When we think about tracking references in a conversation, we usually think about concrete entities. If a text mentions "Elon Musk," we expect subsequent pronouns like "he" or descriptions like "the billionaire" to point back to that same physical person. Traditional natural language processing models excel at this type of entity clustering. As Heeyoung Lee et al. note in Deterministic Coreference Resolution Based on Entity-Centric, Precision-Ranked Rules., rule-based and machine learning systems have long targeted these nominal entity chains.
However, human discourse is rarely limited to tracking objects. We frequently refer back to actions, events, and entire occurrences. Consider the following narrative:
"The company decided to acquire its main competitor. This merger shocked the market, but that was exactly their plan."
In this snippet, the noun phrase "this merger" and the demonstrative pronoun "that" do not refer to a physical object or a person. Instead, they point directly to the action introduced by the verb acquire.
Tracking these dynamic occurrences as referable units—a challenge known as event coreference or event anaphora resolution—requires a fundamental shift in how models represent discourse. Early theoretical foundations recognized this necessity. Damaris M. Ayuso in Discourse Entities in Janus extended early discourse frameworks using intensional logic to treat events and actions as explicit "mental hooks" or referable units. Similarly, Michael Strube 0001 et al. introduced functional centering in Functional Centering to track deictic expressions that refer to entire situations or propositions rather than isolated noun phrases.
The Linguistic Asymmetry of Event Anaphora
Why can't standard entity-centric models resolve these event links? The core issue is a structural and grammatical asymmetry. Entity coreference typically operates within the same part of speech (noun phrase to pronoun, or noun phrase to noun phrase), relying heavily on surface features like gender, number, and lexical overlap.
Event coreference, by contrast, frequently crosses parts of speech. An event is typically introduced to a narrative by a verb, but subsequent references to it are often noun phrases or clauses.
[Verb Trigger] "exploded" ───► [Noun Phrase Anaphora] "the blast" ───► [Demonstrative] "that"
To resolve these connections, architectures must explicitly bridge the gap between verbal actions and nominal descriptions. Bin Chen et al. pioneered a systematic approach to this in Resolving Event Noun Phrases to Their Verbal Mentions, using structural syntactic information via tree kernels and twin-candidate ranking to map event-related nouns back to their preceding verbs. Without these specialized mechanisms, traditional systems fall short. As Veselin Stoyanov et al. observed in Reconciling OntoNotes: Unrestricted Coreference Resolution in OntoNotes with Reconcile, standard pipelines often prune out or ignore verbal mentions entirely, leaving them blind to the distinct linguistic properties of event-based references.
Furthermore, many event references are associative rather than identical, falling under the umbrella of "bridging descriptions." Renata Vieira et al. demonstrated in Towards Resolution of Bridging Descriptions that connecting a verb to a subsequent noun requires tracking discourse focus and nominalization rather than simple lexical matching. When the antecedent is an entire clause, traditional span-to-span architectures break down. To address this, Hideo Kobayashi et al. developed a constrained multi-task learning framework in Constrained Multi-Task Learning for Bridging Resolution that enforces consistency between bridging, coreference, and information status, allowing models to capture abstract clausal antecedents that standard architectures miss.
Scaling to Cross-Document Contexts and Retrieval
Within a single document, tracking event chains is vital for understanding narrative progression. For example, Saliha Azzam et al. showed in Using Coreference Chains for Text Summarization that tracking event-level coreference is essential for summarizers to synthesize the narrative arc of dynamic occurrences like corporate mergers or elections.
However, the problem becomes significantly more complex when scaling cross-document. Early datasets suffered from a selection bias: they primarily featured documents where the main topic was the event itself (descriptive mentions). To force models to learn more realistic discourse patterns, Alon Eirew et al. introduced the WEC-Eng dataset in WEC: Deriving a Large-scale Cross-document Event Coreference dataset from Wikipedia. This dataset captures "referential mentions," where an event is briefly cited as a point of reference within a document covering an entirely unrelated topic.
When searching for an event across millions of documents, exhaustive pairwise clustering becomes computationally impossible. To solve this bottleneck, the authors reframed cross-document resolution as a retrieval task rather than a clustering task. In Cross-document Event Coreference Search: Task, Dataset and Modeling, Alon Eirew et al. adapted Open-Domain Question Answering architectures, embedding fine-grained coreference scoring directly into a Deep Passage Retrieval (DPR) framework to pull all mentions of a single event out of massive, unstructured text repositories.
[Query Event Mention] ──► [DPR Framework + Coreference Scoring] ──► [Retrieve All Matching Mentions Across Corpus]
Ultimately, understanding an event requires grounding it to the real world. Just as traditional Entity Linking grounds a word like "Boston" to a specific map coordinate or Wikipedia page, Xiaodong Yu et al. introduced the task of Event Linking in Event Linking: Grounding Event Mentions to Wikipedia. Their model, EVELINK, grounds event triggers (like the verb "detonated") directly to specific Wikipedia event pages (like "Boston Marathon Bombing"), leveraging the surrounding entity context to disambiguate the action.
Why Event Resolution Matters for Downstream NLU
Failing to resolve event-level references creates a major bottleneck for deep natural language understanding. In fact, Sebastian Martschat et al. revealed in Recall Error Analysis for Coreference Resolution that over half of all persistent recall errors across state-of-the-art coreference systems stem from an inability to resolve abstract references to complex actions and sentences.
When models cannot track these event links, performance degrades across several critical downstream tasks:
- Textual Entailment: Document-level reasoning relies heavily on event tracking. Shachar Mirkin et al. demonstrated in Assessing the Role of Discourse References in Entailment Inference that nearly half of all entailment inferences require resolving discourse references, with roughly one-quarter of those critical links involving verbs and events rather than nouns.
- Machine Translation: Standard translation models often mistranslate ambiguous pronouns when they refer to an entire preceding action rather than a single noun. Asaf Yehudai et al. demonstrated in Evaluating and Improving the Coreference Capabilities of Machine Translation Models that injecting explicit coreference markers during fine-tuning is necessary to help translation models maintain semantic grounding for event-level pronouns.
- Stance Detection: In complex news narratives, tracking "who says what about whom" requires linking pronouns or abstract phrases back to their original event contexts. Xinliang Frederick Zhang et al. showed in Generative Entity-to-Entity Stance Detection with Knowledge Graph Augmentation that grounding these linguistic pointers to their original event contexts is vital for capturing nuanced political sentiment and media framing.
- Information Extraction: To extract clean facts from text, a system must separate the primary narrative from background details. As Ludovic Jean‐Louis et al. argued in Utilisation d'indices temporels pour la segmentation événementielle de textes, resolving entity-to-event associations ensures that extracted facts are correctly attributed to the primary occurrence rather than secondary background details.
Key Foundational Work
Discourse Entities in Janus (1989) — contributed 1 concept elements to this direction.
Distributives, Quantifiers and a Multiplicity of Events (1985) — contributed 1 concept elements to this direction.