Assessing statistical significance of overrepresentation of exceptional words is becoming an important task in computational biology. We show on two problems how large deviation methodology applies. First, when some oligomer $\path$ occurs more often than expected, e.g. may be overrepresented, large deviations allow for a very efficient computation of the so-called $p$-value. The second problem we address is the possible changes in the oligomers distribution induced by the overrepresentation of some pattern. Discarding this noise allows for the detection of weaker signals. Related algorithmic and complexity issues are discussed and compared to previous results. The approach is illustrated with two typical examples of applications on biological data.