What's the potential of capsule network in text-mining?

Question

Capsule networks seem to have been made to try and replace CNNs. It's been experimented with in the past to use CNNs on text data by doing word embeddings (word2vec, GloVe, etc.) and training it to classify text. People have also thrown things like LSTMs to take in the CNN features and give results like that.

With capsule networks looking to find characteristics and properties of the data it sees, is it possible to implement capsule networks to extract features and properties out of text data?

If it is possible, would there be any added benefit to use it over CNNs?

Answer 1

Intuitively capsules seem to be suited better for multidimensional data. Location on a line is trivial in some sense, and text is essentially unidirectional: left to right or right to left. While in multidimensional case it's much more complicated, especially when you work with two dimensional projections where you get overlapping objects etc.

I see one places in text processing for capsules. When you get into semantics, such as dialogues and overlapping story lines. So, one guy's speech starts somewhere, and ends somewhere, and it can overlap with other speakers. So, the capsule could be locating these parts.