In this paper we face the problem of Community Question Answering for the Arabic language. In this setting, a member of the community posts an initial query expressed in Natural Language. Other participants post their own interventions: answers, comments, additional questions, etc. contributing to building a rather tangled thread of nodes containing Natural Language short texts. The task consists in answering the initial query using the thread as the space of possible answers. The task can be approached as a classification, a regression or a ranking problem. In our case we select the set of possible candidates, we assign a relevance score to each candidate and we rank them accordingly.We propose a bunch of unsupervised models and show that a model based on Latent Semantic Indexing approach outperforms state of the art models for this task. We also use transfer learning to power the embeddings layer of various deep learning models and prove that the pairwise approaches outperform their pointwise counterparts. All the proposed models have been evaluated on Semeval 2017 Task 3 Subtask D: Arabic Community Question Answering and achieve state of the art or near performance.