A film item labelled football: a system error and a lesson for Vietnamese sports data

In Nagoya, I usually open the data board before I put the kettle on. That morning the first line in the tagging file was a short item: nineteen information points, and in the domain column two words were written very clearly — football. I read all nineteen. No club. No player. No match, no contract, no possession figure, not a line about tactics or financial structure. What appeared were the names of four filmmakers — Tom Holland, Sabrina Carpenter, Paul King and Kathleen Riley — and a biopic project about Fred Astaire.

In twenty-seven years on the job, I have made it a habit to check every data label by hand before it reaches any table of statistics. One wrong label does not ruin one article. It ruins a whole dataset, and that dataset will quietly decide the headlines of thirty articles to come.

The mislabelled item has very specific content, and that content belongs to cinema. Sabrina Carpenter was reported to star alongside Tom Holland in a biopic about Fred Astaire. Paul King directs and co-writes with Steven Levenson. The screenplay is based on Kathleen Riley's book The Astaires: Fred & Adele, with Riley serving as script and story consultant. The project was first announced nearly five years ago. The background recalls the era when Fred and Adele Astaire were major stars on Broadway and London's West End, and Fred's celebrated dance partnership with Ginger Rogers.

What stands out is the sourcing layer. Of nineteen information points, only six are direct quotes from Paul King; the rest carry no verified source. There is no studio statement, no agent confirmation, no Hollywood trade outlet cited. The outlet is an English-language daily in Pakistan, a redistribution node for international wire content. The story had passed through at least two relays before it reached the tagging system.

Yet the domain column still read football. The failure is not in the summarising layer: the note correctly identifies the item as a news report with an objective stance. The failure sits in exactly one layer — the keyword dictionary. One person's name, one phrase overlapping a sports category, and the door opens.

This is where I should explain why an apparently harmless error bothers me. Dirty data does not do damage by creating false information; it does damage by diluting the weight of true information. When a film item is written into a football dataset, it does not stay put. It gets counted into media-heat indices, linked into entity graphs, used as a training sample for sentiment classification in later runs. Three years from now, when someone asks why a league's attention index spiked in a week with no fixtures, the answer may lie in a film that has not started shooting.

The item itself also carries useful clues for verification work. Six information points duplicate one another: points fourteen through nineteen are three remarks by a single person, reworded into six lines. In a table, those six lines will be counted by a model as six independent signals, and a single statement suddenly carries six times the weight. In Japan I have seen the same thing with transfer news: one player reported by three outlets from exactly one source, and by the next morning all three have become multiple confirmations. The agent only had to make one phone call.

My trade taught me to keep the order of verification strict: official statement, club confirmation, training-ground footage, and only then the agent's word. In that film item, director Paul King's remarks are the only attributed source, while all the casting information sits at reported-but-unconfirmed. That is the same classification I use for transfer news, and it works in any field.

There is one more comparison worth keeping in mind. The information half-life of a casting announcement is short, measured in days to weeks, and transfer news behaves the same way. But what lasts in both cases sits at the edge: a small detail nobody races to publish. For that film, it is the director's remark that no footage of Adele Astaire's dancing survives. For football, it is a midfielder's touch count in the eightieth minute, when the result is already settled. People remember the name of the scorer; I remember the one who put the ball in the right place.

The easiest thing now is to blame the tagging machine. I do not think that is the right diagnosis. An algorithm only repeats exactly what the people who trained it were too lazy to check. If the process never required anyone to open the item and read it before applying a label, then the mislabel is a symptom. The disease is a habit of trusting volume: the more rows of data, the safer we feel, without having checked a single row.

I once sat in a press room in Russia where hundreds of reporters waited for one name. A player who ran 11.8 kilometres and made the most ball recoveries in his team was barely asked about. Attention does not distribute by information value; it distributes by noise. The film item is the same: the part that trended was two young actors' names, the part worth remembering was a gap in an archive.

And here is where I want to be blunt about my own side: agents are the largest hidden cost of the transfer market, not because they are bad people, but because their noise distorts the measure. A casting announcement timed for release follows exactly that logic. Read the timing of an announcement as promotional data, not as evidence. Stability never dazzles, but it keeps things from breaking apart.

In Japan, clubs publish injury updates on a fixed template, weekly, without embellishment. That dullness is a form of infrastructure. For Vietnamese sports data, the task is to move the question from tagging faster to who is accountable when a wrong label slips through. Some matches nobody watches, but the people inside them still have to play them out in full. When the ball stops rolling, I start looking more closely.