Struct HandshakingMatrix 

pub struct HandshakingMatrix {
    pub cells: Vec<HandshakingCell>,
    pub seq_len: usize,
    pub num_labels: usize,
}

Handshaking matrix for joint entity-relation extraction.

Research Alignment (W2NER, AAAI 2022)

From the W2NER paper (arXiv:2112.10070):

“We present a novel alternative by modeling the unified NER as word-word relation classification, namely W2NER. The architecture resolves the kernel bottleneck of unified NER by effectively modeling the neighboring relations between entity words with Next-Neighboring-Word (NNW) and Tail-Head-Word- (THW-*)* relations.”

In TPLinker/W2NER, we don’t just tag tokens - we tag token PAIRS. The matrix M[i,j] contains the label for the span (i, j).

Key Relations

Relation	Description	Purpose
NNW	Next-Neighboring-Word	Links adjacent tokens within entity
THW-*	Tail-Head-Word	Links end of one entity to start of next

Benefits

• Overlapping entities (same token in multiple spans)
• Joint entity-relation extraction in one pass
• Explicit boundary modeling
• Handles flat, nested, AND discontinuous NER in one model

Fields

cells: Vec<HandshakingCell>

Non-zero cells (sparse representation)

seq_len: usize

Sequence length

num_labels: usize

Number of labels

Implementations

impl HandshakingMatrix

pub fn from_dense( scores: &[f32], seq_len: usize, num_labels: usize, threshold: f32, ) -> Self

Create from dense scores with thresholding.

Arguments

• scores - Dense [seq_len, seq_len, num_labels] scores
• threshold - Minimum score to keep

pub fn decode_entities<'a>( &self, registry: &'a SemanticRegistry, ) -> Vec<(SpanCandidate, &'a LabelDefinition, f32)>

Decode entities from handshaking matrix.

In W2NER convention, cell (i, j) represents a span where:

• j is the start token index
• i is the end token index (inclusive, so we add 1 for exclusive end)