T-KAER: Transparency-enabled Knowledge-Augmented Entity Resolution Framework

Environment Setup

First, install dependencies:

conda env create --name kaer_39 --file=environments.yml

Second, install Refined for entity linking:

pip install https://github.com/amazon-science/ReFinED/archive/refs/tags/V1.zip

Third, install Doduo for column annotation:

pip install https://github.com/megagonlabs/doduo.git

Example scripts to implement Doduo: DittoPlus/doduo_scripts/doduo_annotation.py.

Experiments:

• [Experiment I]: Run KAER and Documenting Experimental Process
• [Experiment II]: Evaluating and Analyzing ER results

Experiment I: Run KAER and Documenting Experimental Process

1. Commands and HyperParameters

Entity Resolution by Pre-trained Language Models (PLMs) can be started by running train_ditto.py script under dittoPlus folder.

The command and key hyperparameters can be tuned by users are as follows:

$ cd dittoPlus
$ python train_ditto.py --task {*} --dk {*} --prompt {*} --kbert {*}

• task: dataset folder name (trainset, validset, and testset), all meta-information documented in dittoPlus/configs.json.
• dk: domain knowledge name: {default:none (ditto baseline), sherlock, doduo, entityLinking}
• prompt: prompting methods name: {default: 1 (space), 0: kbert, 2 (slash)}
• kbert: using kbert (constrained pruning method) or not: {default: False, True}

2. Experiment Result: Log File Generated

After the experiment, one log file will be generated and can be found under this directory: dittoPlus/output/.

Experiment II: Evaluating and Analyzing ER results

1. Evaluating script based on the log files: dittoPlus/ev_results.py
2. Compare the performance across the KA methods

Directory and Descriptions

Directory	Contents Descriptions
data	Dataset from The ER-Magellan Benchmark
environment.yml	All Dependencies Required to Run the Experiments {sherlock}
dittoPlus	ditto + Domain Knowledge

Related Papers

[1] Fang, L., Li, L., Liu, Y., Torvik, V. I., and Ludäscher, B. (2023). Kaer: A knowledge augmented pre-trained language model for entity resolution. Knowledge Augmented Methods for Natural Language Processing workshop in conjunction with AAAI 2023. arXiv preprint arXiv:2301.04770.

[2] Li, L., Fang, L., Liu, Y., Torvik, V. I., & Ludäscher, B. (2024). T-KAER: Transparency-enhanced Pre-Trained Language Model for Entity Resolution. IDCC, 18.