i have a problem where disliked stories (disliked text) are still shown (super dislike also doesn’t help):
I believe you’re looking at the Text view. NewsBlur can only train on the Feed view. If you switch to Feed view, you’ll see that those phrases aren’t found in the text. I’ve added code to helpfully display the training, but it’s not applying because the Text view is fetched at the time of you reading the story.
Good catch, and thanks for confirming it’s the Feed view. I tracked this down.
The issue is that the text classifier was doing substring matching against the raw HTML of the story content, but you trained on the visible rendered text. In this case, the ZEIT feed content has inline HTML tags (, ) that split up the phrase you trained on:
The visible text reads: Fragen, Kritik, Anregungen? Sie erreichen uns unter wasjetzt@zeit.de.
But the raw HTML is: `<em>Fragen, Kritik, Anregungen? Sie erreichen uns unter </em><a href="mailto:wasjetzt@zeit.de"><em>wasjetzt@zeit.de</em></a>.`
Those tags in the middle broke the substring match, so the classifier scored it as 0 (neutral) instead of -1 (dislike).
I’ve pushed a fix that strips HTML tags before matching, so text classifiers now match against the same visible text you see and train on. Just deployed, so it should show as hidden now.
Very good for finding the bug that fast and fixing it. Thanks!
1 Like