Language modeling for voice recognition by use of knowledge graph

Abstract

The invention relates to language modeling for voice recognition by use of a knowledge graph. One unrestricted embodiment of the invention discloses language model processing on a received speech by use of a combined language model. The combined language model is applied to evaluating a transcription probability associated with the received speech. The combined language model evaluates naming entity data associated with the received speech by use of a position-based language model and an entity relation probability model. In at least one embodiment of the invention, a final probability that one or more transcriptions comprise the naming entity data is generated by evaluating the naming entity data of the received speech by use of at least one entity knowledge graph and query click log data associated with a naming entity of the knowledge graph. A result is output based on the final probability calculated through the combined language model. The output result comprises one or more transcriptions for ranking probabilities of candidate transcriptions on the basis of the combined voice model. Other embodiments are also disclosed.

Description

technical field [0001] The present invention relates to language models and, more particularly, to language modeling for speech recognition using knowledge graphs. Background technique [0002] Statistical language model is a classical model designed for speech recognition to estimate the prior probability of word strings. In speech recognition systems, the applied language models typically work with very large but limited training corpora. Although modern commercial speech recognition systems try to add sufficient and large training corpora as much as possible, due to constraints such as training performance, model size, and costs (financial and resource, etc.) for collecting and maintaining a training corpus containing named entities Considering that it is difficult to update the training corpus to collect all possible entity information. However, named entities are often present in speech recognition systems, especially spoken dialogue systems. Such entities are very l...

AI Technical Summary

Problems solved by technology

Although modern commercial speech recognition systems try to add sufficient and large training corpora as much as possible, due to constraints such as training performance, model size, and costs (financial and resource, etc.) for collecting and maintaining a training corpus containing named entities Considering that it is difficult to update the training corpus to collect all possible entity information

Images