Semantic Centroids and Hierarchical Density-Based Clustering for Cross-Document Software Coreference Resolution

Julia Matela; Frank Krüger

cs.CLMar 25, 2026

Semantic Centroids and Hierarchical Density-Based Clustering for Cross-Document Software Coreference Resolution

Authors: Julia Matela, Frank Krüger

Organizations: Department of Electrical Engineering and Computer Science, Wismar University of Applied Sciences Wismar, Germany

Abstract

This paper describes the system submitted to the SOMD 2026 Shared Task for Cross-Document Coreference Resolution (CDCR) of software mentions. Our approach addresses the challenge of identifying and clustering inconsistent software mentions across scientific corpora. We propose a hybrid framework that combines dense semantic embeddings from a pre-trained Sentence-BERT model, Knowledge Base (KB) lookup strategy built from training-set cluster centroids using FAISS for efficient retrieval, and HDBSCAN density-based clustering for mentions that cannot be confidently assigned to existing clusters. Surface-form normalization and abbreviation resolution are applied to improve canonical name matching. The same core pipeline is applied to Subtasks 1 and 2. To address the large scale settings of Subtask 3, the pipeline was adapted by utilising a blocking strategy based on entity types and canonicalized surface forms. Our system achieved CoNLL F1 scores of 0.98, 0.98, and 0.96 on Subtasks 1, 2, and 3 respectively.

Figures & tables

Figure 1: System Description

Metric	Task 1	Task 2	Task 3
MUC F1	0.9939	0.9916	0.9912
BCUB F1	0.9905	0.9858	0.9724
CEAFE F1	0.9584	0.9521	0.9218
CoNLL F1	0.9809	0.9765	0.9618

Table 1: System performance across SOMD 2026 Subtasks.

Fraction	n	Embed (s)	KB Match (s)	Canon. (s)	HDBSCAN (s)	Merge (s)	Total (s)
10%	21,995	$2.23\pm 0.30$	$0.60\pm 0.02$	$0.09\pm 0.02$	$14.02\pm 0.18$	$0.12\pm 0.01$	$17.06\pm 0.40$
25%	54,987	$5.54\pm 0.38$	$1.51\pm 0.01$	$0.33\pm 0.31$	$85.94\pm 1.03$	$0.30\pm 0.01$	$93.62\pm 1.15$
50%	109,975	$10.97\pm 0.32$	$3.02\pm 0.01$	$0.73\pm 0.32$	$64.97\pm 0.64$	$0.58\pm 0.02$	$80.27\pm 0.54$
75%	164,962	$16.87\pm 0.27$	$4.47\pm 0.03$	$0.68\pm 0.24$	$142.29\pm 0.77$	$0.90\pm 0.01$	$165.20\pm 0.85$
100%	219,950	$22.32$	$5.89$	$0.80$	$99.01$	$1.16$	$129.17$

Table 2: Mean wall-clock time per pipeline stage at each test-set fraction. Values for 10%–75% are means

\pm

std over six random seeds; the 100% row is the mean of two independent full runs.

Figure 2: Pipeline runtime vs. corpus size for Subtask 3. Each coloured line shows the mean wall-clock time of one pipeline stage across six random test set fractions. The bold black line shows total runtime.

Coreference Resolution

Coreference resolution links expressions that refer to the same entities or events across sentences, documents, languages, and modalities.

See all 25 papers

Do Lexical and Contextual Coreference Resolution Systems Degrade Differently under Mention Noise? An Empirical Study on Scientific Software Mentions
Apr 2, 2026Atilla Kaan Alkan, Felix Grezes, Jennifer Lynn Bartlett +3Coreference Resolution
Harvard-Smithsonian Center for Astrophysics, Cambridge, MA, USA
Closing the Gap at CRAC 2026: Two-Stage Adaptation for LLM-Based Multilingual Coreference Resolution
May 16, 2026Antoine Bourgois, Olga Seminck, Thierry PoibeauMultilingual Language Models Coreference Resolution
Lattice (CNRS UMR 8094 & ENS-PSL & Universit´e Sorbonne Nouvelle), Montrouge, France
Interpretable Coreference Resolution Evaluation Using Explicit Semantics
May 11, 2026Bruno Gatti, Giuliano Martinelli, Roberto NavigliCoreference Resolution
Sapienza NLP Group, Sapienza University of Rome · 2Babelscape

Apr 2, 2026J/K move · Enter open · S save

Do Lexical and Contextual Coreference Resolution Systems Degrade Differently under Mention Noise? An Empirical Study on Scientific Software Mentions

Atilla Kaan Alkan, Felix Grezes, Jennifer Lynn Bartlett, Anna Kelbert+2

Harvard-Smithsonian Center for Astrophysics, Cambridge, MA, USA

Coreference Resolution

Open paper AskPDF arXiv

We present our participation in the SOMD 2026 shared task on cross-document software mention coreference resolution, where our systems ranked second across all three subtasks. We compare two fine-tuning-free approaches: Fuzzy Matching (FM), a lexical string-similarity method, and Context Aware Representations (CAR), which combines mention-level and document-level embeddings. Both achieve competitive performance across all subtasks (CoNLL F1 of 0.94-0.96), with CAR consistently outperforming FM by 1 point on the official test set, consistent with the high surface regularity of software names, which reduces the need for complex semantic reasoning. A controlled noise-injection study reveals complementary failure modes: as boundary noise increases, CAR loses only 0.07 F1 points from clean to fully corrupted input, compared to 0.20 for FM, whereas under mention substitution, FM degrades more gracefully (0.52 vs. 0.63). Our inference-time analysis shows that FM scales superlinearly with corpus size, whereas CAR scales approximately linearly, making CAR the more efficient choice at large scale. These findings suggest that system selection should be informed by both the noise profile of the upstream mention detector and the scale of the target corpus. We release our code to support future work on this underexplored task.

Abstract

Coreference Resolution

Explore similar work

Do Lexical and Contextual Coreference Resolution Systems Degrade Differently under Mention Noise? An Empirical Study on Scientific Software Mentions

Closing the Gap at CRAC 2026: Two-Stage Adaptation for LLM-Based Multilingual Coreference Resolution

Interpretable Coreference Resolution Evaluation Using Explicit Semantics