A Dance Clip Filed Under Football Data: A Labelling Error and What It Costs Vietnamese Football

A Dance Clip Filed Under Football Data: A Labelling Error and What It Costs Vietnamese Football

On Tuesday night, in a small apartment on Nguyen Chi Thanh Street in Da Nang, I opened a batch of 400 clips that an aggregation system had tagged “football” and downloaded to my machine. I watched them the old way: counting bodies in the frame, locating the shape of the block, tracking the runs of the player without the ball. At clip thirty-seven, I stopped. In the frame was a woman dancing, forty-five seconds long, stage lighting, no ball, no player, no referee. The label still read “football”. I watched it three times, as is my rule, and only then wrote two words in my notebook: wrong label.

I did not laugh. In the summer of 2026, when world football froze, I spent four months hand-counting every counter-attacking sequence in 200 V-League and First Division matches from the 2026 season, building a 5,000-row table. I know that one bad cell can skew an entire column. At 67, I also know that my trade — reading a match — begins with trusting the label written above each clip. The ball is dead on the pitch, but intent is alive in every metre of movement; and that intent only appears when the clip is filed in the right place.

The context of a wrong label

The source article I read was an entertainment piece in a general-interest newspaper. Stripped down, it ran as follows: a 45-second clip of a famous singer’s performance resurfaced, spread through aggregator pages, and triggered an online argument about whether the behaviour shown was appropriate for a workplace. A person in the background was described as “possibly an employee”; that person’s identity and employment status were not confirmed. A repost by one aggregator drew more than 1.4 million views. People defended, people attacked, the room split in two, and both halves lacked context.

Nowhere in that text was there a club, a player, a coach, a competition, or a single financial figure. Yet the record was filed under “football” at the first processing layer and sat in the queue waiting to be analysed as a sporting event.

I tell this story not to catch out a label. I tell it because this is a case study that surfaces exactly where my interest lies: how we classify, and then how we trust the classification.

The mechanism behind the failure is simple. An automatic classifier scores a text against a list of member keywords and a confidence threshold. Vietnamese is fertile ground for this kind of error, because short words collide easily: “sân khấu” (stage) contains “sân” (pitch); “đội ngũ nhân viên” (staff team) contains “đội” (team); “bàn làm việc” (desk) contains “bàn” (goal); “phạt tiền” (fine) contains “phạt” (penalty). Set the threshold low enough and an entertainment item slips straight into the football corpus, with nobody accountable for it being there.

Fixing it is no harder: put a gate at the end of the pipeline and ask three questions. Does the text name a club? Does it name a player or a coach? Does it name a competition or a match? Fail all three and the record leaves the corpus, however high the keyword score. I call those three questions the entity test, and it costs far less than cleaning up afterwards.

The second thing worth noting is how the source hedges itself. The key detail — the person in the background “possibly an employee” — is written in conditional language and attributed to another aggregator. That style is familiar to me, because I meet it every week on Vietnamese football pages: “a source close to the deal”, “reportedly”, “in all likelihood”. Conditional language keeps the writer safe and lets the reader fill the gap. Once filled, the reader carries away a conclusion that no source will own.

Analysis: a wrong label does not live alone

Wrong labels are inherited. A record that enters a corpus passes through entity extraction, through sentiment models, through league-monitoring dashboards, through reports sent to sponsors. That 45-second dance clip, with more than 1.4 million views, will be read by a model as “fan interest in football”, then quietly added to a brand-reach figure, then selected by an editor hunting trending topics precisely because it sits in the football folder. Nobody lies anywhere in that chain. A single wrong label is simply replicated five times over.

For Vietnamese football, the risk is not an entertainment clip. The risk is that we already maintain a far larger layer of junk data that generates itself without any algorithm going wrong: clip culture.

A ten-second tackle, cut away from the twenty seconds before it, becomes proof of a centre-back’s laziness. A frozen VAR frame becomes a verdict on a goal. A miss in the 88th minute becomes a story about nerve, when the real cause lies in the rhythm of the run-up and the pause before the whistle — things that only appear when you rewatch from a second camera angle. And this is the part worth saying out loud: most conclusions about Vietnamese football are drawn from clips cut exactly at the point that makes the conclusion easiest.

I ran into this in 2026, at 58. I sat in a Da Nang coffee shop rewatching twelve Hanoi FC matches and logging every transition. The result: 37 percent of their goals came from counter-attacks completed in eight seconds or less, while their defensive line pushed high without cover. I had to draw the 5-3-2 out of possession and the 3-4-3 in possession myself. No data provider sold me a “time to complete the counter” column. I had to count it by hand. Later, head coach Chu Dinh Nghiem himself acknowledged that gap in an interview.

The lesson sits there: the most important column in football data is usually the column nobody has built yet. A ready-made label does not contain an answer; it only contains a question.

In the summer of 2026, with no matches to commentate, I opened my laptop and hand-counted 200 games, logging the first position of the ball, the number of passes, and the time each counter took to complete. From those 5,000 rows I built an index of my own: the useful-space index — the area of the pitch an attacking team occupies in the first three seconds after winning the ball. It sounds dry. But when football returned in September, it let me say on air why SLNA collapsed in the 90th minute, and say it correctly. Three seconds from your own box to the opponent’s is enough to rewrite history.

The same reading applies to things that look unrelated to tactics. Before Belgium met Japan in the 2026 World Cup round of sixteen, I watched five matches from each side and noticed a detail off the pitch: coach Roberto Martinez usually left Fellaini and Chadli on the bench. I said on air that if Belgium fell behind, those two would come on to break Japan’s high press. The match followed the script: Japan led 2-0, the changes came at the 65th and 74th minutes, Belgium won 3-2. The bench list was a label. The intent behind it was the analysis. A team sheet says nothing at all until the ball is lost. The strong attack with the ball; the clever attack with space.

I have built a three-pass viewing routine, and the third pass is the one that most often overturns a conclusion. The first pass is live viewing, to pick up the rhythm. The second is the wide angle, to see the whole block shift. The third looks at one man only: the player furthest from the ball in the situation under the microscope. In one 2026 V-League match, a centre-back was jeered by the whole stand for losing the man who scored. On the third pass, the player who should have tracked with him had abandoned his position a beat earlier, after a sideways pass was lost in midfield. The label “centre-back error” was applied in four seconds. Removing it took twenty minutes.

The same principle applies to the books. In the Vietnamese transfer market, a deal recorded as “free” leaves the transfer-fee column at zero. But signing-on money, agent fees, housing and bonuses flow into the wage column instead. On the balance sheet the bargain looks cheap; across a three-year cash flow it can cost more than buying the player outright. This is the most dangerous sort of label: it is not wrong in its letters, only in its meaning. And it is the hardest to audit, because there is no line to check it against.

Academies are the same story. Most V-League clubs print “youth development” on the crest, but the path from the academy to the first team is an almost empty column. Count the academy graduates who played 450 minutes or more in a single season and the number is tiny — under one in ten at most big academies. The “academy” label serves the club’s brand very well. It serves a 21-year-old with 90 first-team minutes very badly. At the academies of wealthy clubs, most trainees are stored assets, not utilised manpower.

The contrarian angle: we are labelling by hand too

The easiest way to deal with the dance-clip story is to blame the algorithm. Laugh, delete the record, move on. But in that very moment we repeat the same error manually: one source, one clip, one conclusion.

I have imposed a triple-verification rule on myself for years, and I must say plainly that the rule has a trap. Three sources are not automatically three independent sources. In Vietnam, a clip appears on a big outlet, ten small sites quote it by the afternoon, and by evening you have “three sources” — but all three are mirrors facing each other. To verify, you need three sources with separate paths: someone who filmed it, someone who was present, an original record. If all you have is three copies, you have nothing.

The second trap belongs to age. The older I get, the easier the verification rule turns into delay: wait until it is certain, then never write. So I set a clear boundary for myself: three independent sources and I write; short of that, I write the verifiable part and state exactly what is missing inside the piece itself. That is why this article exists — it records a verified fact (a text with no football content was labelled football) and it states the unverified part (the identity of the person in the background).

The third trap is empty ruthlessness. Saying European football is ruthless or Japanese football is patient is always true and never useful. The Japanese taught me patience; the Belgians taught me ruthlessness — but both only mean something when they are brought down to a specific situation. If I say a centre-back stood in the wrong place in the 78th minute, I must draw where he should have stood and who should have covered. If I cannot draw it, my remark is just another clip nobody rewatched.

The last trap is nostalgia. At my age, my head is full of old matches, empty grounds, afternoons with no ball rolling. When the pitch is empty, I find counter-attacks inside my own head — but inside my head, every counter-attack is beautiful. So in every piece I force myself to cite at least one current data point, with a date and a person accountable for it. Memory is not allowed to be the label on the present.

What to look forward to

Next matchday, try one small exercise. Pick a moment your group chat has already labelled and closed: a missed penalty in the 88th minute, a goal ruled offside, a substitution everyone criticised. Rewatch it from a second camera angle, a little slower. If your conclusion changes, you have just measured the price of a label. If it does not change, you have gained another layer of confirmation — and that is still a profit.

I will still be sitting there every night, keeping the notebook. I am old, but I can see the footstep before the touch. The cleanest data column is still the one I count by hand.