Utilizing meta-learning for cross-lingual transfer in few-shot stance detection task

Abstract

With the popularity of the mobile Internet and the rapid development of social networks we can now receive more and more information of which controversial information is often the focus of attention The automatic extraction of semantic information from natural language texts is an important research issue in many practical applications Therefore the stance of controversial information is also an important part of natural language processing research Researchers have achieved excellent results in many stance detections tasks But in the actual application it is difficult for us to have such a large amount of training data Therefore this article focuses on solving the problem of insufficient data in certain stance detection tasks This paper proposes a model based on meta-learning algorithm that uses data from other tasks to enhance the training effect of tasks with insufficient data In the process of meta-training this paper uses a memory network and a cross-lingual pre-training model to enable the model to use different languages different sources of stance detection tasks This paper also explores the effects of different optimization methods and parameter settings of different meta-learning model Therefore the main contributions of this paper are: (1) Integrate the meta-learning algorithm into the position inspection task to improve the effect of the stance detection task with insufficient data (2) We use cross-lingual pre-training model and memory network architecture in the meta-training process and Increased the applicability of training materials

Date of Award	2020
Original language	English
Supervisor	Hung-Yu Kao (Supervisor)