Separate Answer Decoding for Multi-class Question Generation

Question Generation (QG) aims to automati-nerate questions by understanding the semantics of source sentences and target answers. Learning to generate diverse questions for one source sentence with different target answers is important for the QG task. Despite of the success of existing state-of-The-Art approaches, they are designed to merely generate a unique question for a source sentence. The diversity of answers fail to be considered in the research activities. In this paper, we present a novel QG model. It is designed to generate different questions toward a source sentence on the condition that different answers are regarded as the targets. Pointer-Generator Network(PGN) is used as the basic architecture. On the basis, a separate answer encoder is integrated into PGN to regulate the question generating process, which enables the generator to be sensitive to attentive target answers. To ease the reading, we name our model as APGN for short in the following sections of the paper. Experimental results show that APGN outperforms the state-of-The-Art on SQuAD split-l dataset. Besides, it is also proven that our model effectively improves the accuracy of question word prediction, which leads to the generation of appropriate questions.