Arabic morphology is complex, partly because of its richness, and partly because of common irregular word forms, such as broken plurals (which resemble singular nouns), and nouns with irregular gender (feminine nouns that look masculine and vice versa). In addition, Arabic morphosyntactic agreement interacts with the lexical semantic feature of rationality, which has no morphological realization. In this paper, we present a series of experiments on the automatic prediction of the latent linguistic features of functional gender and number, and rationality in Arabic. We compare two techniques, using simple maximum likelihood (MLE) with back-off and a support vector machine based sequence tagger (Yamcha). We study a number of orthographic, morphological and syntactic learning features. Our results show that the MLE technique is preferred for words seen in the training data, while the Yamcha technique is optimal for unseen words, which are our real target. Furthermore, we show that for unseen words, morphological features help beyond orthographic features and that syntactic features help even more. A combination of the two techniques improves overall performance even further.