Log-linear models and latent semantic indexing applied to mwe identification

A short introduction characterizes the task of identification of multiword expressions and their idiosyncratic properties. Then, this document gives a detailed description of loglinear models and latent semantic analysis. The description enumerates components of the models, estimation techniques for the model parameters and addresses the interpretation of the models and their evaluation. We also briefly report on how the models have been earlier used in identification of multiword expressions. Furthermore, we describe how these models can be used in R statistics package. A small case study reports on preliminary experiments on quantifying two linguistic attributes. We explore to what extent these attributes identify mwes and whether a statistical dependence between them can be captured by a loglinear model.