Reuse and Adaptation for Entity Resolution through Transfer Learning

Abstract

Entity resolution (ER) is one of the fundamental problems in data integration, where machine learning (ML) based classifiers often provide the state-of-the-art results. Considerable human effort goes into feature engineering and training data creation. In this paper, we investigate a new problem: Given a dataset DT for ER with limited or no training data, is it possible to train a good ML classifier on DT by reusing and adapting the training data of dataset DS from same or related domain? Our major contributions include (1) a distributed representation based approach to encode each tuple from diverse datasets into a standard feature space; (2) identification of common scenarios where the reuse of training data can be beneficial; and (3) five algorithms for handling each of the aforementioned scenarios. We have performed comprehensive experiments on 12 datasets from 5 different domains (publications, movies, songs, restaurants, and books). Our experiments show that our algorithms provide significant benefits such as providing superior performance for a fixed training data size.

Deep Transfer Learning for Multi-source Entity Linkage via Domain Adaptation

“…• TLER [34] is a non-deep transfer learning framework that defines a standard feature space and reuses the seen data to train models for the new domain.…”

“…A popular approach is to adapt the pre-trained model for the new task through fine-tuning [20], or by adding new functions to specific tasks such as object detection [15]. In terms of EL, TLER [34] is a non-deep method that reuses and adopts seen data from the source domain to train models for the new domain. Auto-EM [40] proposes to pre-train models for both attribute-type (i.e., schema) and attribute value matching based on word-and character-level similarity.…”

“…We define the attribute importance in entity linkage as the high-level transferable knowledge and automatically learn it through a proposed attributelevel attention mechanism (what to transfer). In general, as transfer learning aims to transfers knowledge learned from the domain with abundant training data to a related target domain with limited data, the existing works either rely on increasing the labeling volume by introducing the external data (e.g., public knowledge bases) [40] or reusing the seen training data [34]. On the contrary, AdaMEL adopts domain adaptation (DA) to jointly update the attention scores for attributes in both the seen and unseen data as the basis for entity linkage (how to handle multiple sources), so that the knowledge is adaptive to the continuously incoming data sources.…”

Survive the Schema Changes: Integration of Unmanaged Data Using Deep Learning

“…Entity matching (EM) [16], which is to identify data instances that refer to the same real-world entity, is also related. Some EM works also employ a deep learning-based approach [24], [37], [42], [49], [57], [73], [82]. Mudgal and et al [57] evaluates and compares the performance of different deep learning models applied to EM with three types of data: structured data, textual data, and dirty data (with missing value, inconsistent attributes and/or miss-placed values).…”

“…In the past few years, deep learning (DL) has become the most popular direction in machine learning and artificial intelligence [46], [65], and has transformed a lot of research areas, such as image recognition, computer vision, speech recognition, natural language processing, etc.. In recent years, DL has been applied to database systems and applications to facilitate parameter tuning [47], [71], [76], [81], indexing [21], [43], partitioning [34], [86], cardinality estimation and query optimization [39], [44], and entity matching [24], [37], [42], [57], [73], [82]. While predictions based on deep learning cannot guarantee correctness, in the Big Data era, errors in data integration are usually tolerable as long as most of the data is correct, which is another motivation of our work.…”