Structure-Tracing Evidence Mining (STEM)

Original Code for STEM:"STEM: Structure-Tracing Evidence Mining for Knowledge Graphs-Driven Retrieval-Augmented Generation"

We propose Structure-Tracing Evidence Mining (STEM), a novel framework that reframes multi-hop reasoning as a schema-guided graph search task. First, we design a Semantic-to-Structural Projection pipeline that leverages KG structural priors to decompose queries into atomic relational assertions and construct an adaptive query schema graph. Subsequently, we execute globally-aware node anchoring and subgraph retrieval to obtain the final evidence reasoning graph from KG. To more effectively integrate global structural information during the graph construction process, we design a Triple-Dependent GNN (Triple-GNN) to generate a Global Guidance Subgraph (Guidance Graph) that guides the construction. STEM significantly improves both the accuracy and evidence completeness of multi-hop reasoning graph retrieval, and achieves State-of-the-Art performance on multiple multi-hop benchmarks.

Requirements

pip install -r requirements.txt
redis==7.2.4

Retrieval & Inference Excecutation

Run main.py to initiate subgraph retrieval + LLM prediction, and save the predicted answers in the path specified by the "output" parameter:

python ./main.py \
        --config ../config.yaml \
        --mode build \
        --split test \
        --n_proc 1 \
        --output ./predictions.jsonl

Evaluation Metrics (Hit@1 and F1)

Run eval_results.py to calculate metric scores based on the generated predictions file.

python ./eval_results.py \
        --d ./predictions.jsonl \
        --cal_f1 True \
        --top_k 10

LLM-Driven Reverse Generation: Random Walk with Masking for Subgraph Sampling

python ./scripts/run_graph_sampling.py \
        --d webqsp \
        --output_path ../masked_sample_subgraphs.jsonl \
        --n 50

Training

SGDA & SAGB Training Execution

The training of SGDA​ and SAGB​ relies on the LlamaFactory framework (version 0.9.3). You need to place their respective training data .jsonlfiles into the llamafactory dataset directory and add corresponding entries in the dataset_info.jsonlfile. Then execute the following commands:

For SGDA, execute:

llamafactory-cli train ./train/train_config_sgda.yaml

For SAGB, execute:

llamafactory-cli train ./train/train_config_sagb.yaml

T-GNN Training Execution

Run the following command to launch T-GNN training：

python ./train/tgnn_train.py \
        --config ../config.yaml \
        --train_data_path ../rgd_train_20k.jsonl