Extracting star formation histories from spectra is a process plagued by numerous degeneracies among the parameters that contribute to the definition of the underlying stellar populations. Traditional approaches to overcome such degeneracies involve carefully defined line strength or spectral fitting procedures. However, all these methods rely on comparisons with population synthesis models. This paper illustrates alternative approaches based on the statistical properties of the information that can be extracted from uniformly selected samples of observed spectra, without any prior reference to modelling. Such methods are more useful with large datasets, such as surveys, where the information from thousands of spectra can be exploited to classify galaxies. An illustrative example is presented on the classification of early-type galaxies with optical spectra from the Sloan Digital Sky Survey.