next up previous
Next: Evaluation Up: Robust parsing of word-graphs Previous: Comparison with Dijkstra's algorithm

Best-first methods

Rather than integrating parsing and disambiguation of the word-graph as a single procedure, as we proposed above, it is also possible to try to disambiguate the word-graph first, and then use the parser to parse the resulting path in the word-graph.

We have implemented two versions of this approach. Both versions use the search algorithm discussed above, by applying a method which takes into account the acoustic scores and Ngram scores. One version uses N=2, the other version uses N=3. In section 4 we refer to these two methods as best_1_bigram and best_1_trigram respectively.

We have experimented with such methods in order to evaluate the contribution of grammatical analysis to speech recognition. If, for instance, the integrated method nlp_speech_trigram performs significantly better than best_1_trigram then we can conclude that grammatical analysis improves speech recognition. The results below, however, do not permit this conclusion.