When building large-scale literature graphs like Semantic Scholar (Groeneveld et al. (2018)) or CiteSeerX (Wu et al. (2014)), we cannot rely on raw text strings to identify individuals. Author name disambiguation (AND) is the process of resolving identity conflicts within bibliographic databases, ensuring that every publication is correctly linked to its true author.
Without robust identity resolution, the integrity of the entire scientific record degrades. It introduces significant noise into citation networks, skews metrics like the h-index (Kreiman et al. (2011); McKeown et al. (2016)), and compromises expert-finding systems (Wang et al. (2017)).
The Two Faces of Identity Ambiguity
Name ambiguity in unstructured text manifests as a classic data-clustering challenge with two opposing forces:
- Polysemy (The Merging Problem): Multiple distinct individuals share the exact same name string (e.g., "J. Smith"). This problem is particularly acute in cultures with high homonymy rates or short character lengths, such as Chinese names, where relying on simple string matching like initials and surnames fails completely (Xu et al. (2010)).
- Synonymy (The Splitting Problem): A single individual appears under multiple variations due to typographical errors, changing naming traditions, or varying journal formats (Mdluli et al. (2021); Maxwell-Smith et al. (2022)). This causes academic databases to fragment a researcher's profile into multiple disconnected records (Anthis et al. (2022)).
The goal of an author disambiguation system is to find the mathematical sweet spot that groups all instances of synonymy together while cleanly splitting instances of polysemy.
Disambiguation via Contextual Neighborhoods
To tell two individuals with identical names apart, automated systems look at their surrounding metadata. If one "J. Smith" publishes on deep learning with co-authors from Stanford, and another publishes on marine biology with co-authors from Tokyo, they can be separated based on context.
Disambiguation models leverage several metadata features to construct a unique fingerprint for each publication instance (Mishra et al. (2018)):
- Co-authorship networks: The shared social circle of researchers.
- Institutional affiliations: The organizations or departments where the research was conducted.
- Research sub-fields: The topical keywords and categorical labels of the venue.
However, raw metadata is frequently noisy or underspecified. Affiliations might list generic department names without identifying the parent institution (Zhang et al. (2020)), or a single institution might be represented by multiple fragmented identifiers across different journals (Butler et al. (2017)). When metadata is sparse, simple rule-based matching quickly breaks down.
Graph-Based Identity Resolution
To move beyond fragile, flat string comparisons, modern approaches treat author disambiguation as a graph topology problem. Instead of viewing a paper as an isolated text document, the literature is modeled as a heterogeneous network connecting papers, authors, venues, and institutions (Tang et al. (2007); Buntine et al. (2014)).
The intuition here is that an author's identity is defined by their structural position within this network. Peng (2019) demonstrates how network representation learning can map these multi-relational connections into a continuous vector space. By embedding papers and authors based on their structural proximity, we can use adaptive clustering to group publications.
[Paper A] --- (Co-author: Wang) --- [J. Smith] --- (Topic: AI) ===> Cluster 1 (Tech)
[Paper B] --- (Co-author: Silva) --- [J. Smith] --- (Topic: Biology) ===> Cluster 2 (Bio)
If two ambiguous name instances share a dense local network of co-authors or sit within a continuous self-citation lineage (Luo et al. (2023)), their graph embeddings will naturally pull them together. This topological approach prevents "identity bleed"—the erroneous merging of distinct entities due to superficially similar contexts (Lawrie et al. (2013)).
Balancing Scale and Precision
Disambiguating hundreds of millions of entities requires balancing computational efficiency with model complexity. Comparing every paper to every other paper creates an $O(N^2)$ bottleneck that is impossible to compute at scale.
To solve this, researchers deploy multi-stage pipelines. The authors typically use a fast "blocking" phase to group similar names using multi-stage string processing, drastically reducing the candidate search space (Chin et al. (2013)). Once the data is partitioned into manageable blocks, fine-grained classification models are applied.
Furthermore, Zhang et al. (2020) showed that combining simple heuristic rules with a lightweight convolutional neural network can achieve state-of-the-art accuracy while running significantly faster than heavy graph neural networks. In the absence of universally adopted persistent identifiers, these hybrid approaches allow digital libraries to maintain data integrity across massive, evolving scientific repositories (Romero (2022)).