Hacker News Reader

Even if totally 'accurate' to the dataset, there's the issue that a lot of 'toxicity' and offence is completely culturally based. Take 'cunt' for example. Used in casual, informal conversation here, but deemed incredibly offensive by some Americans. Or, more broadly, 'thumbs up' could mean "Good job" or "Stick it up your arse".

View on HN

Exactly. This is why it's important not just to have language skills when creating these kinds of datasets, but also cultural knowledge and context.

For example, to pick a bit on the Google Emotions dataset again, it's difficult to label this message...

“Also Republicanism is a belief system. It’s taught and handed down like religion. Conservative talk radio is its evangelism.”

...unless you're familiar with US politics. Hence why it was labeled as APPROVAL by the non-US annotators, even though it's criticizing Republicans.

Reply on news.ycombinator.com