Struct SpanRepConfig 

pub struct SpanRepConfig {
    pub hidden_dim: usize,
    pub max_width: usize,
    pub use_width_embeddings: bool,
    pub width_emb_dim: usize,
}

Configuration for span representation.

Research Context (Deep Span Representations, arXiv:2210.04182)

From “Deep Span Representations for NER”:

“Existing span-based NER systems shallowly aggregate the token representations to span representations. However, this typically results in significant ineffectiveness for long-span entities.”

Common span representation strategies:

Method	Formula	Pros	Cons
Concat	[h_i; h_j]	Simple, fast	Ignores middle tokens
Pooling	mean(h_i:h_j)	Uses all tokens	Loses boundary info
Attention	attn(h_i:h_j)	Learnable	Expensive
GLiNER	FFN([h_i; h_j; w])	Balanced	Requires width emb

Recommendation (GLiNER Default)

For most use cases, concatenating first + last token embeddings with a width embedding provides the best tradeoff:

• O(N) complexity (vs O(N²) for all-pairs attention)
• Captures boundary positions (critical for NER)
• Width embedding disambiguates “I” vs “New York City”

Fields

hidden_dim: usize

Hidden dimension of the encoder

max_width: usize

Maximum span width (in tokens)

GLiNER uses K=12: “to keep linear complexity without harming recall.” Wider spans rarely contain coherent entities.

use_width_embeddings: bool

Whether to include width embeddings

Critical for distinguishing spans of different lengths with similar boundary tokens.

width_emb_dim: usize

Width embedding dimension (typically hidden_dim / 4)

Trait Implementations

impl Clone for SpanRepConfig

fn clone(&self) -> SpanRepConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for SpanRepConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for SpanRepConfig

fn default() -> Self

Returns the “default value” for a type.