An N-Best Representation for Bidirectional Parsing Strategy

In speech understanding systems, the interface between acoustic and linguistic modules is often represented by the N best sequences that match the input signal. They compose a set that will be linguistically analyzed in order to find the interpretation of the input. An appropriate representation of the N-Best could make linguistic processing more efficient. Here a representation based on a context-free model is proposed that is obtained by an algorithm inherited by the data compression field. This algorithm is based on the subword tree of the concatenation of the N best sequences. The proposed representation seems particularly appropriate when coupled with a bidirectional parser and some experiments demonstrate that tha approach is worth pursuing. Such experiments focus on the comparison between the proposed representation and a sequential processing of the N hypotheses given by the acoustic module. The comparison takes into consideration the efficiency attained in the two cases, in terms of (partial) analyses constructed by the linguistic module. The obtained results are presented and discussed.