[Paper Review] DeepER -- Deep Entity Resolution

Muhammad Ebraheem, Saravanan Thirumuruganathan+3 more|arXiv (Cornell University)|Oct 2, 2017

Data Quality and Management50 references56 citations

TL;DR

DeepER introduces distributed representations of tuples (DRs) for entity resolution, using word embeddings and LSTM-based compositionality to reduce labeled data, with LSH-based blocking for efficiency. It demonstrates competitive accuracy across benchmarks and multilingual data.

ABSTRACT

Entity resolution (ER) is a key data integration problem. Despite the efforts in 70+ years in all aspects of ER, there is still a high demand for democratizing ER - humans are heavily involved in labeling data, performing feature engineering, tuning parameters, and defining blocking functions. With the recent advances in deep learning, in particular distributed representation of words (a.k.a. word embeddings), we present a novel ER system, called DeepER, that achieves good accuracy, high efficiency, as well as ease-of-use (i.e., much less human efforts). For accuracy, we use sophisticated composition methods, namely uni- and bi-directional recurrent neural networks (RNNs) with long short term memory (LSTM) hidden units, to convert each tuple to a distributed representation (i.e., a vector), which can in turn be used to effectively capture similarities between tuples. We consider both the case where pre-trained word embeddings are available as well the case where they are not; we present ways to learn and tune the distributed representations. For efficiency, we propose a locality sensitive hashing (LSH) based blocking approach that uses distributed representations of tuples; it takes all attributes of a tuple into consideration and produces much smaller blocks, compared with traditional methods that consider only a few attributes. For ease-of-use, DeepER requires much less human labeled data and does not need feature engineering, compared with traditional machine learning based approaches which require handcrafted features, and similarity functions along with their associated thresholds. We evaluate our algorithms on multiple datasets (including benchmarks, biomedical data, as well as multi-lingual data) and the extensive experimental results show that DeepER outperforms existing solutions.

Motivation & Objective

• Reduce human labeling and feature engineering in entity resolution while maintaining high accuracy.
• Render both syntactic and semantic similarity between tuples without extensive feature engineering.
• Provide a holistic, DR-based blocking method to efficiently limit comparisons across all attributes.

Proposed method

• Compute DRs of tuples by either averaging word embeddings or composing via uni-/bi-directional LSTM with LSTM units.
• Train an end-to-end model to tune DRs for a specific ER task to improve accuracy.
• Use similarity vectors between tuple DRs as input to a classifier for match/mismatch decisions.
• Introduce LSH-based blocking that uses DRs of tuples to form blocks considering all attributes.
• Explain how to handle out-of-vocabulary words and scenarios with partial or minimal dictionary coverage.
• Discuss tuning word embeddings via fine-tuning or retrofitting to domain resources.

Experimental results

Research questions

• RQ1 Can DRs of tuples capture both syntactic and semantic similarities for effective ER without heavy feature engineering?
• RQ2 How can ER blocking be achieved efficiently using DRs and LSH to scale to large datasets?
• RQ3 What strategies (averaging vs. LSTM-based composition) yield better ER performance across different datasets?
• RQ4 How can word embeddings be adapted or tuned for domain-specific ER tasks (full, partial, or minimal coverage)?

Key findings

• DR-based tuple representations enable effective similarity measures for ER without extensive manual feature engineering.
• LSTM-based compositional DRs can provide advantages on datasets where word order and attribute interactions matter.
• LSH-based blocking over DRs significantly reduces comparisons while leveraging semantic similarity across all attributes.
• End-to-end tuning of DRs via supervised learning improves ER accuracy on task-specific data.
• Vocabulary retrofitting and domain-specific embedding strategies help address full/partial/minimal coverage scenarios.
• Experiments show DeepER outperforms state-of-the-art solutions on benchmark, biomedical, and multilingual datasets.