Work / Story 5 of 5
Reading a cross-market controversy on Chinese social media
My master's thesis: 5,289 Weibo comments about a Thai performer, a Paris venue and a Chinese-language argument over freedom, read by what people actually saw rather than by how often things were said.
The problem
Brands watching a controversy usually count mentions. But on a platform ranked by likes, a handful of comments carry most of what anyone actually reads, so counting everything equally describes a conversation nobody had. I wanted to know how a shared opinion forms in that gap, using a case that crossed three markets at once.
What I did
I collected 5,289 comments under 219 Weibo posts and sampled by visibility instead of volume: every comment above 500 likes (246 of them, carrying 86.9% of all the likes), plus a random tail of 200 so I could still say something about the quiet majority. I coded 446 comments, close-read 29, and wrote down all 28 borderline judgement calls so another person could check the work. I did not clean the corpus first, because filtering to the relevant comments would have turned rarity into something I had created myself.
The result
Passed public defence at Stockholm University, August 2026. The two vocabularies moved in opposite directions: nation words appeared in 260 comments, class words in only 98, yet class-framed comments drew about three times the likes each (699 against 231). Counting would have called the argument nationalist. Weighting by what people saw showed it was settled by class, and settled by renaming the thing rather than by arguing about it.
What this means for a brand
Share of voice measures who spoke. This measures who was seen, which is the number that predicts what a newcomer to the thread ends up believing. The same design works on a product launch or a brand incident on Chinese platforms: weight by attention, keep the raw corpus, report frequency and reach as two separate figures, and watch which label is winning rather than which way a sentiment score moved. Then leave the decision log, so the insight is something a colleague can check instead of take on trust.
Download the thesis (PDF, 0.8 MB)
the argument was settled by renaming, not by arguing
B. Context
This was my MA thesis in Global Media Studies at Stockholm University, completed in 2026 and passed at public defence in August 2026. The case was a public event: the 2023 performance by Lisa of the K-pop group BLACKPINK at the Crazy Horse cabaret in Paris, and the weeks of argument it set off in Weibo comment sections between September and November 2023. I treated those comment sections as a listening problem: which readings of “freedom” and “art” became the visible common ground, and how. The work sits between academic discourse research and the social listening used in brand and reputation work, and the transferable part is the method rather than the topic.
C. The problem
On Weibo, like-ranking decides which comments a newcomer meets first, so a listening report that counts what is said can miss what is seen. Without a visibility logic, a rare, low-visibility view can end up on the same footing as the reading that thousands of people met at the top of the thread. The cost is a wrong picture of where the argument actually stood.
D. My role and the team
This was individual work. I designed the sampling, ran the collection, wrote the sampling and counting scripts, coded every sampled comment, close-read the extracts and wrote the thesis. A supervisor gave feedback through the department’s course process and an external reader commented on the final revision; neither made coding or analytical decisions. Two examiners assessed the thesis and set four required revisions, which I completed in September 2026. AI tools were used as assistants for checking, drafting support and file verification and are declared in the thesis; a person made every coding and analytical decision.
E. Key decisions
-
Decision: Weight the sample by visibility, not by count: code every comment with at least 500 likes (246 comments carrying 86.9 per cent of all likes), then check the rest with a stratified random tail sample of 200. Why: On a like-ranked platform, visibility is the audience-side fact. In listening terms, this asks who was seen, not who spoke. Alternative I rejected: A simple random sample of all 5,289 comments, which would treat an unseen comment and a top-ranked one as equal evidence.
-
Decision: Keep the corpus uncleaned: no removal of fan-war, off-topic or one-word comments before sampling. Why: Claims about absence, such as how rare an argued defence was, only hold against the record as collected. Cleaning would turn rarity into an artefact of my own filter. Alternative I rejected: Filtering to “relevant” comments first, which is common in listening dashboards and quietly changes the denominator.
-
Decision: Two stages, two jobs: reflexive thematic analysis to map patterns across the coded comments, then critical discourse analysis of 29 extracts to show how the patterns work at the level of wording. Why: The map says where to look; the close reading says what stands there. Alternative I rejected: A single close-reading pass, which cannot say anything about distribution.
-
Decision: Replace a second coder with a documented second pass: a full re-read of all coded rows with a written log of 28 rulings on borderline cases, kept in the Excel ledger. Why: This was a solo project. The log makes each judgment auditable, which is the practical need in a small team. Alternative I rejected: Reporting a reliability statistic that would have implied a coder I did not have.
-
Decision: Report frequency and visibility as two numbers, never one. Why: For the nation and class vocabularies the two measures pointed in opposite directions, and collapsing them would have hidden the finding. Alternative I rejected: A single share-of-voice figure.
F. What I built or produced
- A corpus of 5,289 comments under 219 posts, collected with Octoparse using a workflow written for the study.
- A three-tier sampling design: 246 comments at or above 500 likes, a random tail of 200 in four like bands of 50 with a fixed seed, and a purposive supplement of 2 kept outside all statistics.
- An Excel coding ledger: coding scheme, coded rows for all tiers, theme development sheets, a post reference table and the second-pass decision log.
- A keyword specification for coverage counts of nation and class words across all 5,289 comments, with a sensitivity list.
- A two-stage analysis: 446 comments coded in the statistical tiers (plus 2 purposive extracts outside all counts), 29 close-read, 18 further comments cited as supporting instances, all traceable by row number.
- The thesis itself, 78 numbered pages, including an overview table of data selections and methods added at revision.
G. Result
The findings hold for this corpus and this platform, read like-weighted, and are not offered as a general rule. The contest was settled by renaming rather than by argument: “art” was re-labelled, and among the 100 most-liked comments there was no argued defence of the performance as a free choice. The heaviest critiques conceded “her freedom” and redefined it as something distributed by class. Nation words touched more comments than class words (260 against 98), but class-framed critiques drew roughly three times the likes per comment (699 against 231 on average). The findings chapter is organised by theme.
For a listening or insight role, the transferable capabilities are: designing a sampling logic around reach rather than volume, building and documenting a coding frame, separating mention counts from engagement weight, reading a cross-market controversy (a Thai performer, a French venue, a Chinese platform) without importing one market’s frame, and leaving an audit trail another person can follow. The same design could serve pre-launch listening or reputation monitoring on Chinese platforms; no commercial deployment has been made.
H. What I would do differently
- Collect during the event, not after. The corpus was pulled from hot-sorted threads long after 2023, without timestamps, and 37.2 per cent of comments had no like data. I would run repeated pulls during the controversy with timestamps and record the sort order at each pull.
- Add a second platform. Weibo’s like economy shaped what was visible; a parallel read of at least one other Chinese platform would show whether the renaming pattern is platform-specific.
- Add a formal intra-coder check: a second coding of a fixed subset after a time gap, with agreement reported.
I. Evidence assets
- Thesis PDF: Public link, Tier A once the DiVA record is open access; until then own file.
- Sampling funnel diagram (5,289 to 446 to 29 to 1 finding): Anonymised diagram, Tier B, to be drawn from the thesis figures.
- Coding frame structure (clusters, comment counts and like totals): Anonymised diagram, Tier B.
- Second-pass decision log, structure only: Anonymised description, Tier B.
- Keyword specification word lists: Anonymised description, Tier B.
- Coding ledger in full, raw corpus, any comment original, translation or username, examiners’ assessment and discussant review: Tier C, not shown.
J. Keywords
Social listening; Discourse analysis; Reflexive thematic analysis; Critical discourse analysis; Weibo; Chinese social media; Audience insight; Cross-market communication; Qualitative research design; Sampling design; Reputation monitoring; Research documentation