The Web as a Baseline: Evaluating the Performance of Unsupervised Web-based Models for a Range of NLP Tasks