I don't know about a paper which describes just the very basics but basically what you want to do is:
1. Render your Scene without the transparent object to a RT
2. Render you transparent object and project the scene's RT on it + apply distortions based on the object's normals

This is of course a crude approximation and does not work well with multiple overlapping transparent objects (+ several other limitations) but it's still pretty popular in the real-time realm.
It also looks a lot better if you vary the distortion strength for the different color channels which gives you chromatic aberration (stronger offset for blue, weaker offset for red, green in between).
Depending on your use case you could also use a cubemap which gives you more information of the environment. There should be a shader on the wiki which does this.