This is very similar to the thin glass problem.
It can be fixed by giving it a thickness, (i.e. Solidify Modifier, as already sugested), which in some cases raises some topology problems, besides the doubled mesh.
Or by faking the surface thickness by not using refraction at all.
This last approach raises another problem: IOR inverts automatically on backside faces (which in this case should be treaten as frontside)... Luckly we can solve that.
When a light ray hits a surface with a refraction shader, the ray will be bent dued to the difference of the light speed from one side of the surface to the other. This speed difference is represented by the IOR and ruled by the Snell's law. When the ray is exiting that medium, the difference of the light speed gets inverted, and so the IOR. Cycles does this automatically, and treats objects as solid (not hollow).
If we want hollow objects, we need to model both outer and inner surfaces of that object, and most of the times is exactly what we do. But there's a nice particularity from this phenomena: if both surfaces are parallel (at least where the ray enters and exits), the outgoing direction will be the same as the incoming direction; and thought the complete light path is not a straight line, the difference is quite small, sometimes smaller than the pixel size, and we can discard all that refraction calculation. A very good example of this are glass windows. In 3D, we call this the 'Thin Wall' or 'Thin Glass' effect, and some renders even have a shader just for this purpose.
In Cycles we can do something similar (treat the surface as a thin wall and discard refractions), but we need to specifically tell the engine that the IOR from a face when viewed from the objects inside, is the same as from the outside.
The nodes in the 'Fixed Backside IOR' frame do exactly that: If 'Backside' is 0 (False), we get (1-0)*IOR+(0*1/IOR)=IOR; if it's 1 (True) then (1-1)*IOR+(1*1/IOR)=1/IOR. Cycles will invert this backface value internally and in the end it stays equal to the initial IOR.
After having the IOR fixed, we can deal with the Reflective/Transmissive sides of the Fresnel formula. I didn't use a complete transparent shader for the transmissive part, because the greater the angle between the incoming ray and the surface normal, the greater the distortion and absorbtion on the light on that path. I used the translucent shader to mimic that effect, but it's just an approximation.