An Adversarial Multi-Task Training Method for Spoken Language Understanding

Abstract

The invention discloses an adversarial multi-task training method for spoken language comprehension, including sampling from unlabeled data and labeled data to train and update a language model and a shared space, and labeling the first public feature obtained in the shared space as a language model The task is to train and update the task discriminator and the shared space; take samples from the labeled data to train and update the spoken language understanding model and the shared space, and mark the second public feature obtained by the shared space as a spoken language understanding model task to train and update the described Task discriminator and the shared space. The adversarial multi-task training method for spoken language understanding in the embodiment of the present invention can train the spoken language understanding model based on unlabeled data and labeled data at the same time, thereby avoiding the heavy dependence of the traditional method for training the spoken language understanding model on labeled data , reducing the cost overhead caused by a large number of labeled data.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to an adversarial multi-task training method for spoken language comprehension. Background technique [0002] The spoken language understanding (SLU, spoken language understanding) module is a key component of the target-oriented spoken language dialogue system (SDS, spoken dialogue system), which parses the user's utterance into corresponding semantic concepts. For example, the sentence "show me flights from Boston to New York" could be parsed as (departure city=Boston, arrival city=New York). Generally, it is considered as a slot filling task, assigning each word in an utterance a predefined semantic slot label. [0003] Recent research on statistical slot filling in SLU has focused on recurrent neural networks (RNNs) and their extensions, such as long and short memory networks (LSTMs), codec models, etc. These traditional methods require a large amount of labele...

AI Technical Summary

Problems solved by technology

However, since data annotation is labor-intensive and time-consuming, it is difficult to obtain sufficient in-domain labeled data for training

Images