Natural Answer Generation With Attention Over Multi-instances

Knowledge Base Question Answering(KBQA)is a method that could get answers for given natural language questions and knowledge base through semantic understanding and analysis of the questions,and query and reasoning the knowledge base.Due to its diverse application fields,it is widely used.Our daily life.The traditional knowledge base question answering system usually consists of several consecutive modules.The pipeline system composed of the question semantic representation,semantic matching,query and reasoning can get the answer to the question,but the system is very sensitive to error propagation and the universality is also poor.Every time it is promoted to a new field,it requires a lot of manual r-econstruction.Therefore,the end-to-end Natural Answer Generation(NAG)came into being,which is a hot topic in KBQA.The natural answer generation model is an end-to-end generation model based on deep learning,which can generate natural language answers using structured knowledge bases.However,the current natural question-and-answer model assumes that there is only one answer to a question.In fact,real-world community usually contain multiple answers to users' questions,and many of the answers vary in quality.In response to this problem,this paper proposes a novel approach which organizes the answers into question bags using multi-instance learning principle to dynamically reduce the weights of noisy instances.In particular,four kinds of attention-based algorithms are proposed:instance selection,instance weighting(which weights the instances in a bag with pre-defined weights),selective attention and self-attention mechanisms(which could dynamically weight the instances during the training process)to de-emphasize noisy instances and emphasize instances that contain more information.After that,we also transform and use the training methods of curriculum learning on our task,so that the model first trains with single instance bags and gradually adds multiple instance bags to further improve the performance of our model.As for experiment,we experiment our methods with a public open domain dataset(CQA).Experiments demonstrate that Selective-ATT outperforms the state-of-the-art by 10.69%in the entity accuracy,9.73%in the BLEU score,and 1.21%in the Rouge score on the CQA dataset.