pyHCRF

A hidden (state) conditional random field (HCRF) implementation written in Python and Cython.

This package aims to implement the HCRF model with a sklearn type interface. The model classifies sequences according to a latent state sequence. This package provides methods to learn parameters from example sequences and to score new sequences. See the paper by Wang et al and the report Conditional Random Fields for Noisy text normalisation by Dirko Coetsee.

Example

X = [array([[ 1. , -0.82683403,  2.48881337],
            [ 1. , -1.07491808,  1.55848197],
            [ 1. ,  6.7814359 ,  4.01074595]]),
     array([[ 1. , -3.01165932, -2.15972362],
            [ 1. , -3.41449473, -2.2668825 ]]),
     array([[ 1. , -2.64921323, -1.20159641],
            [ 1. ,  0.31139394,  1.58841159]]),
     array([[ 1. ,  5.85226017,  2.43317499],
            [ 1. , -1.57598266, -2.07585778]]),
     array([[ 1. , -0.32999744, -2.70695361],
            [ 1. ,  0.44311988,  0.36400733]]),
     array([[ 1. , -0.05301562,  3.95424435],
            [ 1. ,  3.04540498, -3.25040276]]),
     array([[ 1. , -4.29117715,  0.9167861 ],
            [ 1. , -3.22775884,  1.83277224]]),
     array([[ 1. , -2.80856491,  1.95630489],
            [ 1. ,  1.62290542, -0.7457237 ]]),
     array([[ 1. , -2.32682366,  2.60844469],
            [ 1. ,  2.12320609,  1.04483217]]),
     array([[ 1. , -4.17616178,  4.09969658],
            [ 1. ,  0.67287935, -5.67652159]])]

y = [0, 1, 0, 1, 1, 0, 1, 0, 0, 0]

from pyhcrf import Hcrf
from sklearn.metrics import confusion_matrix

model = Hcrf(num_states=3,
             l2_regularization=1.0,
             verbosity=10,
             random_seed=3,
             optimizer_kwargs={'maxfun':200})
model.fit(X, y)
pred = model.predict(X)
confusion_matrix(y, pred)
> array([[12,  0],
>        [ 0,  8]])

States

Each state is numbered 0, 1, ..., num_states - 1. The state machine starts in state 0 and ends in num_states - 1. Currently the state transitions are constrained so that, on each element in the input sequence, the state machine either stays in the current state or advances to a state represented by the next number. This default can be changed by setting the transitions and corresponding transition_parameters properties.

Dependencies

numpy, scipy (for the LM-BFGS optimiser and scipy.sparse), and cython.

Installation

Download/clone and run

python setup.py build_ext --inplace
python setup.py install