Two-step TAG parsing revisited

Based on the work in (Poller, 1994) and a minor assumption about a normal form for TAGs, we present a highly simplified version of the twostep parsing approach for TAGs which allows for a much easier analysis of run-time and space complexity. It also snggests how restrictions on the grammars might result in improvements in run-time complexity. The main advantage of a two-step parsing system shows in practical applications like Verbmobil (Bub et al., 1997) where the parser must look at multiple hypotheses supplied by a speech recognizer (encoded in a word hypotheses lattice) and filter out illicit hypotheses as early as possible. The first (context-free) step of our parser filters out some illicit hypotheses fast (O(n3 )); the constructed parsing matrix is then reused for the second step, the complete (O(n6 )) TAG parse.