A framework for the analysis of historical newsreels

The paper addresses how to study audiovisual news—specifically historical newsreels—at scale while preserving their multimodality and societal context. It argues that existing approaches either capture nuanced meaning but lack scalability (qualitative film history) or scale well but miss multimodality and context (computational methods), and thus proposes a unified, multidisciplinary framework. The research question is how to systematically combine qualitative and quantitative methods to analyze large longitudinal collections of newsreels to reveal temporal dynamics, cultural diversity, and potential societal effects. The study motivates the need for long-term, consistent datasets and integrated methods to detect continuities, short-term trends, and contextual meaning, culminating in a framework demonstrated on the Soviet "News of the Day" series (1944–1992).

The literature review synthesizes contributions and gaps across four streams and introduces digital hermeneutics:

• Qualitative film history: Offers rich multimodal interpretation, historical context, and sensitivity to intersubjective meaning-making, but typically focuses on limited periods and is difficult to scale or compare across studies.
• Computational video analysis: Enables large-scale, systematic, and comparable analysis (shot/scene detection, person/object/event recognition, audio analysis, multimodal pipelines). However, it often privileges single modalities, offers limited historical contextualization, and rarely connects recognized features to audience meaning-making.
• Computational film studies: Bridges qualitative and quantitative approaches, analyzing dynamics (e.g., shot length, editing) and combining modalities (image, sound, text). It debates how films create meaning and which cues/units to extract. Methodological concerns include when to apply human interpretation and ensuring that variables map to meaning-making units.
• New Cinema History: Emphasizes production, circulation, and reception contexts over time and space, but engages less with film content. Integration with content analysis is needed for a fuller picture of meaning.
• Digital hermeneutics: Stresses source and method criticism when working with historical digitized materials. It highlights layered data provenance (production, archiving, digitization), temporal shifts in meaning, biases introduced during digitization/metadata creation, and the need to adapt contemporary models to historical data. This perspective cautions that gaps and censorship can be meaningful and must be accounted for in interpretation.

Framework: The proposed Newsreel Framework is a modular, multidisciplinary workflow with three core stages: (1) detect and pair meaning-making units (MMUs) with available or enrichable variables; (2) apply digital data (source) and method criticism to account for layered historical and digitization contexts; (3) conduct quantitative analysis on selected variables, visualize multiple dimensions, and merge findings into contextualized qualitative conclusions. MMUs include images, voice-over narration, acoustic motifs, persons, activities, locations, and topics, alongside contextual factors like sociopolitical circumstances and media ecology. Variables derive from metadata and audiovisual content and can be enriched via manual or computational methods.

Case study data: 1747 issues of the Soviet "News of the Day" (Новости дня) series (1944–1992), digitized by Net-Film, with mp4 videos and metadata (production year, issue number, duration, brief text descriptions in Russian/English, and crew information). The dataset is incomplete (missing issues, low-quality AV, imperfect metadata).

Data enrichment and processing:

• Shot detection and visual embeddings: Performed shot boundary detection (SBD), extracted central frames per shot (avg 117 shots per issue; avg 126 frames ≈ 5 s per shot; total 205,678 frames). Computed ResNet50 (ImageNet-pretrained) embeddings of central frames; projected via t-SNE/UMAP. Used the Collection Space Navigator for interactive exploration.
• Topic/story segmentation and text processing: Split each issue into stories using metadata (12,707 stories total), correcting numbering inconsistencies manually. Applied two approaches: • Topic modeling using embedding-driven clustering (Top2Vec/BERTopic-style pipeline with fastText embeddings and grouped TF-IDF) for exploratory themes. • Zero-shot content classification using LLMs (GPT-3.5-turbo-0301, GPT-4-0301) with eight predefined topics informed by prior literature: USSR politics; sports; military; scientific/industrial progress (incl. space, aviation, construction); USSR economy and industry; USSR agriculture; natural disasters; social issues and lifestyle; plus a misc class. Accuracy on a 100-story annotated test set: 88% (GPT-3.5-turbo-0301) vs 84% (GPT-4-0301). Applied GPT-3.5 to all stories.
• Named Entity Recognition (NER) and geocoding: Extracted place mentions from textual descriptions via NER; enriched with city name derivatives (per Russian grammar, Wiktionary) and geocoded to lat/long. Categorized city mentions: (a) city/citizens; (b) organizations named after city; (c) regions named after capital (e.g., oblast) and their organizations; (d) toponyms outside the city (e.g., Warsaw Pact); (e) not a true mention.
• Crew network and gender inference: Built a directed collaboration graph from directors to crew (hiring/supervision relation). Deduplicated names via Levenshtein distance with manual checking. Inferred assumed gender of directors/crew using Russian gendered surnames; validated qualitatively. Crew data: 1251 unique people, 15,425 links; roles include directors (1740 roles by 104 persons), cinematographers (15,145 roles by 1132 persons), and others (158 roles by 45 persons). Omitted 1944–1953 from network analysis due to inconsistent data.

Source/method criticism: Accounted for layered data formation (production, censorship, archiving, digitization, metadata creation), temporal shifts in meaning (e.g., a horse in 1910 vs 1990), model-domain mismatch (modern-trained models on historical grayscale/low-res footage), and systematic absences (e.g., likely de-Stalinization affecting pre-1954 records).

Framework-level: Demonstrated a pipeline that pairs meaning-making units with variables, integrates source/method criticism, and merges quantitative dimensions into contextual qualitative conclusions.

Case study (News of the Day, 1944–1992):

• Scale and content processing: 1747 issues; 12,707 stories; 205,678 representative frames. Zero-shot topic classification achieved 88% accuracy (GPT-3.5-turbo-0301) on a 100-story test set; applied to full corpus.
• Production volume and story structure: Weekly production is stable from 1954–1986, with sparse pre-1954 data likely due to de-Stalinization and archival gaps. From 1987, annual issue counts drop by about half. Stories per issue gradually decrease across 1954–1986 (fewer, longer stories in the 1970s–1980s).
• Cinematic dynamics: Number of shots per issue decreases over time; mean shot length increases toward the late period—opposite to trends in Hollywood features where shots shorten over the 20th century. Perestroika-era issues often feature long shots of speeches.
• Topic trends and seasonality: Shares of politics, economy/industry, agriculture, and social topics remain relatively stable until mid-1980s, after which social and then political topics expand. Clear annual rhythms: social topics peak around issues 8–9 (near International Women’s Day) and 48–52 (New Year); agriculture peaks around issues 30–40 (August–September harvest).
• Visual characteristics: Storyboards reveal consistent structural patterns (title frames followed by city scenes, then activity scenes) and the introduction of color in the mid-1980s. UMAP clusters of central frames identify recurring visual types: nature; monumental gatherings; people in meetings; close-ups of people at work; industrial production; title/text frames; city views.
• Geography of coverage: Strong emphasis on Europe within and beyond the USSR; Asian USSR east of the Urals is far less covered. Outside the USSR, Warsaw Pact countries account for 36% of mentions despite ~3% of world population (1970), and neutral capitalist neighbors (e.g., Austria, Finland) 9% of mentions (<0.4% of world population). Mentions per year trend downward overall with a sharper decline for foreign cities after ~1960, indicating reduced international coverage. City-level temporal patterns align with historical events (e.g., Krasnoyarsk spike during dam construction; steady prominence of Leningrad/St. Petersburg; rise in Minsk mentions with population growth; decline in Odesa mentions).
• Crew composition and gender dynamics: Average of ~10 people per newsreel; on average ~9 cinematographers and 1 director per issue. Network analysis shows a few prominent, long-career directors supervising multiple teams. Director gender composition exhibits three periods: relative parity (1945–1959), women-dominated directors (1960–1974), and men-dominated directors (1975–1992).

The findings address the central question of how to combine content and context to understand longitudinal dynamics and meaning in audiovisual news. By pairing MMUs with variables and integrating source/method criticism, the framework enables quantification of multimodal features (e.g., shot dynamics, visual motifs, topics, geography, crew structures) and their contextual interpretation.

A key synthesis is the identification of perestroika (mid-1980s to 1991) as a period of system-wide change across all analytical dimensions: halved annual output, fewer and longer stories per issue, lengthened shots dominated by political speeches, concentration of coverage in Moscow and a narrower set of cities, shrinking crews, and increased shares of social and political topics. This period-specific signal, consistently visible across modalities and contexts, indicates profound media-ecosystem and sociopolitical shifts (economic turbulence, decentralization, television’s rise, loosening censorship). The integrated approach reveals plausible interrelations—e.g., staff reductions helping explain fewer issues/shots and geographic concentration—connections that would be less apparent if analyzing any single dimension in isolation.

More broadly, the work shows that large-scale quantitative visualizations can contextualize and test insights typically derived from qualitative scholarship, while qualitative analysis is essential to interpret quantitative patterns, mitigate model and data biases, and connect patterns to historical processes. The framework thus contributes a replicable template for corpus-scale audiovisual news analysis across times and cultures.

The study introduces a multidisciplinary framework for large-scale, longitudinal analysis of historical newsreels that systematically integrates qualitative and quantitative methods. By explicitly pairing meaning-making units with analyzable variables, incorporating digital source and method criticism, and merging multimodal analytical dimensions, the approach maintains cultural complexity while enabling comparison across time and datasets. The Soviet "News of the Day" case provides a proof of concept, highlighting stable post-1954 weekly production, evolving cinematic dynamics (fewer, longer shots), stable yet seasonally patterned content shares until mid-1980s, a strong European/Warsaw Pact focus, and distinct gender eras among directors, with perestroika emerging as a period of systemic transformation across modalities.

Future research directions include: expanding confirmatory analyses and hypothesis testing; systematically evaluating and comparing alternative computational methods and model choices; integrating distribution, exhibition, and reception data; developing techniques for cross-collection harmonization; addressing copyright and access barriers; and leveraging scalable compute infrastructure to process larger, more diverse audiovisual corpora.

• Data completeness and bias: Missing issues, low audio/video quality, and imperfect metadata; pre-1954 gaps likely due to de-Stalinization policies. Some metadata (e.g., descriptions) created during digitization may reflect later archivists’ perspectives.
• Model-domain mismatch: Off-the-shelf models (e.g., ResNet50 trained on modern ImageNet) applied to historical grayscale/low-resolution footage can introduce biases; performance may vary without adaptation.
• Network analysis scope: Crew network omits 1944–1953 due to inconsistent records, potentially biasing longitudinal inferences.
• Scope of value chain: Distribution, exhibition, and reception stages were not analyzed due to lack of data; findings focus on production and content.
• Method selection depth: The framework, while modular, did not exhaustively explore or benchmark all possible analytical methods; the case study emphasizes exploratory over confirmatory analysis.
• Computational and access constraints: Scaling requires high-performance computing; copyright and access restrictions limit comprehensive dataset coverage.