An agent who adaptively asks the user questions to seek information is a crucial element in designing a real-world artificial intelligence agent. In particular, goal-oriented visual dialogue, which locates an object of interest from a group of visually presented objects by asking verbal questions, must be able to efficiently narrow down and identify objects through question generation. Several models based on GuessWhat?! and CLEVR Ask have been published, most of which leverage reinforcement learning to maximize the success rate of the task. However, existing models take a policy of asking questions up to a predefined limit, resulting in the generation of redundant questions. Moreover, the generated questions often refer only to a limited number of objects, which prevents efficient narrowing down and the identification of a wide range of attributes. This paper proposes Two-Stream Splitter (TSS) for redundant question reduction and efficient question generation. TSS utilizes a self-attention structure in the processing of image features and location features of objects to enable efficient narrowing down of candidate objects by combining the information content of both. Experimental results on the CLEVR Ask dataset show that the proposed method reduces redundant questions and enables efficient interaction compared to previous models.