Parsing word graphs using a linguistic grammar and a statistical language model

This paper presents an efficient parsing scheme for word graphs. It combines symbolic information from a linguistic grammar and stochastic information from a statistical language model to find the correct interpretation. A two pass search through the word graph is performed. First a Viterbi-like backward pass computes the exact scores of optimal continuations of partial sentence hypotheses. Then a forward A* tree search uses this information to find the best grammatically correct sentence hypothesis. The parsing algorithm of Tomita is used to ensure that partial sentence hypotheses are grammatically viable. The proposed parsing scheme was successfully tested on word graphs from the German ASL benchmark test. The results indicate that the combination of linguistic and statistical knowledge can considerably improve the recognition accuracy of a speech understanding system