TriaffineSpanLayer<T>
Layers · AiDotNet.NeuralNetworks.Layers
The Triaffine nested-NER classifier (Yuan et al. 2022, “Fusing Heterogeneous Factors with Triaffine Mechanism for Nested Named Entity Recognition”, section 3). The deep triaffine transformation TriAff(u, v, w, W) = W x1 u’ x2 w’ x3 v’ with u’ = [MLP(u); 1], v’ = [MLP(v); 1] and w’ = MLP(w) (Equations 1 to 3) scores, for every span (i, j) and label r, each inside token k against the span’s boundaries; a softmax over the inside tokens gives label-wise span representations (Equations 5 to 7) and the auxiliary scores p_ijr (triaffine scoring, section 3.5). The top-m spans by their best non-None auxiliary score then attend over each other with triaffine attention, and their final scores come from the decomposed triaffine scoring (Equations 8 to 13).
Inheritance: LayerBase<T> → TriaffineSpanLayer<T>
Implements: IShapeContract
How It Works
Shapes: input [B, S, D] (or [S, D]); output [B, S * S, R], the span grid at index
start * S + end, R labels with None first. The top-m spans carry their final scores p^c. In training the other
spans carry their auxiliary scores, so the cross-entropy over the grid trains both predictions (the paper applies the
auxiliary loss to every span and the main loss to the top-m); in inference they are assigned None (section 3.6).
Spans with end before start score None.
Constructors
| Constructor | Summary |
|---|---|
TriaffineSpanLayer(int inputDim, int dim, int numLabels, int topM) |
Creates the classifier. |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<ILayer<T>> GetSubLayers() |
|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
|
void ResetState() |