Struct GraphCorefConfig 

pub struct GraphCorefConfig {
    pub max_iterations: usize,
    pub link_threshold: f64,
    pub transitivity_bonus: f64,
    pub per_shared_neighbor_bonus: f64,
    pub string_similarity_weight: f64,
    pub head_match_weight: f64,
    pub distance_weight: f64,
    pub max_distance: Option<usize>,
    pub include_singletons: bool,
    pub pronoun_proper_bonus: f64,
    pub early_stop: Option<GraphCorefEarlyStopConfig>,
}

Configuration for graph-based coreference resolution.

These parameters control the iterative refinement process. The defaults are based on findings from Miculicich & Henderson (2022):

• max_iterations = 4: Paper found T=4 optimal; more iterations don’t help
• link_threshold = 0.5: Standard classification threshold
• transitivity_bonus = 0.15: Reward for transitive consistency

Tuning Guide

Parameter	Higher Value	Lower Value
link_threshold	Fewer, more confident links	More links, potential noise
transitivity_bonus	Stronger clustering effect	More independent decisions
max_iterations	More refinement passes	Faster, less propagation
head_match_weight	Trust head matches more	Rely more on full string

Fields

max_iterations: usize

Maximum refinement iterations before stopping.

The G2GT paper found T=4 to be optimal on CoNLL 2012. Fewer iterations leave potential coreference links undiscovered; more iterations don’t improve results and waste computation.

link_threshold: f64

Minimum score to create a coreference link.

A pair (mᵢ, mⱼ) is linked if score(mᵢ, mⱼ) + context_bonus > threshold.

transitivity_bonus: f64

Bonus added for transitive consistency.

If mentions A and B share neighbors in the current graph (i.e., both are already linked to some common mention C), this bonus is added to encourage the model to also link A and B directly.

Note: This is our heuristic approximation of G2GT’s graph-conditioned attention. The full G2GT model encodes graph structure as:

Attention(Q,K,V,Lk,Lv) = softmax(Q·(K+Lk)/√d)·(V+Lv)
where Lk = E(G^{t-1})·Wk

We approximate this by explicit score adjustment rather than attention modification.

per_shared_neighbor_bonus: f64

Bonus for each shared neighbor.

Scaled by the number of shared neighbors: total_bonus = shared_count * per_neighbor_bonus

string_similarity_weight: f64

Weight for string similarity in pairwise scoring.

head_match_weight: f64

Weight for head word matching.

The G2GT paper emphasizes head-based matching. When mentions have head_start/head_end set, head matching is used. Otherwise falls back to last word heuristic.

distance_weight: f64

Weight for distance penalty in pairwise scoring.

max_distance: Option<usize>

Maximum character distance to consider (mentions further apart are not linked).

include_singletons: bool

Include singletons (mentions with no coreference) in output.

Default: false (only return multi-mention chains). Set to true for evaluation against datasets that include singletons.

pronoun_proper_bonus: f64

Bonus when a pronoun links to a proper noun.

Pronouns are weak signals alone but should link to antecedents.

early_stop: Option<GraphCorefEarlyStopConfig>

Optional early-stop controls for iterative refinement.

GraphCoref already stops when it reaches a fixed point (Gₜ == Gₜ₋₁). This option additionally stops on:

• Cycle detection (e.g., A→B→A oscillation across iterations)
• Stagnation (edge count stops changing for N iterations)

This is an analogue of “overthinking” / redundancy detection (CoRE-Eval), implemented using observable signals (graph structure) rather than hidden states.

Trait Implementations

impl Clone for GraphCorefConfig

fn clone(&self) -> GraphCorefConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GraphCorefConfig

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

Formats the value using the given formatter.

impl Default for GraphCorefConfig

fn default() -> Self

Returns the “default value” for a type.