Exploiting Qualitative Information from Automatic Word Alignment for Cross-lingual NLP Tasks