← Research Journal

The Word That Stood Out

In one project where I work on content, the client chats in Telegram carry everything at once: edits to articles, questions about the client's website, access to accounts, call times. I had a filter built that reads these chats for me and passes on only what concerns the texts.

On 13 October a client wrote 27 short messages in about two hours. The question was simple: when will the articles appear on our own website? That is a publishing task. It belongs to the people who run the site.

Two readers got it wrong in the same way. I was tagged in the thread as the person who writes the articles. My filter sent me 17 of the 27 messages, because most of them contained the word "articles". In a busy chat, a message goes where its most visible word points.

The fault is not with the person who tagged me. It is with the rules of the game, which nobody had explained: who answers for the texts and who answers for the site. There is no written role description yet, which is normal for a startup, and explaining the rules is the job of whoever hires the team. A filter can route by roles only once the roles are written down.

I answered my part: the texts were ready and approved, and I sent the link to them again. The publishing question stayed with the people whose job it is.

Another stop word will not fix the filter. "Articles" means a text to edit in one message and a page to publish in the next. What tells them apart is everything around the word:

  • the whole thread, not the last five messages: a reply like "Only on our website" means nothing on its own;
  • who is speaking: the client, a manager, a technician, the editor;
  • a run of short messages read as one, because that is how people type in chats;
  • a memory of open questions. This client had first asked about publishing more than two weeks earlier, and the question was never answered.

This is what I argue about search engines and language models too. A keyword is a guess about meaning. A person in a hurry makes that guess, and so does a script. The cure is the same for both: read the context.

← All entries