Research On Few-shot Joint Learning For Dialogue Language Understanding

With the development of artificial intelligence,many smart products have come out,greatly facilitating human life in all aspects of life.The key technologies in these smart products all involve smart dialog systems,and the dialog semantic understanding module is particularly important,which is related to whether smart products can accurately understand human needs.The module has two main subtasks,intent recognition and slot filling,which have a strong correlation and can be jointly modeled to improve performance.However,most of the current deep learning algorithms are inseparable from a large amount of labeled data.When only a small amount of data can be obtained,the performance of the traditional deep learning model will drop sharply.Therefore,t his article aims to explore a few-shot joint dialogue semantic understanding model in order to improve the performance of the task of dialogue semantic understanding in low-resource scenarios.We explored three research directions.First,we explored the d irection of applying the existing joint learning algorithm model to the framework of the few-shot learning algorithm,and found that the simple hard-share joint learning architecture under the framework of the few-shot algorithm has the problem of negative transfer.After the adjustment,an explicit information interaction module between tasks was added to form an effective few-shot joint algorithm model.Secondly,we explore a set of unique few-shot joint algorithm models for the few-shot learning algorithm framework.We use the classic prototype network in the few-shot learning field to design the metric space interaction module and standardize the distribution of prototype points in the metric space.Effectively promoted the information interaction between the two tasks and fully modeled the relevance of the two subtasks in dialogue semantic understanding.Finally,we explored a new training optimization strategy algorithm based on the few-shot joint learning model obtained in research direction1,so that the two tasks can transmit higher confidence information when interacting,so as to better model the correlation between the two tasks.It is worth mentioning that we conducted experiments on datasets in Chinese and English,and constructed the original dataset into a few-shot data set through the miniincluding algorithm.Due to the lack of Chinese dialogue semantic understanding task dataset,we constructed a Chinese dataset.On the Chinese and English dataset,we designed a series of experiments,and the results proved the validity and superiority of the model explored in the above three research directions.