This repo includes the code implementation of the TreeRare and the code for the experiments.
If you find our code or the paper useful, please cite the paper:
@misc{zhang2025treeraresyntaxtreeguidedretrieval,
title={TreeRare: Syntax Tree-Guided Retrieval and Reasoning for Knowledge-Intensive Question Answering},
author={Boyi Zhang and Zhuo Liu and Hangfeng He},
year={2025},
eprint={2506.00331},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2506.00331},
}
git clone git@github.com:billycrapediem/TreeRare.git
cd TreeRare
conda env create -f environment.yml
conda activate TreeRare
pip install -r requirements.txt
mkdir output_dir
mkdir dataWe directly uses BM25 implemented through pyserini as retriver.
.
├── prompt/ # Prompts used for CoT and zero-shot/few-shot settings
│ ├── Ambig_Doc.txt # Zero-shot CoT prompt for AmbigDocQA
│ ├── asqa_prompt.txt # Zero-shot CoT prompt for ASQA
│ └── hotpot_cot.txt # Few-shot CoT prompt for HotpotQA
│
├── script/ # Shell scripts for deployment and running experiments
│ ├── deploy.sh # Script for deploying LLaMA 3.3-70B model
│ ├── df1.sh # Evaluate DF1 performance on ASQA
│ └── experiment.sh # Run all experiments sequentially
│
├── src/ # Source code
│ ├── eval/ # Evaluation scripts
│ │ ├── disambigF1.py # Compute Disambiguation F1 score
│ │ ├── eval_ambigdoc.py # Evaluate performance on AmbigDocQA
│ │ └── eval_multi_hop.py # Evaluate performance on multi-hop QA
│ │
│ ├── ambigdoc_inference.py # Run inference on AmbigDocQA
│ ├── asqa_inference.py # Run inference on ASQA
│ ├── BM25.py # BM25 retriever implementation
│ ├── consituency_tree.py # Constituency parse tree utilities
│ ├── dependency.py # Dependency tree parser
│ ├── dpr.py # Dense Passage Retriever (DPR) code
│ ├── hoppotqa_inference.py # Run inference on HotpotQA (multi-hop)
│ ├── traverse_algo.py # Tree traversal algorithms
│ └── utils.py # Miscellaneous utility functions
In our exerpiment script, all the datasets are inthe ./data folder. And all the model output is under `./output_dir' folder.
Access and download the ASQA dataset here.
Access and download the AmbigDocQA dataset here.
Access and download the HotpotQA dataset here.
Access and download the MuSiQue dataset here.
Access and download the 2WikiMultihopQA dataset here.
before runing experiment, you need to add your own api key and the model name into the experiment.sh file.
sh script/experiment.sh