This paper adresses the problem of modelling paraphrases in a deep linguistic processing framework where the meaning construction component is based on an LFG grammar. We present a syntax-based approach to paraphrase extraction that operates on shallow dependency analyses in a parallel corpus. By means of an XLE-based conversion routine, we generate transfer rules for the automatically acquired semantic correspondences. These rules can be used as an additional component in the rule-based process of meaning construction which will augment the meaning representation with entailments that hold for complex phrasal units.