Topic modeling reveals how Latvian diaspora newspapers preserved cultural identity from the 1940s to 2000s. Our computational analysis of over 7,000 articles uncovers song festivals, Cold War geographies, and unexpected methodological lessons for digital humanities.
Topic Modeling zeigt, wie lettische Diaspora-Zeitungen von den 1940er bis 2000er Jahren kulturelle Identität bewahrten. Unsere computergestützte Analyse von über 7.000 Artikeln enthüllt Liederfeste, Geografien des Kalten Krieges und unerwartete methodische Erkenntnisse für die Digital Humanities.
How to read thousands of newspaper articles?
This was the question we faced when the National Library of Latvia granted us access to their digitized collection of Latvian exile press – 27 newspapers and journals, over 17,000 issues, and more than 780,000 articles published by diaspora communities from Australia to Canada between the 1940s and 2000s (this analysis represents a pilot study focused on English-language materials – roughly 7,000 relevant articles categorized via the Python langdetect library with a confidence level exceeding 70%). Traditional close reading would take ages. Instead, we turned to topic modeling – and specifically BERTopic – to discover what these communities wrote about and how they preserved their identity far from home.

Čikāgas Ziņas, Issue No. 25, May 1978
What Is Topic Modeling?
Topic modeling is a computational method that identifies recurring themes across large text collections. Imagine sorting thousands of documents into thematic folders. To achieve this, instead of you reading each one, an algorithm detects patterns of words that frequently appear together. A cluster containing election, vote, parliament, and party likely represents political coverage; one with church, pastor, congregation, and prayer points to religious life.
Traditional approaches like LDA (Latent Dirichlet Allocation) have been used in digital humanities for over a decade, but often produce topics that are difficult to interpret. BERTopic, a newer technique, leverages transformer-based language models – the same technology behind ChatGPT – to better capture semantic meaning (for an accessible overview of available tools, see Egger & Yu, 2022). It understands that song festival is a single concept, not just two unrelated words. This makes it particularly suited for analysing historical newspapers, where cultural terms and multi-word expressions carry significant meaning.
The results surprised us – not because the algorithm found something entirely unexpected, but because it independently confirmed what historians have known about Baltic diaspora culture, while also revealing patterns we hadn’t anticipated.

Most popular topics (fragment of BERTopic visualization for 1-2-3-gram approach): Culture (Topic 1), Labor Camps (Topic 2), Science and Education (Topic 3), etc. Topic 0 is the largest cluster by document count, aggregating heterogeneous texts that do not fit into specialized clusters. Duplicate singular/plural forms in some topics (e.g., “law”/”citizenship law”) reflect intentionally minimal lemmatization, applied to pre-serve culturally specific vocabulary.
When the Algorithm Rediscovers Song Festivals
The largest thematic cluster in our English-language corpus – more than 200 out of 7,000 articles – centers on music, festivals, songs, opera, concerts, folk traditions, choirs, and dance. For anyone familiar with Baltic culture, this is no surprise. The Latvian Song and Dance Festival, a tradition dating back to 1873, has been a cornerstone of national identity. UNESCO recognized it as a Masterpiece of Oral and Intangible Heritage. But seeing the algorithm arrive at this conclusion independently, purely through statistical patterns in word co-occurrence, was a moment of validation.

Latvia. Riga. VII Latvian Song Festival. 1931. Photographer: unknown. Source: Wikimedia Commons.
When we refined our analysis from single words (unigrams) to three-word phrases (trigrams), the picture sharpened dramatically. Generic terms like “music” and “festival” became “Latvian song festival,” “Latvian community center,” and “Chicago Latvian community.” The exile press wasn’t just discussing music in the abstract – it was documenting specific community gatherings, venues, and cultural institutions that kept traditions alive thousands of miles from Riga.
Other major themes emerged with similar clarity each in its own temporal horizon within the corpus: forced labor camps and Soviet deportations, universities and academic life in exile, NATO membership of host countries and geopolitical advocacy, and church activities. These clusters tell a story of how displaced persons and communities maintained not just cultural practices, but also historical memory, educational institutions, and political engagement.
The Geography of Diaspora Memory
Beyond thematic analysis, we extracted nearly 40,000 geographic place mentions from the corpus. The spatial distribution of exile communities, it turns out, organized itself along three dimensions.
First, the lost homeland: Riga dominates with over 4,000 mentions, later followed by other Latvian cities and regions. The homeland remained central to diaspora consciousness even decades after displacement.
Second, the political adversary: Moscow appears over 2,460 times – more than any city except Riga. The Soviet capital was a constant reference point, whether in discussions of occupation, Cold War politics, or eventual independence movements.
Third, the new homes: the United States leads with mentions of Chicago, Washington, New York, and other cities totaling over 3,000 references. Australia follows (Sydney, Melbourne), then Canada (Toronto, Montreal, Ottawa). These weren’t just places where Latvians happened to live – they were sites of community building, cultural preservation, and political organizing.
What surprised us were the unexpected geographic references. Korea (both) with 621 mentions, Hungary 276, Afghanistan 241 – hot spots of Cold War confrontation. The exile press, it seems, was not merely nostalgic. It engaged actively with global politics, tracking conflicts that resonated with both their experience of Soviet occupation and the Cold War concerns of their host countries.
The Paradox of Perfect Scores
Such metrical results always carry the grain of aggregated errors. This is why digital historians control their results with various metrics. Here we must be honest about a methodological discovery that gave us pause. When we evaluated our topic models using standard coherence metrics – statistical measures of how well topic words hang together semantically, essentially asking „do these words make sense as a group?” – the highest-scoring results came from our bigram model. Good news, we thought initially.
But when we examined what these “coherent” topics actually contained, we found telephone directories, professional listings, and contact information. The algorithm had identified patterns that were indeed highly structured and internally consistent – but not the kind of information we were looking for – it identified a category of text or genre rather than for instance meaning-making.
This is a crucial lesson for computational humanities: statistical validity does not equal historical relevance. A topic capturing “Grand Rapids Mich,” “Silver Spring MD,” and “Los Angeles CA” might score perfectly on coherence metrics because these phrases always appear in the same structured format. These directory-style topics – assuming they represent address and subscription tables – reveal a lot about which exile locations were on the radar of the diaspora community. They could be used to map the geography of dispersal or reconstruct community networks. The lesson is that topic modeling is a serendipitous technology. It holds a lot of answers, if you know how to ask the right questions in relation to the algorithmic technology and the corpus.
The solution to tracing the meaning-making in the communities was combining multiple approaches. Our unigram model provided broad thematic categories with strong coherence. The trigram model captured specific cultural phrases despite lower overall scores. Together, they offered complementary perspectives that neither could provide alone (Burckhardt et al. 2019).
What Comes Next
This analysis represents a pilot study focused on English-language materials – roughly 7,000 articles from a much larger multilingual collection. The full corpus also includes Latvian, Russian, German, and Swedish texts spanning 27 periodicals and over 780,000 articles.
Our next steps involve extending the analysis to the complete multilingual archive using a multilingual transformer model such as multilingual-e5-large, comparing how identity themes manifest across languages, – specifically, how references to the homeland shifted around key political events like the 1990-1991 restoration of independence.
More broadly, we see potential for deeper temporal analysis – how was Latvian culture represented in the 1950s compared to the late 1980s? From there, we plan a comparative analysis across other exile press collections from the former Soviet space. Did Estonian, Lithuanian or Ukrainian diaspora newspapers show similar patterns of cultural preservation? The methods we developed for the Latvian corpus could be adapted for cross-cultural comparison.
The exile press represents a remarkable archive of how displaced communities maintained identity across decades and continents. Computational methods cannot replace careful historical interpretation, and they need rigid methodological and epistemological grounding – our experience with misleadingly high coherence scores illustrates that. When applied carefully, however, they can reveal patterns across thousands of documents that no individual scholar could read in a lifetime – not merely confirming what we already suspect, but generating new hypotheses and unexpected connections that would otherwise remain invisible.
References
Egger R and Yu J (2022) A Topic Modeling Comparison Between LDA, NMF, Top2Vec, and BERTopic to Demystify Twitter Posts. In: Frontiers in Sociology. Volume 7, 2022. https://doi.org/10.3389/fsoc.2022.886498
Burckhardt, Daniel, Alexander Geyken, Achim Saupe, und Thomas Werneke. Distant Reading in der Zeitgeschichte. Möglichkeiten und Grenzen einer computergestützten Historischen Semantik am Beispiel der DDR-Presse. 2019. https://doi.org/10.14765/ZZF.DOK-1345.
Simon Donig is Head of the Research Department Digital History and Information Systems at Herder-Institut Marburg.
Dinara Gagarina and Timur Mitrofanov are research associates at the Department Digital History and Information Systems at Herder-Institut Marburg.
Title picture: European Latvian Song Festival, 1973, Cologne, Germany; from the private collection of Ilze Šakare; National Library of Latvia. Published here with kind permission of the copyright owner.




































