Paraphrases are useful for statistical machine translation (SMT) and natural language processing tasks. Distributional paraphrase generation is independent of parallel texts and syntactic parses, and hence is suitable also for resource-poor languages, but tends to erroneously rank antonyms, trend-contrasting, and polarity-dissimilar candidates as good paraphrases. We present here a novel method for improving distributional paraphrasing by filtering out such candidates. We evaluate it in simulated low and mid-resourced SMT tasks, translating from English to two quite different languages. We show statistically significant gains in English-to-Chinese translation quality, up to 1 BLEU from non-filtered paraphrase-augmented models (1.6 BLEU from baseline). We also show that yielding gains in translation to Arabic, a morphologically rich language, is not straightforward.