Toward a Flexible Framework for Linear Representation Hypothesis Using Maximum Likelihood Estimation