LLM4CD: Leveraging Large Language Models for Open-World Knowledge Augmented Cognitive Diagnosis

Introduction

This repository contains the PyTorch implementation of the paper LLM4CD: Leveraging Large Language Models for Open-World Knowledge Augmented Cognitive Diagnosis.

Requirements

• numpy
• transformers
• torch
• dgl
• pickle

Quick Start

To get started, run the following command in your terminal:

python main.py --dataset ass09 --llm chatglm2 --num_epochs 100 --lr 1e-4 --model lamdacd --batch_size 4000 --encoder graph

Dataset

The public datasets we used are ASSIST09, ASSIST12 and Junyi. For data processing, please refer to data/ass09/log_data.json.

To divide data into train_set, val_set and test_set, run the following command:

python divide_data.py

To get the graph information, please refer to data/ass09/graph

python construct_concept_map.py
python process_edge.py

Text generation

The exercise information and concept information are available at text/ass09, please refer to id_skill_desc_dict.json, id_skillname_dict.json and the augmented question.json.

To construct cognitive text for students, please refer to text/ass09/prompt_generator.py:

python prompt_generator.py

After obtain the text from exercise and student sides, encode the generated text into semantic embeddings:

python encoder.py --model_name chatglm2