This Python script contains several common pre-processing in NLP such as PPMI computation, SVD-based dimensionality reduction, and PLSR-based distribution prediction.

The following packages are required.

• Python 2.7 (not tested with Python 3)
• numpy
• scipy
• sparsesvd
• sklearn
• svmlight-loader

There is no specific installation for svdmi. Once you have all the dependencies installed, you can run svmi as described in the usage section.

Positive Pointwise Mutual Information

$ python svdmi.py -m PPMI -i raw_co-occurrences_matrix_file_name -o ppmi_matrix_file_name

Singular Value Decomposition-based dimensionality reduction (SVD1) and matrix smoothing (SVD2).

For SVD1 mode

$ python svdmi.py -m SVD1 -i matrix_file_name -o dimensionality_reduced_matrix_file_name -n svd_dimensions -p power_to_raise_singular_values

For SVD2 mode

$ python svdmi.py -m SVD2 -i raw_co-occurrences_matrix_file_name -o smoothed_matrix_file_name -n svd_dimensions -p power_to_raise_singular_values

Use -v option to print the reproduction error (Frobenious norm)

Partial Least Square Regression-based distribution prediction.

Training a PLSR model

$ python svdmi.py -m  -m PLSR.train -x x_matrix_file_name -y y_matrix_file_name -n PLSR_components -i model_file_name

Use -v option to print the reproduction error (Frobenious norm)

Predicting using the trained PLSR model.

$ python svdmi.py -m PLSR.pred -x x_matrix_file_name -y predicted_y_matrix_file_name -i model_file_name

Simple BSD

Danushka Bollegala