Optimal Transport-based Alignment of Learned Character Representations for String Similarity

Abstract

String similarity models are vital for record linkage, entity resolution, and search. In this work, we present STANCE-a learned model for computing the similarity of two strings. Our approach encodes the characters of each string, aligns the encodings using Sinkhorn Iteration (alignment is posed as an instance of optimal transport) and scores the alignment with a convolutional neural network. We evaluate STANCE's ability to detect whether two strings can refer to the same entity-a task we term alias detection. We construct five new alias detection datasets (and make them publicly available). We show that STANCE (or one of its variants) outperforms both state-of-the-art and classic, parameter-free similarity models on four of the five datasets. We also demonstrate STANCE's ability to improve downstream tasks by applying it to an instance of cross-document coreference and show that it leads to a 2.8 point improvement in B 3 F1 over the previous state-of-the-art approach.

Named Entity Disambiguation at Scale

“…Finally, [22] proposed an architecture using a Multi-Layered Perceptron (MLP) to recognize toponyms, and similar neural network architecture is used by [27] for entity linking. Our work is similar to these two last studies, but our pair-wise ranking architecture is coupled with a strategy that allows us to leverage the disambiguation to millions of source and formal names by filtering irrelevant formal names out.…”

“…Named Entity Disambiguation (NED) [14,27] is the task of linking textual variations of Named Entities (NE) 1 to their target names, which are usually provided as a list of formal names. For instance, while recognizing "Philip Morris" as an NE is the job of a Named Entity Recognition (NER) system, associating it to "Philip Morris International Inc (PMI)" in a list of formal names as a means of disambiguation is performed via NED.…”

“…The tf-idf model has two tasks; to generate feature vectors for the deep learning model and to set a threshold for limiting the formal names when using the system at inference time. We test the system on four datasets for alias detection [27] and compare the results with several baselines as well as a state-of-the-art NED system.…”

The impact of preprint servers in the formation of novel ideas

“…FACS and Fluorescence Activated Cell Sorting). A simple alias table or string similarity extension (Tam et al, 2019) would be a clear improvement. Leveraging high precision concept extraction systems (King et al, 2020) might improve clustering even more.…”

A Deep Learning Approach to Geographical Candidate Selection through Toponym Matching

“…Even a highly enriched KB will not cover all possible name variations (especially of less popular entities), not to mention spelling mistakes or OCR errors. To address this issue, recent work has addressed this problem by including a noise detector to the entity linking system that operates at a tokenlevel [11], or by learning and aligning character representations for string similarity [20].…”