Model weighting, selection and hypotheses combination for automatic speech recognition and machine translation
AssigneeInternational Business Machines Corporation
InventorsSuleyman S. Kozat, Ruhi Sarikaya
Life of this patent
Estimating the term…
Abstract
A translation method and system include a recognition engine having a plurality of models each being employed to decode a same utterance to provide an output. A model combiner is configured to assign probabilities to each model output and configured to assign weights to the outputs of the plurality of models based on the probabilities to provide a best performing model for the context of the utterance.
Claims20
A translation system comprising a computer useable storage medium including a computer readable program, wherein the computer readable program when executed on a computer, comprises:
- a recognition engine having a plurality of models each being employed to decode a same utterance to provide an output;
- and
- a model combiner, including a probability module configured to compute a probability for each output based on an exponentiated and normalized loss accumulated over time, configured to assign probabilities to each model output and configured to assign weights to the outputs of the plurality of models based on the probabilities to provide a best performing model for the context of the utterance.
A translation method, comprising:
- recognizing an utterance using a plurality of models each being employed to decode the utterance to provide an output;
- assigning probabilities to the outputs based on an exponentiated and normalized loss accumulated over time;
- weighting the models based upon the assigned probabilities;
- using a processor to predict a best performing model based on the weighting of the outputs;
- and
- determining a combination hypothesis for the best performing model and applying at least one of past performance and user input as feedback to adjust the weights for translating a next utterance.
A computer program product comprising a computer useable storage medium including a computer readable program, wherein the computer readable program when executed on a computer causes the computer to perform the steps of:
- recognizing an utterance using a plurality of models each being employed to decode the utterance to provide an output;
- assigning probabilities to the outputs based on an exponentiated and normalized loss accumulated over time;
- weighting the models based upon the assigned probabilities;
- predicting a best performing model based on the weighting of the outputs;
- and
- determining a combination hypothesis for the best performing model and applying at least one of past performance and user input as feedback to adjust the weights for translating a next utterance.
Dependencies are inferred from the claim text.