Lexical-functional Grammar: a Formal System for Grammatical Representation. Robust Speech Repair

After locating and highlighting erroneous sections in the recognizer hypothesis, misrecognitions are corrected. The spoken hypothesis correction method uses N-Best lists for both the initial utterance and the respo-ken section. The N-Best list for the highlighted section of the initial utterance is rescored using scores from decoding the secondary utterance. Depending on the quality of the N-Best lists, most misrecognitions can be corrected. The spelling hypothesis correction method requires the user to spell the highlighted erroneous section. A spelling recognizer decodes the spelled sequence of letters. By means of a language model we restrict the sequence of letters to alternatives found among the N-Best from the located section. To date, we have evaluated our methods over sentences from the Resource Management task. Table 7 shows the improvements in sentence accuracy, based on recordings from one speaker of the February and October 1989 test data. We selected a subset of erroneous utterances; therefore the accuracy of the baseline system is signicantly lower than the 94% performance our system achieves on the whole test set. The results indicate that repeating or spelling a misrecognized subsection of an utterance can be an eective way to repair recognition utterances. Conclusions We described JANUS-2, our multilingual spoken language translation system. We introduced dierent learning approaches that reduce hand-tuning eorts, yield better word accuracy and even accelerate recognition speed. All of these techniques can be applied in several languages and help making signicant advances towards building a multilingual translation system for spontaneous human-to-human dialogs. Beyond recognition of spontaneous speech JANUS-2 provides a framework for investigating important areas like robust parsing, machine translation of spoken language and developing methods to recover from recognition and parsing errors.