ERBOT

ERBOT is an R-based entity-resolution research program spanning multiview representation, scalable blocking, threshold selection, clustering, and evaluation integrity. Its public artifact is erbot, an MIT-licensed R package built as an end-to-end toolkit for practical entity resolution: taking records from ingestion through feature construction, candidate generation, matching, clustering, and evaluation in one inspectable pipeline.

What the public toolkit does

As documented in the repository, the package covers:

• Feature construction — TF–IDF representations with singular value decomposition, and embedding-based features, so records can be compared through more than one view of the data.
• Blocking — candidate generation utilities that keep comparison budgets manageable as record counts grow.
• Multi-field similarity — combining similarity metrics across fields of different types rather than relying on a single distance.
• Clustering — multiple strategies for forming entities, including K-means, Louvain community detection, hierarchical clustering, and partitioning around medoids.
• Tuning and evaluation — automated parameter search, evaluation against ground truth, and PDF reports that record parameters, optimization curves, and performance metrics alongside results.

Why it is built this way

Entity resolution decisions are rarely wrong in one obvious place; they go wrong quietly, in thresholds, blocking choices, and evaluation shortcuts. The program treats the pipeline as an object of study: every stage should be replaceable, comparable, and auditable, and every reported result should carry the configuration that produced it.

Status and limitations

ERBOT is an active research program, not a finished product. Interfaces are still evolving, and no benchmark results are published here — performance figures only make sense alongside the dataset and evaluation setting that produced them, and that packaging is ongoing work. The public repository reflects the parts of the program that are stable enough to share.