Overall Research Flowchart (Large Image) Prompt: Title: Overall Framework for Link Prediction in Temporal Knowledge Graphs Based on Hypergraphs and Dual Paths Objective: To visualize the entire research process, from data input to prediction output, highlighting the dual-path collaboration and contrastive learning mechanism. Core Elements and Process: 1. Input Layer: • Box element: Label "Temporal Knowledge Graph Data (e.g., ICEWS14, GDELT)", containing entity, relation, and timestamp triples. • Arrow pointing to "Data Preprocessing Module", indicating format standardization. 2. Semantic Initialization Module: • Box element: Label "Pre-trained Language Model (e.g., BERT)", input is entity/relation text description, output is "Semantic Initial Embedding". • Sub-process: Text Encoding → Linear Projection → Generate Initial Embedding Vector. 3. Dual-Path Construction Module: • Parallel Dual Branches: ◦ Local Query View Path: Based on the query entity and timestamp, construct a "Local Dynamic Hypergraph" (hyperedge encodes historical facts), output "Local Short-Path Embedding". ◦ Global Context View Path: Expand the time window, construct a "Long-Term Hypergraph" through "Multi-Hop Sampling", output "Long-Path Temporal Embedding". • Each branch contains hypergraph convolution steps: spatial dimension (multi-head attention aggregation) and temporal dimension (enhanced temporal encoder + decay weight). 4. Fusion and Optimization Module: • Box element: "Dual-Path Embedding Fusion", weighted combination of local and long-range embeddings through a gating mechanism. • Connect to "Contrastive Learning Constraint": Positive sample pairs (local/global embeddings of the same query) and negative sample pairs (embeddings of different queries), optimized through contrastive loss. • Final Output: "Link Prediction Results" (e.g., Hits@1, MRR metrics). Style Requirements: • Use rectangular boxes to represent modules, diamond boxes to represent decision points (e.g., decay threshold), and arrows to indicate the direction of the process. • Color coding: Local path uses blue tones, long path uses green tones, and the fusion part is emphasized with yellow. • Label key terms (e.g., "Dynamic Hypergraph", "Multi-Hop Sampling"), font is Times New Roman, 10-12pt. Local Short-Path Temporal Embedding Diagram (Small Image 1) Prompt: Title: Hypergraph Convolution and Temporal Embedding Process for Local Query View Objective: To refine the embedding generation process of the local path, highlighting the spatio-temporal two-dimensional processing of hypergraph convolution. Core Elements and Process: 1. Input: • Access from the "Local Query View Path" of the overall flowchart, the input is "Query Entity + Timestamp" and "Historical Facts within the Local Time Window". 2. Hypergraph Construction: • Diagram: Nodes represent entities, hyperedges (elliptical) enclose multiple
Research Background: Given that vulnerabilities are unavoida...
![Figure 1: Structural limitations of traditional passive Timeline Summarization (TLS) pipelines. The illustration visualizes why current methods produce brittle narratives. Stage 1: Input: A chaotic cloud of massive unstructured text serves as the noisy data source. Stage 2: Process: A passive retrieval mechanism selects fragments based on surface similarity, ignoring contextual gaps. These fragments are forced through a fixed schema module that lacks flexibility. Stage 3: Output: The result is a brittle timeline, depicted as a broken bridge spanning a chasm, exhibiting macro-level causal gaps with large physical voids between milestones, labeled as missing intermediate stages, representing broken logical flow, and micro-level element gaps with faded, semi-transparent nodes with placeholders such as [Date: ?], labeled as underspecified milestones. Summary: This passive pipeline yields weakly verifiable narratives, motivating our proposed Active Timeline Completion (ATC) framework.](/_next/image?url=https%3A%2F%2Fpub-8c0ddfa5c0454d40822bc9944fe6f303.r2.dev%2Fai-drawings%2FYTKKT4KX6I1Z0I9zVD8Dvk3dLNiZdiB4%2Ff8b117a6-3860-4361-9ba9-68a3f0ad3b68%2Ffa327c56-a2aa-42f4-bb4c-4bc614c51c33.png&w=3840&q=75)
![Method 2: Dynamic Coefficient Learning Based on Attention Mechanism
- **Core Idea**: Introduce an attention layer to allow the model to automatically learn the weights of different tokens in emotion regulation. Implementation steps:
(1) Construct Token Feature Representation: Extract the Token state H=[h1, h2, ....hn] (where n is the number of Tokens) from a certain layer of the model, and use the function vector (A - B) as the 'emotion query vector'.
(2) Calculate Token-Emotion Attention Weights: Calculate the matching degree between each Token and the emotion query vector through scaled dot-product attention:
ai = Softmax (huaye)
where d is the dimension of hi.
(3) Adaptive Injection of Function Vectors: Apply hi+ai.(A-B) to each Token state. This is the idea for my part.](/_next/image?url=https%3A%2F%2Fpub-8c0ddfa5c0454d40822bc9944fe6f303.r2.dev%2Fai-drawings%2F21yePdfXfrXMwKl3x4fhPD0TH1d7yghe%2F4a7680a8-334b-44a9-bf77-6adf2b1dd5e7%2Fcd765bfe-d81d-4f08-b6e5-510889b6cb6c.png&w=3840&q=75)
