Reasoning on Graphs (RoG)

Official Implementation of "Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning".

Reasoning on graphs (RoG) synergizes LLMs with KGs to enable faithful and interpretable reasoning. We present a planning-retrieval-reasoning framework, where RoG first generates relation paths grounded by KGs as faithful plans. These plans are then used to retrieve valid reasoning paths from the KGs for LLMs to conduct faithful reasoning and generate interpretable results.

Requirements

pip install -r requirements.txt

Pre-trained weights

Our code will automatically download the model weight from the huggingface.

You can find the pre-trained weights here.

Datasets

Our code will automatically download the data from the huggingface.

RoG-WebQSP
RoG-CWQ

Subgraph Extraction

We extract the subgraphs from the Freebase following previous studies. The code can be found here.

Inference

Requirements: Any GPU with at least 12GB memory.

Step1: Planning (Generate relation paths)

Run: ./scripts/planning.sh

python src/qa_prediction/gen_rule_path.py \
        --model_name RoG \
        --model_path rmanluo/RoG \
        -d {RoG-webqsp,RoG-cwq} \
        --split test \
        --n_beam 3

Generated rules will be saved at: results/gen_rule_path/{dataset}/{model_name}/{split}

Step2: Reasoning (Generate answers with RoG)

Run: ./scripts/rog-reasoning.sh

python src/qa_prediction/predict_answer.py \
        --model_name RoG \
        --model_path rmanluo/RoG \
        -d {RoG-webqsp,RoG-cwq} \
        --prompt_path prompts/llama2_predict.txt \
        --add_rul \
        --rule_path {rule_path} \

Answers will be saved at: results/KGQA/{dataset}/{model_name}/{split}

Plug-and-play Reasoning (Generate answers with different LLMs)

Note: you need to set your openai key at .env to use ChatGPT.

Run: ./scripts/plug-and-play.sh

python src/qa_prediction/predict_answer.py \
        --model_name {gpt-3.5-turbo,alpaca,llama2-chat-hf,flan-t5} \
        -d {RoG-webqsp,RoG-cwq} \
        --prompt_path {prompt_path} \
        --add_rule \
        --rule_path {rule_path}

Interpretable Reasoning

Run: python scripts/interpretable_example.py

from transformers import pipeline, AutoTokenizer
import torch

MODEL_PATH_OR_NAME="rmanluo/RoG"

tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH_OR_NAME, use_fast=False)
model = pipeline("text-generation", model=MODEL_PATH_OR_NAME, tokenizer=tokenizer, device_map="auto", torch_dtype=torch.float16)

print("====EXAMPLE 1: ====")

INPUT_TEXT_1 = """Based on the reasoning paths, please answer the given question and explain why 

Reasoning Paths: 
Northern District -> location.administrative_division.first_level_division_of -> Israel -> government.form_of_government.countries -> Parliamentary system

Question: 
What type of government is used in the country with Northern District?"""

outputs = model(INPUT_TEXT_1, return_full_text=False)
print(outputs[0]['generated_text'])

Training

Training Datasets

You can download the processed datasets from RoG_train_data.tar.tz. Unzip the files and put them under datasets/ folder.

Process datasets

1. Build question to relation path pairs.

python src/align_kg/build_align_qa_dataset.py -d {RoG-webqsp,RoG-cwq} --split {train,validation,test}

2. Build joint-training datasets.

python src/joint_training/preprocess_align.py
python src/joint_training/preprocess_qa.py

3. Build interpretable examples.

python src/joint_training/generate_explanation_results.py

Training RoG

2 A100-80GB GPUs are required for training RoG.

Run: ./scripts/train.sh

Results

Bibinfo

If you found this repo helpful, please help us by citing this paper:

@inproceedings{luo2024rog,
title={Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning},
author={Luo, Linhao and Li, Yuan-Fang and Haffari, Gholamreza and Pan, Shirui},
booktitle={International Conference on Learning Representations},
  year={2024}
}