Actually, the IP network works over MPLS. The idea of MPLS is that a series of labels can be applied to a given packet (or frame) that can subsequently be used to switch it through a network. In the case of an L3VPN that means that rather than the traditional mechanism of looking at the destination IP address, routing devices look at one or more previously applied labels to make forwarding decisions.
The key point in the above is that the actual -contents- of the packet aren't actually considered. Once a given packet is labeled the intervening devices simply forward it based on whatever LSP has been signaled. In the case of an L3VPN, the packet is a fully formed IP packet. In the case of an L2VPN a frame from a particular interface has a label added and is forwarded. This might be a full Ethernet frame (with or without an 802.1q header), an HDLC frame from a serial link, one or more cells from an ATM PVC, etc..
One of the contrasts between L2 and L3 VPN is the mechanism used to signal and set up the overlay network. L3VPN (RFC2547bis) extends the BGP protocol to allow PE's to signal which routes are available within which VPN's. There are more possible ways to put together a layer 2 network (i.e. straight point-to-point links, multipoint, translational, etc) and there are also more mechanisms in use in the industry used to signal these various topologies.
As for links - google is obviously your friend, but if there's some specific aspect of L2VPN design, architecture or setup it would make things easier for us to direct you.