4 ms·
I think the issue is that text doesn't exist in a vacuum, but the corpus that the model is learning from does. A piece of human writing exists for a particular
by vladTheInhaler 6y ago
I think the issue is that text doesn't exist in a vacuum, but the corpus that the model is learning from does. A piece of human writing exists for a particular reason - to persuade, to inform, to ask a question, etc - and its value is judged on its ability to perform that task. But that's not a quality that is evident from the text itself, only from looking at the world outside the text. This suggests to me some limits on this kind of passive self-supervised approach. Perhaps it could be improved by augmenting the text with other forms of data? For instance predicting video from text and vice versa. But I think that to learn a true "theory of mind", it needs to use text like an agent - to influence its environment, not merely predict it.
- dannyw 6y agoIt could also be a result of training data. If every page is weighted equally, you'd expect SEO spam and even autogenerated content to far surpass high quality content in volume. I would like to see a GPT model where training data is weighted by credibility / authority (e.g. using Pagerank).
- vladTheInhaler 6y agoMy understanding is that GPT-2 was actually trained on a dataset that was designed to avoid those pitfalls. They followed all the links posted to Reddit that had more than a couple karma, under the theory that the content was at least slightly interesting to some actual humans, as opposed to a giant blob of search keywords or what have you.