https://www.semanticscholar.org/reader/a97da58551262ce74a06367f996f806faf04392c

The paper claims to cluster encoded representations and try to map the concepts.

So bascally they try to get the representations. The represnetaions are then clustered. Each cluster is an encoded concept. Then they align those clusters with human concepts.

so can i take that and apply somewhere else ?

So what are my take aways from this ? i didnt find it interesting. They used a technique to bring out the activation

Summary: So what the paper is essentially doing is getting the activations of each layer of Bert from thousands of sentences. Then clustering the activations at each layer across thousands of words. So each layer basically enriches the previous tokens. we are trying to cluster concepts from those layers and then checking the alignment. In Bert layer 1, each word would be updated based on its other word attention, so what we essentially get is a better updated word vector representation. Then they cluster all the words. Then theyu look at each cluster and see the properties of each cluster. if that cluster all ends with ‘er’ then its a suffix. if it ends with ‘un’ then it encoded suffix. The limitations the paper mentions is that there is a bug chunk which remains unexplained. And these are encoder models.