Research On Continual Learning Knowledge Based Question Answering Systems

Question answering system has been widely used in daily life and industry,such as intelligent customer service,intelligent speaker and so on.However,most of the current QA systems are static,and once trained and deployed,they can not be updated again.And the corpus of the interaction between users and QA systems can not been fully utilized.Even if the interaction corpus is obtained,due to the catastrophic forgetting characteristic of deep neural network,traning already-trained models on the new data directly will lead to the performance drop of the model on the previous data.Therefore,in this paper,we construct a continual learning framework for question answering systems to process and store the interactive data,and propose a continual leanring relation extraction algorithm,and also propose an algorithm for searching adversarial examples to enhance the model’s robustness.The main works of this paper are as follows:We propose a continual learning framework for knowledge based question answering system.We add the input text classification module,question generation module and other modules related to process user feedback into the framework.Thus the question answering system can inqure users appropriately,process user feedback,construct triples and examples from user answers.Triples are stored in knowledge base and examples are used for training models.The experimental results show that the continual learning framework for question answering systems proposed by this paper can improve the performance using the interactive information,and can learning continuously.In order to obtain the continual learning ability,the question answering system needs not only the continual learning framework,but also the models in the system having the continual learning ability.Because the system will continue to generate new data after its deployment,the models in the systems,such as relation prediction model,need to be trained on new data to adapt to the new data.However,training an already-trained model on new data will lead to catastrophic forgetting problem.To solve this problem,we propose a continual learning relation extraction algorithm based on models’ linear connectivity.When the model is trained on a new task,the algorithm optimizes the parameters to ensure that the loss on the linear path between the final model parameters and the previous model parameters does not increase,so as to ensure that the model will not forget learned tasks while learning the new task,and then realize the continual learning.The proposed algorithm also train model models on replay examples and the current task data alternately to consolidate the model knowledge and alleviate forgetting.Experiments on several datasets including Few Rel show that continual learning relation extraction algorithm proposed by this paper outerperfoms other algorithms and have better continual learning effect.Although the continual learning question answering system can get corpus data continuously after deployment,the data corresponding to each single task is small.In order to solve the problem that a single task in continual learning scenario has few examples and the trained model is not roubst,an adversarial algorithm based on quantum behavior particle swarm optimization is proposed.Compared with other adversarial exmaples search algorithms,the quantum behavior particle swarm optimization algorithm used in this paper has stronger global search ability.The algorithm fisrt find the substitute word based on sememes to form the examples searching space.Then the quantum behavior particle swarm optimization algorithm is modified according to the discreteness of the search space.The modified QPSO algorithm is used to search the adversarial examples.And the randomness is introduced by combining mutation operation.The experiments on several datasets show that the algorithm proposed by this paper is effective in search adversarial examples,and case study intuitively shows that the adversarial examples obtained by the proposed algorithm have better quality.This paper studies the continual learning question answering systems from three aspects:the framework of continual leanring question answering systems,the algorithm of continual learning relation extraction and the model’s robustness.Experiments on the corresponding datasets have proved the effectiveness.The research in this paper is useful for the application of question answering systems in real scene,but more related research is required to promote the further implementation of question answering systems.