How good is a span of terms?: exploiting proximity to improve web retrieval