Ethnicity Sensitive Author Disambiguation Using Semi-supervised Learning

Gilles Louppe, Hussein T. Al-Natsheh, Mateusz Susik, Eamonn James Maguire

TL;DR

This paper tackles the challenge of accurately identifying scientific publications by the same author, even with name variations and homonyms. They introduce an improved semi-supervised learning approach that leverages a massive dataset of crowdsourced author annotations, incorporating ethnicity-sensitive features to enhance disambiguation for non-Western names.

What Problem Does This Paper Address?

In academic databases, a significant challenge is author name disambiguation: correctly grouping publications written by the same individual. This is complicated by authors using different name variations throughout their careers (synonymy) and multiple authors sharing the same name (polysemy). This problem is particularly acute for researchers from non-Western cultures due to naming conventions and transliteration issues. Accurate disambiguation is crucial for reliable search results, bibliometrics, and scientific network analysis. While manual curation and persistent identifiers (like ORCID) exist, fully automated solutions are needed for scalability and immediate impact.

Methods

The authors propose a semi-supervised author disambiguation pipeline with three main components:

  1. Blocking: Signatures (pieces of information identifying an author on a publication) are initially grouped into smaller blocks to reduce computational complexity. They move beyond simple "Surname and First Initial" (SFI) blocking by proposing phonetic-based blocking using algorithms like Double Metaphone or NYSIIS. This aims to increase recall by grouping name variants that sound similar. They also introduce a two-phase blocking strategy for authors with multiple surnames.
  2. Linkage Function: A supervised learning model is trained to determine the probability that two signatures belong to the same author. They experiment with various classifiers (Random Forests, GBRT, Logistic Regression) and a rich set of features. A key innovation here is the introduction of ethnicity-sensitive features. These features, derived from census data and an SVM classifier, estimate the probability of a name belonging to a specific ethnic group. By multiplying these probabilities for a pair of signatures, the linkage function becomes sensitive to author origin, improving disambiguation for non-Western names. They also emphasize the importance of building a balanced training set with both easy and difficult cases.
  3. Semi-supervised Clustering: Within each block, signatures are clustered using hierarchical agglomerative clustering. They explore different "cut-off" strategies (no cut, global cut, block cut) to determine how to form final clusters, leveraging the semi-supervised nature of their annotated data.

Datasets

The primary dataset used is a large collection of over 1 million crowdsourced annotations from the INSPIRE portal (a digital library for high-energy physics). This dataset comprises approximately 10 million signatures and represents over 15,000 distinct individuals. The authors highlight that this dataset is released publicly, serving as a new benchmark for author disambiguation research.

Key Insights and Ideas

  • Phonetic Blocking Boosts Recall: Using phonetic representations of names for blocking significantly improves the ability to group name variants, leading to higher recall in disambiguation.
  • Ethnicity Matters for Disambiguation: Incorporating ethnicity-sensitive features is crucial for improving disambiguation accuracy, especially for non-Western names, by making the linkage function more nuanced.
  • Balanced Training Data is Key: Simply sampling pairs randomly is insufficient. Building a training set that includes a balance of easy and difficult cases (e.g., oversampling ambiguous pairs) leads to a more robust linkage function.
  • Semi-supervised Learning is Powerful: Leveraging a large set of partially annotated data allows for effective training of linkage functions and guided clustering, outperforming purely unsupervised methods.
  • Most Names Aren't Ambiguous: The "no cut" clustering strategy, which assumes all signatures within a block belong to the same author, still yields surprisingly good results, indicating that a large proportion of author names are not inherently ambiguous.

Key Results

The proposed method, combining phonetic blocking (specifically NYSIIS) with a Random Forest linkage function trained on balanced, blocked data, and using block cut-off strategies for clustering, achieves a B3 F-measure of 0.9862. This represents a significant improvement over baseline methods. The ethnicity-sensitive features were found to be among the most important predictors for the linkage function, alongside co-authorship, affiliation, and full name similarity.

Limitations and Future Work

The authors acknowledge that handling signatures with inconsistent surnames or inconsistent first given names (cases 4 and 5 from Section 3.1) remains an open challenge for blocking. They also suggest further work on improving phonetization functions tailored for author disambiguation and exploring better ethnicity-sensitive features. The potential benefits of adapting the pipeline to specific cultural contexts where disambiguation is known to be less efficient are highlighted as a promising future direction.

Summary updated

BibTeX
@inproceedings{louppe2016ethnicity,
  author = {Gilles Louppe and Hussein T. Al-Natsheh and Mateusz Susik and Eamonn James Maguire},
  title = {{Ethnicity Sensitive Author Disambiguation Using Semi-supervised Learning}},
  booktitle = {KESW},
  year = {2016},
  doi = {10.1007/978-3-319-45880-9\_21},
}
Figure