Exploring Key Point Analysis with Pairwise Generation and Graph Partitioning

Requirements

• torch 2.0.1
• transformers 4.30.2
• networkx
• sentence_transformers
• rouge_score
• scikit-learn

Key Point Generation

• ArgKP with Flan-T5-base: scirpts/run_argkp_base_muc5.sh
• ArgKP with Flan-T5-large: scirpts/run_argkp_large_muc5.sh
• QAM with Flan-T5-large: scirpts/run_QAM_base_muc5.sh
• QAM with Flan-T5-large: scirpts/run_QAM_large_muc5.sh

Graph Partitioning

cd GraphPartitioning && python graph_clustering --dataset ArgKP/QAM --output_file {file path to generated key point output file, under GraphPartitioning/eval_outputs/}

Note

We use BLEURT to score the relevance between two sentences. However, unlike our code, which is based on PyTorch, BLEURT requires a TensorFlow environment. To avoid package dependency conflicts, we install an additional TensorFlow environment and then invoke BLEURT through HTTP requests. You can run bleurt_app.py in the TensorFlow environment.