Source-linked AI summary
Framing and Agenda-setting in Russian News: a Computational Analysis of Intricate Political Strategies
Anjalie Field, Doron Kliger, Shuly Wintner, Jennifer Pan, Dan Jurafsky, Yulia Tsvetkov
TL;DR
The paper addresses how to identify subtle media manipulation beyond censorship and fake news. It combines agenda-setting and framing with cross-lingual methods to study Russian media, finding that U.S. reporting distracts from Russian economic downturns.
Problem
Subtle media manipulation, such as distracting audiences through topic selection or viewpoint emphasis, lacks systematic identification methods.
Method
The paper combines agenda-setting and framing concepts with economic metrics and cross-lingual projection of framing annotations to analyze Russian media.
Results
Izvestia reports negative U.S. events as a distraction from Russian economic downturns, while the framing lexicons correlate 0.75 with a related topic model.
Takeaways & Limitations
The approach provides a starting point for automatically identifying and analyzing subtle media manipulation strategies in news.
Takeaways & Limitations
Existing supervised framing methods focus on English because annotated data for other languages is unavailable, and the analysis primarily investigates U.S. coverage.
Abstract
from arXiv · showhide
Amidst growing concern over media manipulation, NLP attention has focused on overt strategies like censorship and "fake news'". Here, we draw on two concepts from the political science literature to explore subtler strategies for government media manipulation: agenda-setting (selecting what topics to cover) and framing (deciding how topics are covered). We analyze 13 years (100K articles) of the Russian newspaper Izvestia and identify a strategy of distraction: articles mention the U.S. more frequently in the month directly following an economic downturn in Russia. We introduce embedding-based methods for cross-lingually projecting English frames to Russian, and discover that these articles emphasize U.S. moral failings and threats to the U.S. Our work offers new ways to identify subtle media manipulation strategies at the intersection of agenda-setting and framing.
1 Introduction
The paper examines subtle media manipulation through agenda-setting and framing, applying economics- and political-science-based techniques to Russian news. It studies Izvestia coverage from 2003–2016 to identify how topic selection and presentation may shape public attention.
- 1 Introduction: The paper addresses subtle media manipulation strategies that are harder to detect than fake news or censorship, focusing on agenda-setting and framing.Agenda-setting concerns which topics receive coverage, while framing concerns which aspects are highlighted to promote particular interpretations.
- 1 Introduction: Its contributions include automating agenda-setting detection with economic metrics, projecting framing annotations across languages, and using those annotations to analyze manipulation.
- 1 Introduction: The study analyzes 2003–2016 coverage from Izvestia, a widely read Russian newspaper strongly influenced by the government.
- 1 Introduction: The authors investigate whether Russian economic performance relates to coverage of foreign events, primarily the United States as Russia’s main rival.
- 1 Introduction: The paper reports that economic indicators precede increases in U.S. news coverage, treating this pattern as agenda-setting, and then examines how those articles frame the U.S.
2 Agenda-Setting
The agenda-setting analysis finds that U.S. coverage in Izvestia rises as Russian economic conditions worsen. Monthly Granger-causality tests indicate that stock-market declines precede increased U.S. coverage within one month.
- 2.1 Correlations: r=-0.83: annual U.S. coverage and Russian GDP show a strong negative Pearson correlation, with U.S. mentions increasing as economic indicators deteriorate.The 2008 exception coincides with the U.S. financial crisis and presidential election, which may explain unusually high U.S. coverage despite peak Russian GDP.
- 2.1 Correlations: The analysis uses article-level and word-level U.S. coverage measures, compares them with GDP and RTSI, and evaluates yearly, quarterly, and monthly resolutions.
- 2.1 Correlations: Strong negative correlations persist across article-level and word-level coverage metrics, GDP and RTSI indicators, and yearly, quarterly, and monthly time slices.
- 2.2 Granger Causality: The Granger-causality analysis models whether past economic values improve prediction of U.S. coverage, with significance determined by the lagged economic coefficient.
- 2.2 Granger Causality: A one-month RTSI lag significantly Granger-causes both word-level and article-level U.S. coverage, with a negative coefficient indicating that market declines precede increased coverage.Two-month RTSI lags are not significant, suggesting that coverage changes follow stock-market changes within one month.
3 Framing Analysis
The paper develops an interpretable, cross-lingual framing method because existing automated approaches are corpus-specific, difficult to interpret, or limited to English. It derives and expands frame lexicons from English annotations for Russian analysis of Izvestia.
- 3.1 Background on Framing Analyses: Existing framing methods are difficult to transfer across corpora or languages because topic models are corpus-specific and supervised approaches require annotations unavailable outside English.The paper therefore seeks an interpretable method applicable across languages.
- 3.2 Framing Analysis Methodology: The method derives English frame lexicons from Media Frames Corpus annotations using pointwise mutual information and filters words by document frequency.For each frame, the 250 highest-scoring words form the base lexicon before query expansion.
- 3.2 Framing Analysis Methodology: For Russian, the method translates the base lexicons, restricts vocabulary to frequent corpus words, and expands them with embedding neighbors around each lexicon’s center.The Russian procedure uses 200-dimensional CBOW embeddings, cosine distance, and manually chosen thresholds.
- 3.2 Framing Analysis Methodology: The final lexicons contain between 100 and 300 words per frame and include terms related to Russian culture and politics.Table 4 provides examples of source MFC lexicons and their transferred Izvestia counterparts.
- 3.2 Framing Analysis Methodology: Documents receive a frame when they contain at least three lexicon words, with the most frequent frame assigned as primary.Documents lacking three words from any framing lexicon are assigned the primary frame “Other.”
4 Evaluation of Framing Lexicons
The framing lexicons are evaluated on English and Russian data, showing strong performance against supervised or topic-model baselines while revealing limitations for abstract or infrequent frames.
- English Evaluations: Primary-frame classification outperforms Card et al. (2016) and is comparable to Ji and Smith (2017), while transferring across domains and languages without new annotations.The evaluation uses 10-fold cross-validation on the Immigration subset of the Media Frames Corpus.
- English Evaluations: The method outperforms a fully supervised bag-of-words baseline for identifying all frames except two frames, with higher F1 for Morality when that frame is more frequent.The exception frames are External Regulation and Reputation and Morality; on the Samesex subset, Morality reaches F1 0.65.
- Russian Evaluations: All reported accuracies exceed random guessing, with no soft accuracy below 60%; only Fairness and Equality and Morality average 60% or less.These abstract frames also show larger gaps between hard and soft accuracy, reflecting annotator disagreement.
- Qualitative Comparison to Structured Topic Models: Framing lexicons better capture reader-conceptualized frames, whereas structured topic models better identify fine-grained, corpus-specific topics.The comparison includes examples where lexicons connect vocabulary to higher-level frames and topic models identify concrete subjects such as sports.
- Qualitative Comparison to Structured Topic Models: The lexicons correlate with related topic-model outputs, including a 0.75 correlation for Legality, Constitutionality, and Jurisdiction vocabulary and a legislation topic.Lexicon assignments also have higher precision because they rely directly on words occurring in documents.
5 Identifying Media Manipulation
The paper combines nPMI-based frame analysis with downturn-sensitive lexical comparisons to identify how U.S.-focused coverage distracts from negative Russian economic events. It finds recurring portrayals of U.S. immorality, insecurity, and military weakness.
- 5.1 Salient frames: nPMI measures which frames are overrepresented in U.S.-focused news, with strongest associations including External Regulation and Reputation, Morality, Political, Public Sentiment, and Security and Defense.nPMI ranges from -1 to 1; 1 denotes complete co-occurrence and 0 complete independence.
- 5.1 Salient frames: The analysis treats articles as U.S.-focused when they mention the U.S. at least twice and as employing a frame when they use at least three words from its lexicon.These operational thresholds support the frame-association analysis.
- 5 Identifying Media Manipulation: A correlation of -0.49 links Security and Defense AgendaLex frequency in U.S.-focused articles to the RTSI, with words becoming more common after downturns and less common after upturns.AgendaLex identifies words whose salience changes across months following stock-market declines versus growth.
- 5 Identifying Media Manipulation: As the stock market declines, U.S.-focused articles increasingly emphasize threats to the U.S. and use framing strategies that distract from negative economic events.The identified strategies include villainizing the U.S., describing threats to it, and promoting the Russian military over the U.S. military.
- 5.3 Examples of framing during downturns: Articles villainize the U.S. by describing its military actions as immoral and portraying U.S. troops or campaigns negatively, making Russian actions appear favorable by comparison.Examples invoke “Nazi,” “barbaric,” and “horror and outrage,” alongside Morality and Legality, Constitutionality, Jurisdiction frames.
- 5.3 Examples of framing during downturns: Other articles portray the U.S. as unsafe or militarily inferior, while presenting Russia as safer or stronger by comparison.Examples describe threats to the U.S., downplay U.S. military threats to Russia, or emphasize allegedly superior Russian technology.
6 Related Work
The paper extends research on Russian media manipulation and automated framing by combining semi-supervised lexicon construction with economic indicators. Unlike prior work that uses news to predict economic changes, it uses economic data to examine subsequent changes in news coverage.
- 6 Related Work: Prior studies identify confusion and selection attribution in Russian state media, while related work finds analogous information-flooding strategies in China and Venezuela.Selection attribution assigns negative economic events to foreign entities and positive events to Russian officials.
- 6 Related Work: The framing method uses statistical metrics and word embeddings to generate corpus-specific lexicons, avoiding the training-data requirements of fully supervised methods and interpretability challenges of topic models.The approach is semi-supervised and designed for languages lacking substantial annotated resources.
- 6 Related Work: The study reverses the usual text-to-economy direction by using economic indicators to show changes in news articles.It applies Granger causality to connect economic time series with subsequent news patterns.
7 Conclusions
The paper shows that NLP can study subtle media manipulation by combining agenda-setting, framing, economic indicators, and cross-lingual framing projection. In Izvestia, increased reporting on negative U.S. events is presented as distraction from Russian economic downturns.
- 7 Conclusions: The study uses agenda-setting and framing theories with cross-lingual framing projection to analyze subtle distraction strategies in Russian media.It combines economic indicators with frame analysis to examine how topic selection and topic presentation work together.
- 7 Conclusions: Izvestia reports negative events in the U.S. as a way of distracting attention from economic downturns in Russia.The conclusion identifies distraction as the paper’s focal manipulation strategy.
- 7 Conclusions: The approach is positioned as a starting point for detecting nuanced framing, projections of power, over- and under-represented events, and bias in news reports.These directions are offered for further research on automating media-manipulation analysis.
Appendix A
Appendix A describes the Izvestia corpus and its preprocessing pipeline. The dataset contains 118,532 articles and uses named-entity grouping, tokenization, and lowercasing.
- Appendix A: The corpus contains 1,013,024 types, 87,761,626 tokens, and 118,532 articles, averaging 718 articles per month.These statistics summarize the Izvestia corpus used in the study.
- Appendix A: Preprocessing identifies named entities, collapses variants such as “U.S.A.” and “Americans” into one label, then tokenizes and lowercases the text.Country mentions outside the U.S. and Russia receive less extensive normalization.
Appendix B
Appendix B reports soft and hard accuracy scores alongside average precision for Russian framing lexicon intruder detection.
- The appendix evaluates a Russian framing lexicon intruder detection task.
- Performance is reported using soft accuracy scores expressed as percentages.
- The evaluation also includes hard accuracy and average precision.