Figure 3
The illustration presents a three-step workflow for document re-ranking using a capsule network with convolution and cross-attention mechanisms. The diagram includes section titles and annotations in Chinese, corresponding to each step, alongside English labels within the components. The layout flows from left to right in Steps 1 and 2, and then from right to left in Step 3. Step 1: Multi-scale Document Capsule Construction. On the left, a vertical stack of boxes represents candidate documents labeled “Candidate Document 1”, “Candidate Document 2”, ellipsis, and “Candidate Document K”. A rightward arrow leads to a feature extraction module labeled “L-layer Multi-scale Convolution”, shown as stacked layers with interconnected nodes. This module produces capsule feature vectors labeled “C subscript 1”, “C subscript 2”, ellipsis, and “C subscript L”. These capsule vectors are aggregated into a combined representation shown as a rectangular block. The combined output is then split into multiple feature subsets labeled “S superscript 1”, “S superscript 2”, ellipsis, and “S superscript K”. Step 2: Query-Aware Cross-Attention Mechanism. The split feature sets, along with a query vector labeled “v subscript q”, are input into a transformer-style attention module. Inside this module, operations are arranged vertically as “Attention”, “Add and Norm”, “Feed Forward”, another “Add and Norm”, and “Softmax”. Arrows indicate a top-to-bottom flow through these stages. The output consists of refined feature representations corresponding to the input feature sets. Step 3: Dynamic Routing Re-ranking. On the right, the refined feature sets labeled “S superscript 1” to “S superscript K” are transformed into primary capsules labeled “u subscript 1”, “u subscript 2”, ellipsis, and “u subscript K”. Each capsule is associated with a routing coefficient labeled “c subscript 1”, “c subscript 2”, ellipsis, and “c subscript K”. These capsules are combined at a central aggregation node and passed through a transformation labeled “Squash”, which normalizes the outputs. This produces high-level capsules labeled “v subscript 1”, “v subscript 2”, ellipsis, and “v subscript K”. On the far left, a vertical list of candidate documents represents the final re-ranked output. An arrow labeled “Filtering and Re-ranking” indicates that the high-level capsule outputs determine the final ordering of the documents.CapsGCN-Rank framework