ABL: A Naive Implementation of Auto Labelling System

This is my graduation project v1 (metaphase).

Project Structure

abl-text/
├── abl/                    # Core framework
│   ├── datasets/            # Dataset definitions and loading
│   ├── models/              # Model components
│   ├── nn/                  # Text perception modules
│   └── utils/               # Utility functions
├── scripts/                 # Training and evaluation scripts
└── requirements.txt         # Package dependencies

Data Format

This framework expects data in the following format:

• train.jsonl, val.jsonl, test.jsonl: Each line contains a JSON object with text, label, and optional concept annotations
• meta/labels.json: Mapping of label names to indices
• meta/concepts.json: Mapping of concept names to indices

Usage

Training

To train a model(By myself simple test):

python scripts/text_reasoning_train.py \     
    --config configs/config.json \     
    --dataset-config configs/dataset_config.json \     
    --output-dir experiments/text_reasoning \     
    --epochs 50     --batch-size 32     --lr 1e-4     --device cuda

Visualization training process

To show the training process:

tensorboard --logdir experiments/text_reasoning/tensorboard

Evaluation

To evaluate a trained model:

python scripts/text_reasoning_test.py \     
    --config configs/config.json \     
    --dataset-config configs/dataset_config.json \     
    --checkpoint experiments/text_reasoning/checkpoints/model_best.pth \     
    --output-dir experiments/text_reasoning/results \     
    --batch-size 32     --device cuda     --visualize     --interpret-rules

Key Components

• Text Encoder: Uses a pre-trained BERT model to encode text.
• Concept Embedding: Detects concepts in text and represents them as embeddings.
• Quasi-Symbolic Reasoning: Bridges neural perception with symbolic reasoning.
• Reasoning: Applies rule-based reasoning to connect concepts to labels.
• Program Executor: Executes symbolic programs for interpretable predictions.