Data augmentation method and system for spoken language semantic understanding

Abstract

An embodiment of the present invention provides a data enhancement method for semantic understanding of spoken language. The method includes: defining an atomic template matching table, in which each atomic template is a natural language description corresponding to a structured basic granularity label; determining a semantic representation according to a given dialog action, and decomposing the semantic representation of the dialog action According to the atomic template matching table, find the structured basic granularity label corresponding to the unit semantic label, and then determine the atomic template matching the basic granularity label, and use the atomic template to convert the given dialogue action into an atomic sample ; using a neural network-based sentence generation model to combine atomic sample sets into natural sentences. The embodiment of the present invention also provides a data enhancement system for semantic understanding of spoken language. The embodiments of the present invention require data-enhanced dialogue action input to generate brand new tags, and the generated new sentences are more in line with natural sentences.

Description

technical field [0001] The invention relates to the field of speech recognition, in particular to a data enhancement method and system for spoken language semantic understanding. Background technique [0002] SLU (Spoken Language Understanding, Spoken Language Understanding) is a key component of a spoken dialogue system, which parses user utterances into corresponding semantic representations in a narrow domain. Typical semantic representations of SLU can be semantic frames or dialogue behaviors. [0003] Deep learning has achieved great success in the field of SLU, but it requires a large amount of labeled data, which limits the scalability of SLU models. Despite the progress and huge research activities in semi-supervised learning and domain adaptation, deep SLU models A large amount of labeled data is still required for training. In order to obtain a large amount of data, the following techniques are usually used, including: zero-shot learning (zero-shot learning), inpu...

AI Technical Summary

Problems solved by technology

[0006] In order to at least solve the problems in the prior art that label collection is required, new labels cannot be generated, the calculation process has high space-time complexity, and the overall performance of data-augmented sentences is poor.

Images