Structured lexical similarity via convolution kernels on dependency trees