The package neermatch provides a set of tools for neural-symbolic
entity reasoning and matching. It is designed to support easy set-up,
training, and inference of entity matching models using deep learning,
symbolic learning, and a hybrid approach combining both deep and
symbolic learning. Moreover, the package provides automated fuzzy logic
reasoning (by refutation) functionality that can be used to examine the
significance of particular associations between fields in an entity
matching task.
The project is financially supported by the Deutsche Forschungsgemeinschaft (DFG) under Grant 539465691 as part of the Infrastructure Priority Programme “New Data Spaces for the Social Sciences” (SPP 2431).
The package has also an R implementation available at
r-neer-match.
The package is built on the concept of similarity maps. Similarity maps are concise representations of potential associations between fields in two datasets. Entities from two datasets can be matched using one or more pairs of fields (one from each dataset). Each field pair can have one or more ways to compute the similarity between the values of the fields.
Similarity maps are used to automate the construction of entity matching models and to facilitate the reasoning capabilities of the package. More details on the concept of similarity maps and an early implementation of the package’s functionality (without neural-symbolic components) are given by (Karapanagiotis and Liebald 2023).
The training loops for both deep and symbolic learning models are
implemented in tensorflow (Abadi et al.
2015). The pure deep learning model inherits from the
keras model class (Chollet et al. 2015). The
neural-symbolic model is implemented using the logic tensor network
(LTN) framework (Badreddine et al.
2022). Pure neural-symbolic and hybrid models do not inherit directly
from the keras model class, but they emulate the
behavior by providing custom compile, fit, evaluate, and
predictmethods, so that all model classes in neermatch have a
uniform calling interface.
In addition, the package offers explainability functionality customized for the needs of matching problems. The default explainability behavior is built on the information provided by the similarity map. From a global explainability aspect, the package can be used to calculate partial matching dependencies and accumulated local effects on similarities. From a local explainability aspect, the package can be used to calculate local interpretable model-agnostic matching explanations and Shapley matching values.
Implementing matching models using neermatch is a three-step process:
- Instantiate a model with a similarity map.
- Compile the model.
- Train the model.
To train the model you need to provide three datasets. Two datasets should contain records representing the entities to be matched. By convention, the first dataset is called Left and the second dataset is called Right dataset in the package’s documentation. The third dataset should contain the ground truth labels for the matching entities. The ground truth dataset should have two columns, one for the index of the entity in the Left dataset and one for the index of the entity in the Right dataset.
from neer_match.similarity_map import SimilarityMap
from neer_match.matching_model import NSMatchingModel
import tensorflow as tf
# 0) replace this with your own data preprocessing function
left, right, matches = prepare_data()
# 1) customize according to the fields in your data
smap = SimilarityMap(
{
"title": ["jaro_winkler"],
"platform": ["levenshtein", "discrete"],
"year": ["euclidean", "discrete"],
"developer~dev": ["jaro"]
}
)
model = NSMatchingModel(smap)
# 2) compile
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001))
# 3) train
model.fit(
left, right, matches,
epochs=51, batch_size=16,
log_mod_n=10,
)| Epoch | BCE | Recall | Precision | F1 | Sat |
| 0 | 7.1111 | 1.0000 | 0.2500 | 0.4000 | 0.7330 |
| 10 | 6.4167 | 0.0000 | nan | nan | 0.7512 |
| 20 | 4.9687 | 0.0000 | nan | nan | 0.8263 |
| 30 | 3.4703 | 0.0000 | nan | nan | 0.8527 |
| 40 | 1.2650 | 1.0000 | 0.9750 | 0.9873 | 0.8942 |
| 50 | 0.6165 | 1.0000 | 0.9750 | 0.9873 | 0.9171 |
Training finished at Epoch 50 with DL loss 0.6165 and Sat 0.9171
You can obtain the sources for the development version of neermatch
from its github
repository.
git clone https://github.com/pi-kappa-devel/py-neer-match
To build and install the package locally, from the project’s root path, execute
python -m build
python -m pip install dist/$(basename `ls -Art dist | tail -n 1` -py3-none-any.whl).tar.gz
Online documentation is available for the release version of the package.
Make sure to build and install the package with the latest modifications
before building the documentation. The documentation website is using
sphinx. The build the documentation, from
<project-root>/docs, execute
make html
The logo was designed using Microsoft Designer and GNU Image Manipulation Program (GIMP). The hexagon version of the logo was generated with the R package hexSticker. It uses the Philosopher font.
Several state-of-the-art entity matching (EM) systems have been developed in recent years, utilizing different methodologies to address the challenges of EM tasks. Below, we highlight some of the most recent, best-performing and/or most recognized EM systems:
-
HierGAT: HierGAT introduces a Hierarchical Graph Attention Transformer Network to model and leverage interdependence between EM decisions and attributes. It uses a graph attention mechanism to identify discriminative words and attributes, combined with contextual embeddings to enrich word representations, enabling a more nuanced and interconnected approach to EM (Yao et al. 2022).
-
Ditto: Ditto leverages pre-trained Transformer-based language models to cast EM as a sequence-pair classification task, enhancing matching quality through fine-tuning. It incorporates optimizations such as domain-specific highlighting, string summarization to retain essential information, and advanced data augmentation to improve training, making it both efficient and effective for large-scale EM tasks (Li et al. 2020).
-
CorDEL: CorDEL employs a contrastive deep learning framework that moves beyond twin-network architectures by focusing on both syntactic and semantic matching signals while emphasizing critical subtle differences. The approach includes a simple yet effective variant, CorDEL-Sum, to enhance the model’s ability to discern nuanced relationships in data (Wang et al. 2020).
-
DAEM: This approach combines a deep neural network for EM with adversarial active learning, enabling the automatic completion of missing textual values and the modeling of both similarities and differences between records. It integrates active learning to curate high-quality labeled examples, adversarial learning for augmented stability, and a dynamic blocking method for scalable database handling, ensuring efficient and robust EM performance (Huang et al. 2023).
-
AdaMEL: AdaMEL introduces a deep transfer learning framework for multi-source entity linkage, addressing challenges of incremental data and source variability by learning high-level generic knowledge. It employs an attribute-level self-attention mechanism to model attribute importance and leverages domain adaptation to utilize unlabeled data from new sources, enabling source-agnostic EM while accommodating additional labeled data for enhanced accuracy (Jin et al. 2021).
-
DeepMatcher: This framework is one of the first to introduce deep learning (DL) to entity matching, categorizing learning approaches into SIF, RNN, Attention, and Hybrid models based on their representational power. It highlights DL’s strengths in handling textual and dirty EM tasks while identifying its limitations in structured EM, offering valuable insights for both researchers and practitioners (Mudgal et al. 2018).
-
SETEM: SETEM introduces a self-ensemble training method for EM to overcome challenges in real-world scenarios, such as small datasets, hard negatives, and unseen entities, where traditional Pre-trained Language Model (PLM)-based methods often struggle due to their reliance on large labeled datasets and overlapping benchmarks. By leveraging the stability and generalization of ensemble models, SETEM effectively addresses these limitations while maintaining low memory consumption and high label efficiency. Additionally, it incorporates a faster training method designed for low-resource applications, ensuring adaptability and scalability for practical EM tasks (Ding et al. 2024).
-
AttendEM: AttendEM introduces a novel framework for entity matching (EM) that enhances transformer architectures through intra-transformer ensembling, distinct text rearrangements, additional aggregator tokens, and extra self-attention layers. Departing from the focus on text cleaning and data augmentation in existing solutions, AttendEM innovates within the base model design, offering a distinct approach to pairwise duplicate identification across databases (Low, Fung, and Xiong 2024).
Pantelis Karapanagiotis (maintainer)
Marius Liebald (contributor)
Feel free to share, modify, and distribute. If you implement new features that might be of general interest, please consider contributing them back to the project.
The package is distributed under the MIT license.
Abadi, Martín, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, et al. 2015. “TensorFlow: Large-scale Machine Learning on Heterogeneous Systems.” https://www.tensorflow.org/.
Badreddine, Samy, Artur d’Avila Garcez, Luciano Serafini, and Michael Spranger. 2022. “Logic Tensor Networks.” Artificial Intelligence 303: 103649. https://doi.org/10.1016/j.artint.2021.103649.
Chollet, François et al. 2015. “Keras.” https://keras.io.
Ding, Huahua, Chaofan Dai, Yahui Wu, Wubin Ma, and Haohao Zhou. 2024. “SETEM: Self-ensemble Training with Pre-trained Language Models for Entity Matching.” Knowledge-Based Systems 293 (June): 111708. https://doi.org/10.1016/j.knosys.2024.111708.
Huang, Jiacheng, Wei Hu, Zhifeng Bao, Qijin Chen, and Yuzhong Qu. 2023. “Deep Entity Matching with Adversarial Active Learning.” The VLDB Journal 32 (1): 229–55. https://doi.org/10.1007/s00778-022-00745-1.
Jin, Di, Bunyamin Sisman, Hao Wei, Xin Luna Dong, and Danai Koutra. 2021. “Deep Transfer Learning for Multi-Source Entity Linkage via Domain Adaptation.” In Proceedings of the VLDB Endowment, 15:465–77. https://doi.org/10.14778/3494124.3494131.
Karapanagiotis, Pantelis, and Marius Liebald. 2023. “Entity Matching with Similarity Encoding: A Supervised Learning Recommendation Framework for Linking (Big) Data.” http://dx.doi.org/10.2139/ssrn.4541376.
Li, Yuliang, Jinfeng Li, Yoshihiko Suhara, AnHai Doan, and Wang-Chiew Tan. 2020. “Deep Entity Matching with Pre-Trained Language Models.” Proceedings of the VLDB Endowment 14 (1): 50–60. https://doi.org/10.14778/3421424.3421431.
Low, Jwen Fai, Benjamin C. M. Fung, and Pulei Xiong. 2024. “Better Entity Matching with Transformers Through Ensembles.” Knowledge-Based Systems 293 (June): 111678. https://doi.org/10.1016/j.knosys.2024.111678.
Mudgal, Sidharth, Han Li, Theodoros Rekatsinas, AnHai Doan, Youngchoon Park, Ganesh Krishnan, Rohit Deep, Esteban Arcaute, and Vijay Raghavendra. 2018. “Deep Learning for Entity Matching: A Design Space Exploration.” In Proceedings of the 2018 International Conference on Management of Data, 19–34. https://doi.org/10.1145/3183713.3196926.
Wang, Zhengyang, Bunyamin Sisman, Hao Wei, Xin Luna Dong, and Shuiwang Ji. 2020. “CorDEL: A Contrastive Deep Learning Approach for Entity Linkage.” In 2020 IEEE International Conference on Data Mining (ICDM), 1322–27. IEEE. https://doi.org/10.1109/ICDM50108.2020.00171.
Yao, Dezhong, Yuhong Gu, Gao Cong, Hai Jin, and Xinqiao Lv. 2022. “Entity Resolution with Hierarchical Graph Attention Networks.” In Proceedings of the 2022 International Conference on Management of Data, 429–42. Philadelphia PA USA: ACM. https://doi.org/10.1145/3514221.3517872.