Semantic Typing and Grounding of Pronominal and Nominal Mentions Across Divergent Coreference Schemas

The problem

Current entity recognition and extraction pipelines exhibit sharp performance drops when evaluated on schemas that include pronominal and nominal mentions, dropping to as low as 42% recall on datasets like ACE 2005. Because most taggers assume proper named entities, models fail to type clusters composed entirely of pronouns or misclassify ambiguous pronouns (such as mistaking collective 'we' as person rather than organization) in the absence of cross-sentence antecedent context. Consequently, roughly 10% or more of coreference mentions remain completely ungrounded or unassigned, preventing reliable document-level entity extraction.

Why it matters

Enables information extraction systems to resolve and semantically type document-level mentions regardless of whether they appear as proper names, nominal phrases, or pronouns.

Ways to approach it

Prior-work checks are free with an account. Results someone already ran are shown to everyone.

1. 1

   Conduct a cross-dataset robustness evaluation across OntoNotes, ACE 2005, and OKE benchmarks to measure typing and grounding error distributions across mention types (proper, nominal, pronominal) and evaluate baseline drop-offs under divergent annotation conventions.

   Check for prior work
2. 2

   Develop a joint mention-typing and coreference model that explicitly propagates entity-type posteriors along cross-sentential coreference chains to ungrounded pronouns and nominals, measuring MUC-F1, grounding accuracy, and pronominal type error rates.

   Check for prior work
3. 3

   Design a schema-harmonization mapping method that aligns divergent mention boundaries (e.g., inclusion/exclusion of determiners and relative pronouns like 'which') without requiring manual re-annotation, measuring zero-shot cross-schema recall transfer.

   Check for prior work

Why it might fail

Long-context instruction-tuned large language models may dissolve this limitation via prompt-based extraction if they can resolve pronoun grounding and typing zero-shot without specialized architectures. Furthermore, persistent structural ambiguities in historical dataset annotation conventions may enforce a ceiling on cross-dataset schema transfer.

Evidence

Each paper's own statement of the limitation, verbatim.

• ToMMeR - Efficient Entity Mention Detection from Large Language ModelsACL 2026

  Fails to transfer to schemas with divergent annotation conventions, such as ACE 2005 (42.0% recall) which includes determiners and pronouns.

• Interpretable Coreference Resolution Evaluation Using Explicit SemanticsACL 2026

  Fails to assign semantic types to clusters composed exclusively of pronouns, leaving approximately 10% of coreference mentions unlabeled across datasets.

  Restricted to English-language coreference datasets due to reliance on English CNER taggers.

• Mahānāma: A Unique Testbed for Literary Entity Discovery and LinkingEMNLP 2025

  Only named entities are annotated — pronouns and common nouns are excluded — so the dataset is not suitable for comprehensive coreference resolution

• Entity Disambiguation via Fusion Entity DecodingNAACL 2024

  Underperforms on OKE15/OKE16 datasets because the model is not trained to handle coreference annotations (pronouns and common nouns linked to entities)

• Who Are You Referring To? Coreference Resolution In Image NarrationsICCV 2023

  Performance is low in absolute terms (MUC-F1 19.19), and pronouns with weak language priors (e.g., 'which') are often missed; grounding accuracy is only ~29%

• A Neural Layered Model for Nested Named Entity RecognitionNAACL 2018

  High frequency of type errors for pronouns (e.g., 'we' as ORG vs PER) due to a lack of cross-sentence co-reference information.

Nearest existing work

• Evaluating the Impact of a Hierarchical Discourse Representation on Entity Coreference Resolution PerformanceNAACL 2021
• Incorporating Centering Theory into Neural Coreference ResolutionNAACL 2022
• Analyzing and Visualizing Coreference Resolution ErrorsNAACL 2015
• Conundrums in Entity Coreference Resolution: Making Sense of the State of the ArtEMNLP 2020
• Incorporating Context and External Knowledge for Pronoun Coreference ResolutionNAACL 2019
• ExtEnD: Extractive Entity DisambiguationACL 2022
• Revisiting Joint Modeling of Cross-document Entity and Event Coreference ResolutionACL 2019
• Incorporating Syntax and Semantics in Coreference Resolution with Heterogeneous Graph Attention NetworkNAACL 2021
• Sentence-Incremental Neural Coreference ResolutionEMNLP 2022
• Neural Cross-Lingual Coreference Resolution And Its Application To Entity LinkingACL 2018
• Analysing Coreference in Transformer OutputsEMNLP 2019
• Coreference Resolution with Entity EqualizationACL 2019
• BERT for Coreference Resolution: Baselines and AnalysisEMNLP 2019
• Knowledge-aware Pronoun Coreference ResolutionACL 2019
• Lexical Features in Coreference Resolution: To be Used With CautionACL 2017

Related open problems