DBLPLink: Scholarly Entity Linking

• DBLPLink is a web-based entity linking system that maps natural-language scholarly queries to DBLP identifiers for authors, titles, venues, and institutions.
• It employs a two-stage pipeline combining T5-based span detection with Elasticsearch candidate retrieval and neural re-ranking using Siamese networks and KG embeddings.
• The system evolves with DBLPLink 2.0, which integrates LLMs for zero-shot mention extraction and supports dynamic RDF-based scholarly graphs.

Searching arXiv for DBLPLink and related papers to ground the article with current citations. DBLPLink is a web-based entity linking system for the DBLP scholarly knowledge graph that maps entities mentioned in natural-language questions to the corresponding DBLP identifiers. It is designed for scholarly queries in which mentions may denote authors, publication titles, venues, or institutions, and in which ambiguity arises because labels can be repeated across different DBLP entities. DBLPLink is presented as the first working entity linker for a scholarly knowledge graph of this kind, and it combines generation-based span detection, database retrieval, and neural re-ranking in an end-to-end demonstrator (Banerjee et al., 2023). A later system, DBLPLink 2.0, adapts the idea to the RDF-based DBLP 2025 graph and replaces the earlier supervised KG-embedding re-ranker with a zero-shot, LLM-centered architecture (Banerjee et al., 30 Jul 2025).

1. Scholarly entity linking in DBLP

DBLPLink addresses a specialized form of entity linking: not merely identifying names in text, but linking spans in scholarly questions to the correct entities in DBLP. The target knowledge graph focuses on computer science publications, authors, venues, and institutions, so the system must distinguish between entity classes that are semantically close and often lexically overlapping. In this setting, the linking problem is shaped by repeated labels, ambiguous author names, publication titles that resemble free text, and venue names that can appear in multiple forms (Banerjee et al., 2023).

The system is situated within a broader entity-linking landscape in which linking has been studied extensively for general-purpose knowledge graphs such as Wikidata and for specialized domains such as biomedicine or finance. DBLPLink targets DBLP specifically, which makes it a scholarly-domain linker rather than a general-purpose one. This suggests that the system’s design is driven by the structure of scholarly metadata and by the distribution of entity mentions in question-answering data over DBLP.

2. Core pipeline and mention extraction

DBLPLink follows a two-stage pipeline. In the first stage, a text-to-text pre-trained LLM, specifically T5, is fine-tuned to extract entity label spans and their types from the input question. The models used in the paper are T5-small and T5-base, trained on the DBLP-QuAD dataset. The paper describes this stage as producing “entity labels and types from the input question,” framing entity linking as generation rather than token-level tagging (Banerjee et al., 2023).

Once labels and types are generated, candidate generation is performed by free-text search over an Elasticsearch instance containing DBLP entity URLs together with their labels. Retrieved entities are then filtered by the predicted types. This type filtering narrows the search space before disambiguation, which is important in a scholarly graph where similar or identical labels can occur across different entity classes.

The system’s default operating mode is deliberately conservative. In normal demo operation, the top-ranked candidate after label-based retrieval is used as the final linked entity. Only if the top candidate has the same label as another candidate does the system proceed to explicit disambiguation. This means that DBLPLink does not treat every case as a hard semantic matching problem; instead, it exploits the observation that many instances can be resolved by label matching alone, and invokes heavier neural ranking only when ambiguity is explicit.

3. Neural disambiguation and representation design

The disambiguation stage uses a Siamese neural network for similarity learning between the question and candidate entities. The question and each candidate are projected into a shared embedding space, but the representation is not purely textual. The system constructs a 969-dimensional vector that mixes contextual text semantics, knowledge-graph structure, and surface-form similarity (Banerjee et al., 2023).

For a question, the first 768 dimensions are occupied by the BERT embedding, and the remaining 201 dimensions are filled with zeros. For an entity candidate, the first 768 dimensions are the BERT embedding of the entity label, and the next 200 dimensions are reserved for entity embeddings. The paper considers three KG embedding families in those 200 dimensions: TransE, DistMult, and ComplEx. The final 969th dimension stores a string similarity score between the entity label and the input question.

Training is performed with positive and negative sample pairs using a triplet ranking loss with an $L_2$ distance metric. At inference time, the question and candidate are vectorized and passed through the trained Siamese network, after which cosine distance between the resulting embeddings is computed. The candidate pair with the lowest distance is selected as the best match. The training and inference procedures therefore use different distance notions: an $L_2$-based ranking loss during training and cosine distance during inference.

Conceptually, the model learns a compatibility function over question-entity pairs. In DBLPLink, that compatibility function is not derived from textual similarity alone. It is explicitly conditioned on BERT encodings, KG embeddings, and a string-similarity feature, making the re-ranker a hybrid semantic-structural matcher rather than a pure dense retriever.

4. Evaluation and empirical behavior

The system is evaluated on the 2,000-question test split of DBLP-QuAD, with F1 score as the main metric. The paper reports results for three ranking modes: “label sorting,” in which the top candidate from label-based retrieval is selected directly; “conditional-disambiguation,” in which re-ranking is triggered only if two candidates share the same label; and “hard-disambiguation,” in which re-ranking is always applied after candidate generation (Banerjee et al., 2023).

Setting	T5-small	T5-base
Label sorting	0.698	0.698
Conditional: TransE	0.700	0.701
Conditional: ComplEx	0.692	0.692
Conditional: DistMult	0.699	0.701
Hard: TransE	0.511	0.521
Hard: ComplEx	0.482	0.484
Hard: DistMult	0.537	0.547

Several conclusions follow directly from these results. First, the size difference between T5-small and T5-base does not materially change performance. The paper attributes this to the character of the label generation task, which is closer to copying relevant pieces from the input than to exploiting deeper pretraining knowledge. Second, hard disambiguation performs substantially worse than plain label sorting. This shows that in DBLP, exact or near-exact string matching is often more reliable than KG-embedding-based similarity. Third, conditional disambiguation yields a small improvement or at least does not hurt, because ambiguity is relatively rare in the DBLP-QuAD test set.

Among embedding methods, DistMult is strongest in the pure hard-disambiguation setting. The authors interpret this as consistent with the structure of DBLP, where many relations are naturally 1-to-$N$, such as one author linked to many papers. In the same interpretation, TransE is better suited to relations closer to one-to-one, while ComplEx is often advantageous for symmetric relations. A plausible implication is that re-ranking quality in scholarly knowledge graphs depends not only on embedding capacity, but also on the relational profile of the graph being linked.

5. Demonstration system and interface design

DBLPLink is not only a model but also a web application intended for interactive use and qualitative inspection. The interface has three main sections. Section A is the input area, where a question can be typed or selected from a dropdown menu, and where the span-detection model and embedding method can be chosen. Section B displays the linked results. For each detected span, it shows the top-ranked entity together with its DBLP label and type, a hyperlink to the corresponding DBLP entity page, and a distance metric indicating how close the match is to the input question; smaller distance means a better match. It also provides collapsible “Ranked Entities” panels that reveal the top 10 ranked candidates for each span. Section C allows certain result combinations to be hidden in order to reduce clutter on the screen (Banerjee et al., 2023).

The demonstrator is designed for comparison rather than opaque prediction. Users are meant to vary T5-small versus T5-base and compare TransE, DistMult, and ComplEx, then inspect how the linked outputs change. This makes the system useful both as a public demo and as an instrument for examining model behavior, particularly in cases where label matching and semantic disambiguation diverge.

6. Limitations, significance, and later redesign

The paper identifies DBLPLink as a pragmatic hybrid linker rather than a universal semantic disambiguator. Its best behavior is obtained when label sorting is preserved and re-ranking is used only conditionally. The learned KG-embedding similarity is not strong enough to outperform simple string-based ranking across the board. The approach also depends on the quality of the generated label spans and on the presence of suitable labels in the Elasticsearch index: if the span detector misses the correct mention, downstream linking cannot recover it. Likewise, the re-ranker is designed primarily for cases in which candidate labels collide, not for arbitrary semantic retrieval (Banerjee et al., 2023).

These limitations are also central to the system’s significance. DBLPLink establishes an empirical baseline for scholarly entity linking over DBLP and demonstrates that, in this domain, surface-form evidence remains unusually strong. Its contribution is therefore both infrastructural and methodological: it provides a usable end-to-end linker for a scholarly KG and an experimental result showing that neural re-ranking should complement rather than replace lexical matching.

DBLPLink 2.0 revises this design for the RDF-based DBLP 2025 graph, which adds a new entity type, dblp:Stream, for publication venues such as conferences, journals, series, and repositories. Instead of retraining KG embeddings and supervised re-rankers, the newer system uses prompted LLMs for mention and type extraction, type-specific Elasticsearch retrieval, and KG-neighborhood-based re-ranking. Its central scoring method asks an LLM whether a candidate supported by a one-hop neighborhood triple is the correct entity and uses the log-probability of the token “yes” before generation, aggregated by mean pooling over triples. The paper reports that Qwen-2.5-3B provides the best balance of size and performance in the public demo, with F1 = 0.4400, MRR = 0.5388, Hits@1 = 0.4900, Hits@5 = 0.5900, and Hits@10 = 0.6200 on 100 sampled DBLP-QuAD questions (Banerjee et al., 30 Jul 2025).

The relationship between the two systems is not a simple version increment. DBLPLink 1.0 is a supervised, hybrid linker centered on T5, Elasticsearch, and Siamese re-ranking with KG embeddings. DBLPLink 2.0 is a zero-shot redesign motivated by schema change and by the need to support the new dblp:Stream inventory without retraining multiple components. Taken together, they document an evolution from supervised re-ranking over a fixed scholarly KG to evidence-based LLM scoring over an evolving RDF graph.