Mapping post-election Facebook discourse in Bulgaria: narrative framing, producer types, and heuristic disinformation indicators across two reach strata

This study examines Bulgarian-language Facebook posts about the parliamentary elections that were surfaced through Meta’s Content Library in the week following the 19 April 2026 election. The present analytical corpus covers two reach strata – Shard A (13,700 – 99,800 views) and Shard B (100,000 – 1,000,000 views) – and contains 769 posts with a combined 42.24 million views, collected between 20 and 27 April 2026.

The lower inclusion threshold of 13,750 views was not chosen arbitrarily. It was set to approximate the number of votes required for one seat in the National Assembly in the April 19 parliamentary election¹, so that entry into the corpus marked an audience scale roughly equivalent to one parliamentary seat quota.

The material was collected manually from the Meta Content Library search interface rather than from Meta’s downloadable research dataset. The query used the Bulgarian term “изборите”², Facebook post results, the period 20 – 27 April 2026, and image-text search enabled, as shown in the collection protocol screenshot. This design choice followed a prior comparison between two Meta Content Library outputs: the downloadable research dataset, which is restricted by follower thresholds, and the full on-screen search results, which are not limited in the same way by producer size.³

The analysis was executed as a staged coding pipeline. The pipeline comprised five sequential steps. Step 0 established the reach-stratified corpus and performed producer profiling, assigning each post to one of four producer-type categories on the basis of verified or heuristic identification. Step 1 computed content-type and post-length distributions and applied word-count-based length tiers (short: ≤ 50 words; medium: 51–200 words; long: 201–500 words; article_length: > 500 words). Step 2 applied rule-based narrative coding across eight narratives (N1–N8), each defined by a keyword set and a minimum-match threshold, and produced a post-level co-occurrence matrix. Step 3 computed five heuristic disinformation indicators (C1–C5) and an aggregate disinfo_score, with risk tiers assigned as low (score 0), medium (score 1–2), or high (score ≥ 3). Step 4 cross-tabulated narratives, producer types, content types, and length tiers and produced the shard-level intermediate output files cited in section 1 of each shard report. All indicator values derived from steps 2–4 are automated-unverified heuristic proxies; only counts and arithmetic are verified.

What was measured and why

The primary unit of analysis is the individual Facebook post. Three dimensions were measured for each post: reach (views as reported in Meta's Content Library, the most direct available proxy for audience exposure), textual properties (post length in words, presence of external links, use of capitalisation), and content properties (narrative framing, heuristic indicators of disinformation risk). Reach was chosen rather than engagement metrics such as reactions or shares because the research question concerns the scale of audience exposure to specific framings, not the intensity of individual users' responses. Post length was measured because it bears directly on the character of the discourse: whether circulating content was declarative and headline-like or argumentative and elaborated. Narrative framing was measured because framing — in Robert Entman's formulation — involves selecting aspects of perceived reality and making them salient in ways that promote particular problem definition, causal interpretation, moral evaluation, and/or treatment recommendation⁴. Identifying which frames were predominant, and which producers carried them, is therefore analytically prior to any assessment of disinformation risk.

Producer typing was included as a cross-cutting variable because the structural position of the source — professional media, political actor, or individual citizen account — conditions how a frame circulates and what authority it carries. The same choice also bears on the corpus design itself: as the comparison between Meta's downloadable dataset and the manual collection demonstrates, restricting analysis to large verified accounts excludes a structurally significant part of the information environment. Measuring producer type in a corpus that retains sub-threshold accounts makes it possible to ask whether the reach and framing patterns differ between institutional and non-institutional producers — a question that is simply not answerable from the downloadable dataset alone.

How it was measured

Post-length tiers were assigned on the basis of word counts computed from the post text field: short (1–50 words), medium (51–200 words), long (201–500 words), and article-length (over 500 words). Thresholds were set to distinguish minimal declarative posts from extended analytical ones and are not derived from an external standard; they are descriptive conventions used consistently across both shards. Narrative coding was implemented as a keyword-match pipeline: for each of the eight narratives (N1–N8), a lexical set was defined and a minimum-match threshold applied to control precision. A post receives a narrative flag if the number of distinct keyword matches in its text meets or exceeds the threshold for that narrative. Confidence values reflect keyword density relative to threshold and are capped at 1.0; they do not represent probabilistic estimates validated against human judgement.

Disinformation-indicator scoring applied five surface-level heuristic criteria (C1–C5) to each post: absence of source attribution, link-in-comments evasion, sustained all-caps use, presence of phrases signalling unverifiable extraordinary claims, and urgency or amplification language. The aggregate disinfo_score is the sum of active criteria and it is not a validated disinformation classification. All narrative flags and disinformation scores are automated-unverified and require manual review before any interpretive conclusion can be drawn from them. This review was conducted on a limited basis within the scope of this study, primarily to determine whether this approach is valid and whether automated categorization has value that warrants further investigation (it has).

Narratives, framing, and disinformation: working definitions

A narrative in the present study denotes a recurrent, identifiable framing pattern — a cluster of co-occurring references, actor characterisations, and evaluative terms that organises the representation of a political event in a consistent direction. This is close to what Miskimmon, O'Loughlin, and Roselle⁵ call a 'strategic narrative': a means by which political actors attempt to construct shared meaning of the past, present, and future in order to shape the behaviour of other actors. The distinction from frame is largely one of scale and temporality: frames are micro-level devices operating within individual texts; narratives are patterns that persist across multiple texts and producers over time.

The relationship between narrative and propaganda is similarly not one of identity. Propaganda — in the Oxford English Dictionary's formulation, 'the systematic dissemination of information, esp. in a biased or misleading way, in order to promote a particular cause or point of view, often a political agenda'⁶ — involves narrative, but not every narrative is propaganda. Propaganda requires a systematic communication intent; a citizen posting a long commentary on an election result is deploying narrative framing, not necessarily conducting a propaganda operation. The analytical distinction that matters for this study is therefore a three-tier one: narrative (a recurrent framing pattern, value-neutral in itself), propaganda (narrative deployed systematically and intentionally to promote a political cause), and disinformation (content that is false or misleading and deliberately disseminated to cause harm or achieve political effect). A post can instantiate all three, or only one, or none; the conditions cannot be read off from surface-level keyword matches.

Narratives are not equivalent to disinformation, and the presence of a narrative flag in this dataset does not constitute evidence of false or misleading content. Political narratives — including partisan, evaluative, or mobilising ones — are the ordinary material of democratic communication. They may be selective in emphasis, they may favour one political actor over another, and they may serve the strategic interests of whoever promotes them; none of these features makes them disinformation per se. Following the now widely adopted formulation of Wardle and Derakhshan⁷, disinformation specifically designates content that is false and deliberately created to harm a person, social group, organisation, or country. This requires both a falsity condition and an intent condition — neither of which can be established by keyword detection alone. The MediaWell project's working definition similarly frames disinformation as ' a rhetorical strategy that produces and disseminates false or misleading information in a deliberate effort to confuse, influence, harm, mobilize, or demobilize a target audience'⁸ — a definition that foregrounds intentionality as the key differentiating variable.

How the narratives in this study were identified

Eight narrative categories were derived inductively from a preliminary reading of the corpus, guided by the immediate political context of the April 19 elections: electoral outcome framing (N1), which captures posts centred on results and their interpretation; pro-Radev victory framing (N2), which covers pro-Progressive Bulgaria and Radev-aligned evaluations of the outcome; anti-GERB / anti-Borissov framing (N3); pro-Russia / geopolitical reorientation framing (N4), covering posts that contextualise the election within a broader geopolitical realignment; electoral integrity and vote-buying (N5); anti-PP-DB framing (N6); international and regional reactions (N7); and conspiracy and sensationalist framing (N8). Each category was operationalised as a keyword set applied at a minimum-match threshold. The categories are not mutually exclusive: a post about Radev's victory can simultaneously frame the result in geopolitical terms and include an electoral integrity claim, as the co-occurrence data suggest (automated-unverified).

This approach identifies the presence of a thematic framing pattern in a post but it does not assess the truth value of the content, the intent of the producer, or the effect on the audience. Narrative flags should therefore be read as indicators of thematic salience, not as disinformation labels.

The heuristic disinformation indicators (C1–C5) are a separate, parallel layer that targets specific surface-level properties associated — in prior research and platform-level detection practice — with low-quality or deceptive information: absence of source attribution, link-burying, typographic emphasis, sensationalist phrasing, and urgency language. Neither layer alone, nor the two combined, can substitute for manual content analysis and fact-checking.

The main research questions were fourfold: which narratives dominated election-related Facebook content after the vote; which producer types carried the largest share of reach; how much of the observed material displayed weak sourcing or other heuristic indicators of disinformation risk; and what additional structure became visible when audience interaction intensity and co-occurrence patterns were examined. The study also asked a more basic methodological question: what is lost when Bulgarian-language election discourse is approached only through Meta’s downloadable research subset rather than through the full visible search results.

Limitations

First, the corpus is reach-stratified and does not represent the entire Facebook post universe. Second, producer typing, narrative coding, disinformation indicators, and long-form genre labels rely partly on heuristic rules rather than manual validation. Third, the dataset does not capture native reshare chains, and attribution embedded only in images or video may remain invisible to text-based coding. Fourth, the additional co-occurrence analyses are exploratory because they rely on surface-form tokenization and a compact actor lexicon rather than lemmatized or syntactically parsed text.

Executive summary

💡Load the interactive dashboard of the top 50 posts by reach here

Across the two analyzed reach strata, the post-election Facebook discussion was structured above all by two dominant narrative frames: electoral outcome framing (N1) and pro-Radev victory framing (N2)⁹. In Shard B, these two narratives appeared in 43 and 42 posts respectively and accounted for 48.7% and 46.0% of views, while in Shard A they appeared in 234 and 242 posts and accounted for 34.5% and 37.5% of views. Taken together, the two dominant frames accounted for 17.38 million views for electoral-results framing and 17.51 million views for pro-Progressive Bulgaria / pro-Radev victory framing across the two shards.

In both shards, Citizen/unclassified producers carried the largest share of circulation. They accounted for 74.7% of posts and 82.4% of views in Shard B, and 71.0% of posts and 70.3% of views in Shard A. This pattern means that post-election reach cannot be adequately mapped through established media pages alone, because a large share of observed reach was generated by pages and profiles outside the mainstream media category.

The evidence does not support a characterisation of this corpus as uniformly disinformation-saturated; the dominant register is a broad layer of political commentary in which victory interpretation, actor conflict, and institutional accusation are more salient than overtly conspiratorial content. In Shard B, the narrative field was already concentrated around N1 and N2, with smaller but visible anti-GERB, electoral-integrity, and geopolitical strands. In Shard A, the same structure persisted at larger scale, with especially strong overlaps between electoral outcome framing and pro-Radev victory framing (122 posts), between pro-Radev victory framing and anti-GERB / anti-Borissov framing (69 posts), and between pro-Radev victory framing and pro-Russia / geopolitical reorientation framing (53 posts), indicating that outcome framing, actor blame, and geopolitical positioning were often embedded in the same posts rather than appearing as separate discursive streams.

The additional Shard A analyses refine this picture. The reactions-to-views analysis showed that disproportionately high audience response was often attached to lower-reach posts that were not the most heavily flagged under the heuristic risk model; in the below-median-reach, top-decile-engagement subset, risk tiers were predominantly low or medium. This decoupling between engagement intensity and heuristic risk score is methodologically relevant: it suggests that reach-weighted risk profiling alone may miss content that generates strong audience responses at mid-range or lower reach.

The PMI-weighted lexical co-occurrence analysis identified a dense institutional cluster around “Кандев”, “Дечев”, “Гюров”, “Сарафов”, “главен секретар”, and “МВР”, alongside a geopolitical cluster¹⁰. Actor-pair extraction likewise showed frequent and stronger-than-chance co-occurrence among pairs such as DPS – GERB (69 posts, PMI 2.07) and Borisov – Peevski (65 posts, PMI 2.30), suggesting that recurring actor constellations were a constitutive part of the discourse rather than incidental mentions.

The main analytical conclusion is therefore twofold. First, the dominant post-election dynamics in the observed Facebook corpus were organized around electoral outcome interpretation and the political meaning of Radev’s victory, not around a single monolithic disinformation theme. Second, the method matters: once smaller producers are retained in the corpus rather than filtered out by follower thresholds, the structure of visibility changes substantially, and the resulting map of election-related discourse becomes more socially and politically heterogeneous.

These findings are provisional in one further respect: the analysis covers only Shards A and B. Cross-shard synthesis, including comparison of narrative prevalence, producer-type distributions, and disinformation-indicator profiles across the full reach spectrum, remains outstanding and is scheduled as the next analytical phase.

Part I – Shard B: High reach (100,000–1,000,000 views)

1. Shard overview

This shard contains 87 posts from 69 unique producers, all with views between 100,000 and 1,000,000.

Total views in Shard B: 19,653,500.

Processing mode: full 5‑step pipeline.

2. Corpus overview table

All values in this section are directly computed from the Shard B dataset (verified).

3. Producer type distribution

Producer typing combines verified identifications and heuristic assignments. Counts and view statistics are verified; producer type status for Citizen/unclassified is estimated.

Verification summary: 14 posts from verified producers and 73 posts from heuristic‑typed producers.

4. Content type distribution

All statistics in this section are verified.

5. Daily volume and reach

All values are verified.

6. Post length and long‑form genre analysis

From the early 2010s onwards, a widespread assumption among digital publishers held that audiences lacked the attention span for long-form text on social media, an assumption reinforced by the architecture of platforms such as Twitter, whose original 140-character limit functioned as a structural argument about what public communication should look like online. Publishers responded by shortening articles, fragmenting arguments, and treating brevity as a proxy for relevance widely documented in digital journalism research¹¹. The data in this shard run directly counter to that assumption. In Shard B — the highest-reach stratum in the corpus, covering posts with between 100,000 and 1,000,000 views — the two longest length tiers (long: 201–500 words; article_length: > 500 words) together account for 39.0% of posts and 48.7% of total shard views. Posts averaging 866 words reached audiences of 261,500 views on average. Extended analytical or opinion text, produced predominantly by Citizen/unclassified accounts rather than by professional media pages, was circulating at scales comparable to mainstream broadcast content. The length distribution therefore warrants attention not as a curiosity but as evidence that the post-election Facebook discourse in this reach stratum was substantially argumentative and discursive in character, not merely declarative or image-driven.

6.1 Length tier distribution (all posts)

All statistics in this subsection are verified.

6.2 Length tier × producer type (post counts)

Counts and views are verified. Producer type is estimated for Citizen/unclassified; all counts and view statistics are directly computed

6.3 Long‑form genre and article‑length posts

In this shard, all long‑form posts (> 200 words) are currently set to GENRE_UNCLASSIFIED; any genre‑based conclusions are automated‑unverified.

Article‑length posts (> 500 words) in Shard B are:

Counts and scores are verified; the risk interpretation is automated‑unverified.

7. Amplification and duplicate clusters

All counts and view sums in this section are verified; amplification cluster types are automated‑unverified.

Cluster summary:

Total views in non‑primary copies: 825,900; share of shard views: 4.2% (verified).

8. Narrative coding results

All narrative flags in this section are automated‑unverified. Counts and arithmetic are verified, but the semantic assignment of narratives is based on keyword rules, not manual coding.

Narrative co‑occurrence (counts of posts where both narratives fire) shows strong overlaps between N1 and N2, and frequent co‑occurrence of N1/N2 with N3 and N4 in long analytical or opinion posts (automated‑unverified).

Narrative density distribution (number of narratives per post, verified) is:

• 0 narratives: 21 posts
• 1 narrative: 28 posts
• 2 narratives: 21 posts
• 3 narratives: 7 posts
• 4 narratives: 7 posts
• 5 narratives: 3 posts

9. Disinformation indicator scoring

All criteria and disinfo scores in this section are automated‑unverified. Counts and sums are verified, but indicators are heuristic proxies, not validated disinformation labels.

9.1 Criterion counts (C1–C5)

9.2 Disinfo score distribution

There are no posts with disinfo_score ≥ 3 in Shard B (verified).

10. Cross‑tabulation key findings

• Finding 1 (verified for counts, automated‑unverified for narrative labels): High‑reach posts in this shard frequently combine outcome framing (N1) and pro‑Progressive Bulgaria victory narratives (N2), each appearing in about half of posts (43 and 42 posts respectively) and accounting for 48.7% and 46.0% of shard views.
• Finding 2 (verified for counts, estimated for producer types): Anti‑GERB (N3) and Anti‑PP‑DB (N6) narratives are present but more selective, with N3 appearing in 15 posts (15.6% of views) and N6 in 4 posts (6.3% of views), all N6 posts originating from Citizen/unclassified producers.
• Finding 3 (verified for counts, automated‑unverified for narrative labels): Pro‑Russia / geopolitical reorientation narratives (N4) occur in 11 posts, representing 12.7% of shard views; mainstream media posts in N4 achieve higher average reach (about 306,500 views/post) than Citizen/unclassified posts (about 197,200 views/post).
• Finding 4 (verified for scores, automated‑unverified for their interpretation): All posts in this shard have disinfo_score 0 or 1; 56.3% of posts and 60.8% of views are low‑risk (score 0), 43.7% of posts and 39.2% of views are medium‑risk (score 1), with no high‑risk posts (score ≥ 3) under the current indicator scheme.
• Finding 5 (verified for counts, automated‑unverified for narrative and genre labels): Article‑length posts (> 500 words) are concentrated among Citizen/unclassified producers (8 of 9 posts) and often carry multiple narratives, yet all have disinfo scores 0 or 1, indicating low or medium risk in this scheme.

Part II – Shard A: Mid-reach (13,700–99,800 views)

1. Shard overview

This shard contains 682 posts from 423 unique producers, all with views between 13,700 and 99,800.

Total views in Shard A: 22,587,800.

Date range in Shard A: 2026-04-20 – 2026-04-27.

Processing mode: full 5-step pipeline plus additional analyses.

2. Corpus overview table

All values in this section are directly computed from the Shard A dataset (verified).

3. Producer type distribution

Producer typing combines verified identifications and heuristic assignments. Counts and view statistics are verified; producer type status for Citizen/unclassified and Political commentator/blogger is estimated.

Verification summary: producer-type assignment for the majority of Citizen/unclassified pages remains heuristic and estimated; counts and reach statistics are verified.

4. Content type distribution

All statistics in this section are verified.

5. Daily volume and reach

All values are verified.

6. Post length and long-form genre analysis¹³

The length distribution in Shard A presents a more evenly divided picture than Shard B. Short posts (221, 32.4%) and article-length posts (82, 12.0%) coexist in a corpus where medium and long posts together account for 55.5% of posts and 56.5% of views. Article-length posts (> 500 words, averaging 805 words) account for 12.9% of views — slightly above their 12.0% share of posts — indicating that extended text sustained at least proportional reach even at mid-range visibility levels. As in Shard B, these posts are concentrated almost entirely among Citizen/unclassified producers (70 of 82), and none crosses a disinfo score of 2

6.1 Length tier distribution (all posts)

All statistics in this subsection are verified.

6.2 Length tier × producer type (post counts)

Counts are verified; producer type interpretation is estimated where Citizen/unclassified appears.

6.3 Long-form genre and article-length posts

Long-form genre labels (genre_commentary, genre_prose, genre_roundup, genre_activation, genre_unclassified) are automated-unverified heuristic assignments. Numeric post counts and scores are verified.

Article-length posts (> 500 words) in Shard A include (top 20 by wordcount shown):

Counts and scores are verified; the risk interpretation is automated-unverified.

7. Amplification and duplicate clusters

All counts and view sums in this section are verified; amplification cluster types are automated-unverified.

Shard A contains 53 amplification clusters. A selection of structurally notable clusters is shown below.

Total views in non-primary copies: 1,745,100; share of shard views: 7.7% (verified).

8. Narrative coding results

All narrative flags in this section are automated-unverified. Counts and arithmetic are verified, but the semantic assignment of narratives is based on keyword rules, not manual coding.

Narrative co-occurrence (counts of posts where both narratives fire) shows strong overlaps between N1 and N2 (122 posts), N2 and N3 (69 posts), N2 and N4 (53 posts), and N1 and N3 (50 posts), indicating that evaluative and actor-framed narratives are frequently embedded within outcome and victory discourse (automated-unverified).

Narrative density distribution (number of narratives per post) is:

• 0 narratives: 271 posts
• 1 narrative: 212 posts
• 2 narratives: 110 posts
• 3 narratives: 54 posts
• 4 narratives: 27 posts
• 5 narratives: 7 posts
• 6 narratives: 1 post

9. Disinformation indicator scoring

All criteria and disinfo scores in this section are automated-unverified. Counts and sums are verified, but indicators are heuristic proxies, not validated disinformation labels.

9.1 Criterion counts (C1–C5)

9.2 Disinfo score distribution

There are no posts with disinfo_score ≥ 3 in Shard A (verified)

10. Cross-tabulation key findings

• Finding 1 (verified for counts, automated-unverified for narrative labels): N1 (electoral results framing) and N2 (pro-Progressive Bulgaria / pro-Radev victory) are the two most prevalent narratives, appearing in 234 and 242 posts respectively, and accounting for 34.5% and 37.5% of shard views. This structural dominance is consistent with the shard's post-election timing.
• Finding 2 (verified for counts, automated-unverified for narrative labels): N2 and N1 co-occur in 122 posts, the single largest narrative overlap in the shard, indicating that outcome framing and pro-Radev victory framing are frequently combined rather than mutually exclusive in mid-reach content.
• Finding 3 (verified for counts, estimated for producer types): Anti-GERB (N3) and electoral integrity (N5) narratives appear in 92 and 81 posts respectively and account for 13.7% and 11.9% of views. N6 (anti-PP-DB) is narrowly distributed with 15 posts (2.5% of views). All N3, N5, and N6 posts originate predominantly from Citizen/unclassified producers.
• Finding 4 (verified for counts, automated-unverified for narrative labels): N4 (pro-Russia / geopolitical reorientation) occurs in 73 posts (10.5% of views). Political actors achieve higher average reach in N4 (approximately 49,125 views per post) than Citizen/unclassified posts (approximately 30,905 views per post), indicating that geopolitical framing from institutional actors travels further per post. This is consistent with top-down amplification of geopolitical frames and it requires verification of producer identities.
• Finding 5 (verified for scores, automated-unverified for their interpretation): All posts in this shard have disinfo_score 0, 1, or 2, with no high-risk posts (score ≥ 3). 58.1% of posts and 58.4% of views are low-risk (score 0), 41.9% of posts and 41.6% of views are medium-risk (score 1 or 2), under the current indicator scheme.
• Finding 6 (verified for counts, automated-unverified for narrative and genre labels): Article-length posts (> 500 words) total 82 posts (12.0% of the corpus) and 12.9% of shard views. They are concentrated among Citizen/unclassified producers (70 of 82 posts) and have risk profiles of mostly low or medium, with none crossing score 2.

11. Additional analysis: reactions-to-views ratio and lexical co-occurrence

This section contains analyses run specifically for Shard A and not implemented on Shard B. All indicators in this section are automated-unverified; all count and ratio values are verified.

11.1 Reactions-to-views engagement ratio

The engagement-by-reach metric is defined as:

applied to posts with non-zero views. The dataset exposes reactions as a single aggregate total (no split between comment types or reaction types). Median non-zero reach in Shard A is 24,950 views. The top-decile engagement rate threshold is 0.0462 reactions per view. Posts defined as disproportionately reactive are those with views at or below the median and engagement rate at or above that threshold.

Top 10 posts by reactions-to-views ratio in the high-engagement, below-median-reach subset (verified for all numeric values; producer typing estimated):

The high-engagement, below-median-reach subset consists almost entirely of Citizen/unclassified producers. Risk tiers are mostly low or medium; no high-risk (score ≥ 3) post appears in this subset. This indicates that disproportionate audience interaction in Shard A is not concentrated in the most heavily flagged posts; lower-reach content can sustain strong interaction intensity without crossing the high-risk threshold under the current C1–C5 scheme (automated-unverified for interpretation).

11.2 Lexical co-occurrence: PMI-weighted pairs

Word-level co-occurrence was computed at post level over the top-200-token vocabulary (lowercased, punctuation-stripped, minimum token length 3, no lemmatization). Pointwise mutual information (PMI) is defined as:

where probabilities are estimated from post-level presence counts. Only pairs with raw co-occurrence ≥ 5 posts are retained.

Raw token frequencies confirm that the corpus is election-centred and actor-centred: the top tokens are "това" (989 occurrences), "изборите" (748), "като" (742), "радев" (709), and "българия" (594).

Top 15 PMI-weighted co-occurring pairs (verified for counts and PMI values; semantic interpretation automated-unverified):

The highest-PMI cluster centres on MVR institutional roles and personnel (Кандев, Дечев, Гюров, Сарафов, главен секретар, МВР). The pair "русия – украйна" (PMI 3.27) indicates a consistently co-embedded geopolitical framing. The pair "БСП – възраждане" (PMI 2.44, in the full table) also stands out as a stronger-than-chance association.

PMI results should be read as exploratory and must not be treated as validated semantic relationships. No lemmatization or stopword filtering was applied, so some associations involve function-word co-occurrence that reflects discourse structure rather than topical proximity.

11.3 Actor and party co-occurrence

Actor co-occurrence was computed using a short lexicon of single-token forms mapped to actor labels (radев, borisov, peevski, pp_db, gerb, bsp, vazrajdane, dps, progressive_bg). The mapping is heuristic and partly over-inclusive; results are automated-unverified.

Actor post counts (number of posts containing at least one token mapped to the label): radев 234, gerb 127, peevski 100, progressive_bg 93, borisov 90, dps 88, bsp 55, vazrajdane 48 (verified for the specific tokenized forms used).

Top 15 actor/party pairs by co-occurrence count (verified for counts; PMI values verified as computed; interpretation automated-unverified):

Highest PMI values (strongest-than-chance associations) include: borisov – peevski (2.30), dps – gerb (2.07), dps – vazrajdane (2.12), bsp – vazrajdane (2.44), bsp – dps (1.82), bsp – gerb (1.77), gerb – vazrajdane (1.75). These consistently elevated PMI values among opposition-aligned and coalition-opposite pairs suggest that these actor constellations are discussed together more frequently than expected from their individual prevalence alone, consistent with coalition or "status quo" framing patterns (automated-unverified for interpretation).

12. Caveats register

Note on scope: this caveats register applies to both Shard B and Shard A.

1. Citizen/unclassified producer verification status: the majority of producers in this shard remain heuristically typed; aggregates including this category are estimated for producer typing (automated-unverified for interpretation).
2. C1 false positive floor: "No source attribution" may flag Photo-type and Video-type posts where sources appear in images or video rather than in text, leading to systematic over-flagging (automated-unverified).
3. C4 low-hit interpretation: the low C4 count (4 posts) does not imply that few unverifiable or extraordinary claims are present; it indicates only that the specific C4 phrase list was matched rarely (automated-unverified).
4. Narrative confidence saturation: mean confidence values are capped at 1.0 by design and reflect keyword density relative to thresholds, not human validation (automated-unverified).
5. Narrative flags are automated-unverified: the semantic assignment is based on keyword rules, not manual coding.
6. Amplification cluster detection lower bound: amplification clusters are identified only via text similarity (Jaccard ≥ 0.74) and do not capture all forms of coordination, including organic sharing and image-only duplication (automated-unverified).
7. Absence of organic reshare tracking: the dataset does not record native reshares; reach figures reflect only the posts in this corpus, not downstream sharing chains (verified for scope, automated-unverified for interpretation).
8. Reach-stratified shard: Shard A covers only posts with approximately 13,700–99,800 views and is not directly generalisable to lower-reach content (below 13,700 views) or to the high-reach stratum (Shard B) without cross-shard synthesis (verified for reach bands, automated-unverified for generalisation).
9. Long-form genre classification: long-form genres (genre_commentary, genre_prose, genre_roundup, genre_activation, genre_unclassified) are assigned by heuristic rules and are automated-unverified; any conclusions about genre should be treated as exploratory.
10. Engagement ratio and lexical co-occurrence analyses: the reactions-to-views ratio is based on a single aggregated reactions field; no distinction between reaction types or comments is possible. Lexical co-occurrence is unlemmatized, uses document-window co-occurrence without syntactic filtering, and the actor lexicon is a shallow heuristic. All additional-analysis results are automated-unverified (section 11).
11. Non-Bulgarian posts: 12 posts in Macedonian are included in all cross-shard corpus totals but should be excluded from Bulgarian-only aggregates in downstream synthesis. Narrative flags and disinformation scores for those posts are as computed and are included in the full intermediate files.

13. Working plan for improvement and verification

Verification tasks

• Verify Citizen/unclassified producer identities across both shards through external sources (media registries, official party pages, publicly listed profiles) to upgrade producer-type labels from estimated to verified. In Shard B, 73 of 87 posts carry heuristic producer types; in Shard A, the majority of 484 Citizen/unclassified posts remain unverified.
• Conduct manual review samples of posts flagged under N3 (anti-GERB), N5 (electoral integrity), and N6 (anti-PP-DB) in both shards to validate framing direction and confirm keyword rules capture the intended evaluative stance.
• Review article-length posts in both shards to assign validated long-form genre labels (commentary, opinion, roundup, mobilisation), replacing current genre_unclassified and heuristic automated-unverified assignments.

Analytical extensions

• Apply image OCR and video frame analysis to identify source attributions in visuals, reducing C1 false positives. In Shard A, Photos account for 55.1% of posts; in Shard B, Photos and Videos/reels together account for 72.4%.
• Extend the Shard A additional analyses (reactions-to-views ratio, PMI-weighted lexical co-occurrence, actor/party pair co-occurrence) to Shard B.
• Apply Bulgarian morphological lemmatisation and a domain-appropriate stopword list to the co-occurrence pipeline to sharpen topical signal and reduce noise from high-frequency function words.
• Conduct close reading of the MVR/institutional personnel cluster identified in Shard A (Кандев, Дечев, Гюров, Сарафов, главен секретар, МВР) to determine whether co-occurring tokens reflect a coherent institutional-critique narrative or distinct framings sharing surface vocabulary.

Cross-shard synthesis tasks

• Apply the full pipeline (Steps 0–4) and the additional analyses run on Shard A to a lower-reach stratum (posts with fewer than 13,700 views), once such a corpus is assembled.
• Once the lower-reach stratum is processed, conduct a cross-shard synthesis comparing narrative prevalence, producer-type distributions, disinformation-indicator profiles, amplification cluster density, and long-form genre distributions across both reach strata and the lower-reach corpus.
• Document the collection methodology for the lower-reach stratum using the same collection protocol applied to Shards A and B.

Pipeline and documentation tasks

• Reconcile the N1 threshold difference between shards (threshold 1 in Shard B; threshold 2 in Shard A) by re-running Shard B Step 2 at threshold 2 and reporting both versions, so cross-shard narrative counts rest on a consistent basis.
• Separate the 12 Macedonian-language posts in Shard A from Bulgarian-language aggregates in all cross-shard synthesis outputs, and assess whether their narrative and risk profiles differ systematically from the Bulgarian-language majority.

Appendix

Borchgrevink-Brækhuus, M. (2026). An experiential view of what short digital news practices mean from an audience perspective. Journalism Studies. Advance online publication. https://doi.org/10.1080/1461670X.2026.2633581;
Carlson, M. (2018). Confronting measurable journalism. Digital Journalism, 6(4), 406–412. https://doi.org/10.1080/21670811.2018.1445003;
Ferrer-Conill, R., & Tandoc, E. C., et al. (2023). Popularity-driven metrics: Audience analytics and shifting opinion power to digital platforms. Journalism Studies. https://doi.org/10.1080/1461670X.2023.2167104; Kristensen, L. M. (2023). Audience metrics: Operationalizing news value for the digital newsroom. Journalism Practice, 17(5). https://doi.org/10.1080/17512786.2021.1954058;
Pew Research Center. (2016, May 5). Long-form reading shows signs of life in our mobile news world. https://www.pewresearch.org/journalism/2016/05/05/long-form-reading-shows-signs-of-life-in-our-mobile-news-world;
Sangiorgio, E., Etta, G., Di Marco, N., et al. (2026). Examining the relationship between content length and engagement with news outlets on multiple social media platforms. Technological Forecasting and Social Change. https://doi.org/10.1016/j.techfore.2026.S0040162526001459;
Steensen, S., et al. (2024). Quality journalism in social media – What we know and where we need to dig deeper. Journalism Studies. https://doi.org/10.1080/1461670X.2024.2314204;
Toff, B., et al. (2024). The power of numbers: Four ways metrics are transforming the news. Digital Journalism. https://doi.org/10.1080/21670811.2024.2323655;

Van Remoortere, A., et al. (2024). Dissecting social media journalism: A comparative study across platforms, outlets and countries. Journalism Studies. https://doi.org/10.1080/1461670X.2024.2324318

1 An approximation based on the reported valid vote total https://btvnovinite.bg/bulgaria/sociolozi-blizo-3-3-mln-balgari-sa-glasuvali-na-parlamentarnite-izbori.html?campaignsrc=clipboard

2 Like all other analyses conducted by the BROD team as part of its monitoring of the campaign for the parliamentary elections in Bulgaria on April 19, 2026: https://brodhub.eu/en/research/ The search used the definite form of 'elections' in Bulgarian, which covers the full range of references to the April 19 vote without restricting to a specific party or candidate name

3 BROD. 2006. Monitoring of public and political discourse related to the elections on Facebook during the week leading up to Election Day https://brodhub.eu/en/research/monitoring-of-public-and-political-discourse-related-to-the-elections-on-facebook-during-the-week-leading-up-to-election-day/

4 Entman, R. 1993. Journal of Communication, 43(4), 51–58, DOI: 10.1111/j.1460-2466.1993.tb01304.x.

5 Miskimmon, A., O'Loughlin, B., & Roselle, L. 2013. Strategic narratives: Communication power and the new world order. Routledge.

6 Propaganda at Oxford English Dictionary: https://www.oed.com/dictionary/propaganda_n#

7 Wardle, C. & H. Derakshan (September 27, 2017) Information Disorder: Toward an interdisciplinary framework for research and policy making, Council of Europe (PDF)

8 Spies, S. 2019. Defining “Disinformation”. MediaWell https://mediawell.ssrc.org/research-reviews/defining-disinformation/

9 N1 captures factual or evaluative reporting of results (including critical or neutral framings), while N2 specifically centres Radev and Progressive Bulgaria as winners

10 These are the names of senior Ministry of Interior and prosecution officials whose appointments and roles were publicly contested in the period surrounding the election

11 The literature broadly suggests that while the production side has clearly moved toward shorter, more digestible formats under social media pressure, the empirical evidence on audience behaviour is more ambivalent — engagement data often favours longer, higher-quality content, creating a productive tension at the center of this research area. See Appendix for recommended readings (bibliography)

12 People add links in the comments, rather than the original post, primarily to avoid algorithmic penalties on platforms like Facebook and LinkedIn. Social media algorithms often suppress posts containing external links because they drive users away from the app, causing posts with links to receive significantly lower engagement.

13 See the Post length and long form genre analysis description on page 8

Methodological note:

This study presents a broad overview, some interesting findings, and, above all, a methodology that can be replicated and has proven its merits as an approach to analyzing political and public discourse in the context of elections.

The data analysis was assisted by Perplexity Pro and Claude Sonnet 4.6 Thinking.
The report explicitly highlights all instances where data categorization was automated and performed heuristically by the model; despite sample checks and manual reclassification, the reliability of the categorization remains problematic (especially regarding the categorization of producers due to the large volume of posts and producers). The limitations of computational linguistics are also explicitly stated.