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.
Figures & tables
Statistic
Subtask 1
Subtasks 2 & 3
(gold)
(predicted)
Corpus
Documents
973
967
Sentences with mentions
2,153
2,140
Mention instances
2,974
2,860
Unique surface forms
837
791
Table 1: Corpus statistics for the SOMD 2026 training sets. Subtask 1 uses gold-standard mentions; Subtasks 2 and 3 share a common training set of automatically predicted mentions. Cross-doc rate (non-singletons) excludes singleton clusters, which require no linking decision.
Noise Type
Example
Original
We used MATLAB for statistical analysis and visualization.
Boundary
We used MATLAB for statistical analysis and visualization.
Modification
We used the MATLAB for statistical analysis and visualization.
Mention
We used Python for statistical analysis and visualization.
Substitution
We used NumPy for statistical analysis and visualization.
Table 2: Illustration of noise injection methods applied to a software mention. The original mention is shown in blue , while noisy mentions are shown in red . Boundary modification extends or truncates mention spans; mention substitution replaces the mention string with another software name sampled from the training set. Each noise type simulates a different failure mode of upstream mention detection. All methods are tested at noise rates of 0%, 10%, 25%, 50%, 75%, and 100%.
Subtask 1
Subtask 2
Subtask 3
System
CoNLL
B 3
CEAF e
MUC
CoNLL
B 3
CEAF e
MUC
CoNLL
B 3
CEAF e
MUC
System A †
0.98
0.99
0.96
0.99
0.98
0.99
0.95
0.99
0.96
0.97
0.92
0.99
System B †
0.92
0.94
0.86
0.97
0.92
0.93
0.85
0.98
‡
‡
‡
‡
Fuzzy Matching
0.95
0.95
0.91
0.98
0.95
0.94
0.91
0.99
0.94
0.94
0.90
0.99
Context Aware
0.96
0.96
0.93
0.98
0.96
0.96
0.93
0.99
‡
‡
‡
‡
Table 3: Official test-set results for SOMD 2026. CoNLL is the average of MUC, B 3 , and CEAFe F 1 . Bold = best overall. Underline = best among our systems. † = other participants. ‡ = no submission on Codabench.
Injected Noise Rate
Noise Type
System
∅
25%
50%
75%
100%
Δ
Subtask 1 (gold mentions)
Boundary Modification
Fuzzy Matching
0.70
0.65
0.60
0.54
0.50
0.20
Context Aware
0.82
0.80
0.79
0.77
0.75
0.07
Mention Substitution
Fuzzy Matching
0.70
0.49
0.34
0.23
0.18
0.52
Context Aware
0.82
0.57
0.39
0.25
0.19
0.63
Table 4: CoNLL F 1 under all noise types at increasing noise rates on the training set. Δ = F 1 at 0% − F 1 at 100% (lower = more robust). Baseline scores differ from official test-set results (Table 3 ) as noise experiments are conducted on the training set. Underline = best F 1 ; bold Δ = most robust system per noise type.
Subtask
System
Time (s)
CoNLL F 1
F1/sec
Subtask 1
FM
0.60 ± 0.00
0.95
1.584
CAR
4.45 ± 0.69
0.96
0.215
Subtask 2
FM
47.06 ± 0.41
0.95
0.020
CAR
45.39 ± 5.14
0.96
0.021
Table 5: Inference time and precision–speed trade-off on official test sets of Subtasks 1 and 2. Efficiency = CoNLL F 1 / inference time (s). Measured on an Intel x86_64 CPU (14 cores, 33.3 GB RAM) over 5 runs.
Deployment condition
Recommended
Small corpus, high detector quality
FM
Large corpus ( n≫ 1k mentions)
CAR
Boundary errors dominate
CAR
Mention identity errors dominate
Improve detector
Table 6: Practical system selection guide based on deployment context.
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.
Julia Matela, Frank Krüger
Department of Electrical Engineering and Computer Science, Wismar University of Applied Sciences Wismar, Germany
We present our submission to the LLM track of the 2026 Computational Models of Reference, Anaphora and Coreference (CRAC 2026) shared task. With an average CoNLL F1 score of 74.32 on the official test set, our system ranked first in the LLM track, and third overall. Our system is based on the Gemma-3-27b model, fine-tuned using a two-stage strategy with a multilingual base adapter followed by dataset-specific adapters. We represent mention spans by their headword using an XML-inspired format with local reindexing and annotate documents iteratively. These design choices proved effective across languages, document lengths, and annotation guidelines.
Antoine Bourgois, Olga Seminck, Thierry Poibeau
Lattice (CNRS UMR 8094 & ENS-PSL & Universit´e Sorbonne Nouvelle), Montrouge, France
Coreference resolution is typically evaluated using aggregate statistical metrics such as CoNLL-F1, which measure structural overlap between predicted and gold clusters. While widely used, these metrics offer limited diagnostic insights, penalizing errors without revealing whether a system struggles with specific semantic categories, such as people, locations, or events, and making it difficult to interpret model capabilities or derive actionable improvements. We address this gap by introducing a semantically-enhanced evaluation framework for coreference resolution. Our approach overlays Concept and Named Entity Recognition (CNER) onto coreference outputs, assigning semantic labels to nominal mentions and propagating them to entire coreference clusters. This enables the computation of typed scores aimed at evaluating mention extraction and linking capabilities stratified by semantic class. Across our experiments on OntoNotes, LitBank, and PreCo, we show that our framework uncovers systematic weaknesses that remain obscured by aggregate metrics. Furthermore, we demonstrate that these diagnostics can be used to design targeted, low-cost data augmentation strategies, achieving measurable out-of-domain improvements.
Bruno Gatti, Giuliano Martinelli, Roberto Navigli
Sapienza NLP Group, Sapienza University of Rome · 2Babelscape