On parsing direction and tree search in the LITHAN speech understanding system

Abstract

We compared two parsing directions in a search space, one which starts the parsing from left to right, and another which first identifies the predicate and then starts the parsing from left to right. The latter method was found to be superior. The predicate can be easily found in Japanese, since it is located at the end of a sentence. For this reason, LITHAN adopts the latter method. LITHAN searches a few best paths with backtracking in parallel using the following algorithm (α-β-γ search). Expand the best a nodes at one time. If the number of new nodes which have been expanded by a selected node exceeds a preset threshold (β), only the best β nodes are kept and others are pruned. If the total number of newly generated word sequences and nonexpanded word sequences exceeds a preset threshold (γ) only the best γ word sequences are kept, and others are removed. Repeat these consecutive processes until a complete sentence is generated. We evaluated this method by simulations and experiments, and obtained good performances both in the search efficiency and in the recognition rate.