As a key task of question answering, question retrieval has attracted much attention from the communities of academia and industry. Previous solutions mainly focus on the translation model, topic model, and deep learning techniques. Distinct from the previous solutions, we propose to construct fine-grained semantic representations of a question by a learned importance score assigned to each keyword, so that we can achieve a fine-grained question matching solution with these semantic representations of different lengths. Accordingly, we propose a multi-view semantic matching model by reusing the important keywords in multiple semantic representations. As a key of constructing fine-grained semantic representations, we are the first to use a cross-task weakly supervised extraction model that applies question-question labelled signals to supervise the keyword extraction process (i.e. to learn the keyword importance). The extraction model integrates the deep semantic representation and lexical matching information with statistical features to estimate the importance of keywords. We conduct extensive experiments on three public datasets and the experimental results show that our proposed model significantly outperforms the state-of-the-art solutions.