In standard natural language processing, coreference resolution is often treated as a matching game between nouns. If a text says, "The trophy didn't fit into the brown suitcase because it was too large," a model must decide if "it" refers to the trophy or the suitcase. This is entity-level coreference, where a pronoun points to a specific, concrete noun phrase (NP).
However, human language is rarely that tidy. In a sentence like, "John sold his car, and that was a mistake," the word "that" does not refer to John or the car. It refers to the entire event of selling the car. This is abstract anaphora (or discourse deixis), where referents are not physical objects but propositions, situations, or actions.
The Problem of the Missing Anchor
The fundamental difficulty in resolving abstract anaphora is that there is often no single "anchor" word in the preceding text to point to. Traditional models rely on identifying noun phrases as candidate antecedents. But as Donna K. Byron demonstrated, in spoken dialogue, between 22% and 50% of pronouns refer to non-nominal entities like clauses or entire sentences.
When a pronoun refers to an abstract concept, the "antecedent" is a semantic segment rather than a lexical one. Jingxuan Yang and colleagues found that while word-level attention mechanisms excel at finding concrete referents, they struggle with abstract ones because these referents lack a specific lexical head. This gap makes it nearly impossible for systems to maintain a coherent "mental model" of a conversation, a problem Massimo Poesio identified as early as 1993 in Assigning a Semantic Scope to Operations.
Shell Nouns as Semantic Containers
One way researchers have tackled this is through the study of shell nouns—words like fact, issue, decision, or reason. These nouns act as "shells" that contain complex propositional content.
To solve the lack of labeled data for these abstract references, Varada Kolhatkar, Heike Zinsmeister, and Graeme Hirst introduced a "silver standard" training method in Interpreting Anaphoric Shell Nouns using Antecedents of Cataphoric Shell Nouns as Training Data. They realized that cataphoric shell nouns (where the content follows the noun, e.g., "The fact that it rained...") are easy for machines to label automatically. By training on these easy examples, they could teach models to resolve the much harder anaphoric cases (where the noun points backward to a previous sentence).
From Rules to Neural Ranking
The evolution of this field mirrors the broader shift in AI from symbolic logic to neural architectures:
- Functional Centering: Early researchers like Michael Strube and Udo Hahn moved away from rigid syntax toward "functional" roles, trying to track how the focus of a discourse shifts across sentences in Functional Centering.
- Feature-Rich ML: Christoph Müller and Michael Strube later developed machine learning systems that used decision trees to handle the "messiness" of spoken dialogue, where "it" and "that" are ubiquitous.
- Neural Mention-Ranking: Modern approaches, such as the work by Ana Marasovic and Anette Frank, use Siamese LSTM networks to rank potential antecedents. By generating massive amounts of artificial training data, they allow models to learn the semantic "shape" of an event that might serve as a referent.
Why It Matters: Beyond Simple Coreference
Resolving these abstract links is not just a linguistic curiosity; it is a bottleneck for high-level reasoning. Shachar Mirkin and Ido Dagan showed in Assessing the Role of Discourse References in Entailment Inference that nearly half of all textual entailment inferences require resolving discourse references. If a system cannot link a pronoun to a preceding event, it cannot understand the logic of the document.
Similarly, in Machine Translation, Asaf Yehudai and colleagues have noted that while models are getting better at gender agreement for nouns, they still fail to ground pronouns in preceding actions. This leads to translations that are grammatically correct but logically disconnected from the narrative.
The current frontier, represented by projects like LiDARR and the CODI-CRAC Shared Tasks, involves building end-to-end neural models that treat discourse deixis as a fundamental part of document structure. By moving from "word-to-word" matching to "word-to-proposition" mapping, researchers are attempting to give AI the ability to track not just who is being talked about, but what is happening.