Sentence Transformer Workflow

Sentence Transformer Workflow

Step 1: Input Sentence Pair

• Input consists of two sentences: A and B.
• Both are processed independently using the same BERT model (a twin/siamese network).

Step 2: Embedding Extraction

• Sentences A and B are passed separately through BERT.
• Each yields a fixed-size embedding vector: $a=\text{Embed}(A),b=\text{Embed}(B)$.

Step 3: Compute Difference and Combine

• Compute absolute difference: |a - b|.
• Form a combined vector: $[a; b; |a - b|].

Step 4: Feedforward Neural Network (FFNN)

• Pass the combined vector through a two-layer FFNN.
• Output is a set of raw logits (real-valued scores for each class).

Step 5: Classification via Softmax

• Apply softmax to logits to get class probabilities.
• The class with the highest probability is selected.