Hey,

Recently I have been working a lot with semantic sentence encoding. For example for some semantic search use cases or semantic clustering.

As a sort of demo and way to explore the potential and limitations of semantic encoding, I thought it would be fun to build this small side project together with a colleague.

It's basically a HN clone based of VueJS' HN clone with the small twist of having the top 3 semantically most similar together with the current HN stories.

I have been using this for some time for myself now and I find sometimes some fun stories come up. It doesn't always work so great, but see for yourself.

I added a feature for ranking how similar you perceive the story to be. If enough people make use of the ranking feature it could be possible that a dataset of similar HN-stories title pairs would come together (I'm thinking of sth. like the GLUE STS-B dataset).

If that was the case I would, of course, make it available publicly.

Thanks for reading and please let me know your thoughts :)

/Philipp

Good idea Philipp. The results are not as great, BERT based similarity is still very far off human performance.

I like the idea of the time machine and it’s very interesting to see how history plays out in HN stories. Looking forward to a future time machine with updated technology.