We develop a finite-state transducer for translating unranked trees into general graphs. This work is motivated by recent progress in semantic parsing for natural language, where sentences are first mapped into tree-shaped syntactic representations, and then these trees are translated into graph semantic representations. We investigate formal properties of our tree-to-graph transducers and develop a polynomial time algorithm for translating a weighted language of input trees into a packed representation, from which best-score graphs can be efficiently recovered.