the sum of the individual effects:

(5) $$ \begin{align} \text{Synergy} = \text{EMM}_B - \text{EMM}_O - \text{EMM}_R + \text{EMM}_N. \end{align} $$

This contrast tests whether sharing both locations produces an additional boost in similarity beyond the sum of their individual effects. A positive synergy term indicates that the effects of shared origin and residence are superadditive, while a negative synergy term indicates interference between the two factors.

Complementary analyses and robustness checks

We complement our primary fixed-effects models with two alternative approaches documented in the Appendix of the Supplementary Material. First, we estimate Mundlak-corrected random effects models (Mundlak Reference Mundlak1978; Wooldridge Reference Wooldridge2010) that partially pool information across interviews while controlling for interview-level means. Second, we implement Bayesian versions of these models using the brms package (Bürkner Reference Bürkner2017), which provides full posterior distributions for parameters and naturally handles uncertainty quantification.

These alternative specifications serve as robustness checks, providing complementary perspectives on our research questions. The consistent qualitative results across modeling approaches strengthen confidence in our findings.