rustling.seq_feature

Configurable feature templates for sequence labeling models.

Package Contents

class rustling.seq_feature.SeqFeatureTemplate

A single feature template for sequence labeling.

Do not instantiate directly. Use seq_obs() or seq_label() factory functions instead.

rustling.seq_feature.seq_obs(*positions: int, transform: str | None = None) → SeqFeatureTemplate

Create an observation feature template.

Parameters:

• *positions – One or more relative positions in [-4, +4]. A single position extracts a unigram; multiple positions form an n-gram (e.g., seq_obs(-1, 0) is a bigram).

• transform – Optional transform applied to each observation. "first_char" or "final_char". Defaults to identity.

Returns:

A feature template for use in the features parameter of model constructors.

Raises:

ValueError – If no positions given, position out of range, or unknown transform.

rustling.seq_feature.seq_label(*positions: int) → SeqFeatureTemplate

Create a label feature template.

Label features look back at previously predicted labels. Only supported for averaged perceptron models (not HMM).

Parameters:

*positions – One or more negative relative positions in [-4, -1]. E.g., seq_label(-1) uses the label one step back.

Returns:

A feature template for use in the features parameter of model constructors.

Raises:

ValueError – If no positions given, position out of range [-4, +4], or position is non-negative.