Syntactic Constraints for Pronominal Anaphora Resolution

24 concepts15 papers1 proposal

When a natural language processing system encounters a pronoun like him, itself, or it, it faces a massive search space of potential antecedent noun phrases across a discourse. A naive approach might rely entirely on semantic compatibility—evaluating whether a candidate noun fits the verb's expectations (e.g., "can a rock walk?"). However, as Benjamin Heinzerling et al. demonstrate in Revisiting Selectional Preferences for Coreference Resolution, relying purely on fine-grained semantic preferences yields modest performance gains at a high computational cost, leaving significant ambiguity when multiple preceding entities are semantically plausible.

Instead of trying to guess the correct antecedent from a sea of possibilities, constraint-based reference resolution works upside down: it uses the rigid geometry of sentence structure as a powerful negative filter to instantly eliminate impossible candidates.

Syntax as a Negative Filter

The core intuition behind structural pronoun resolution is that syntax provides a set of absolute boundaries. Before a system ever computes complex discourse semantics or pragmatic context, it can use structural rules to prune the search space. As Shalom Lappin et al. demonstrate in A Syntactic Filter on Pronominal Anaphora for Slot Grammar, a syntactic filter applied directly to head-argument structures can quickly rule out impossible antecedents, leaving a much smaller, cleaner set of candidates for downstream semantic processing.

The foundational framework for these rules is often drawn from Noam Chomsky’s Government-Binding (GB) Theory, specifically Binding Theory. As Alessandra Giorgi et al. emphasize in A Computational Approach to Binding Theory, the primary function of Binding Theory is not to output definitive referential assignments, but rather to restrict them by identifying grammatical violations.

By formalizing Principles A, B, and C of Binding Theory into computational algorithms, systems can systematically rule out invalid antecedents based on the linguistic domain they occupy. For example, Robert J. Kuhns in A PROLOG Implementation of Government-Binding Theory and Roland Stuckardt in Anaphor Resolution and the Scope of Syntactic Constraints show that applying these domain-specific constraints provides a robust, efficient framework for identifying valid pronoun-antecedent pairs, far outperforming approaches that rely solely on shallow semantic information.

The Geometry of Coreference

Why does syntax succeed where simple word proximity fails? Because anaphoric binding is governed by structural domains and tree geometry rather than linear distance. As Mary Dalrymple et al. show in Modeling syntactic constraints on anaphoric binding, an anaphor's grammatical structure dictates the valid locations and types of its potential antecedents across long-distance dependencies.

Consider the classic contrast:

  • John saw himself. (Obligatory coreference: himself must refer to John)
  • John saw him. (Obligatory non-coreference: him cannot refer to John)

To model these relationships computationally, systems must look at tree configurations like c-command (where a node dominates its siblings and their descendants). Manfred Pinkal argues in On the Syntactic-Semantic Analysis of Bound Anaphora that phenomena like weak crossover prove purely semantic rules are insufficient; c-command constraints are strictly necessary to explain why a pronoun cannot be bound by a noun phrase in a syntactically higher position, even when semantically plausible.

Similarly, Marcos Garcia et al. in An Entity-Centric Coreference Resolution System for Person Entities with Rich Linguistic Information integrate morpho-syntactic constraints to prevent incorrect links in complex sentences, explicitly blocking object pronouns from coreferring with subjects when the structural configuration forbids it.

       S
      / \
    NP   VP
    |   /  \
  John V    NP
       |    |
      saw  him (Cannot be John due to local binding domain rules)

Before deep structural trees are even evaluated, surface-level morpho-syntactic features act as immediate, low-cost constraints. Barbara Dunin-Keplicz in Towards Better Understanding of Anaphora outlines how number, gender, and case agreement establish a strict "area of reference," categorizing potential links into obligatory coreference, obligatory non-coreference, and admissible coreference before semantic interpretation begins.

Operationalizing the Constraints

How do we embed these rules into actual language processing systems? Historically, researchers have approached this in two ways:

  1. Post-Grammatical Filters: Parsing the sentence into a tree structure first, and then applying structural filters to rule out invalid pronoun-NP pairs (e.g., Lappin et al., 1990).
  2. Active Binding Machines: Instead of treating constraints as a separate step after parsing, António Branco in Binding Constraints as Instructions of Binding Machines conceptualizes nominals as active "binding machines." In this view, pronouns and reflexives actively execute structural instructions and contribute to semantic interpretation during the parsing process itself, integrating constraints directly into grammatical processing.

Beyond Simple Sentences

These structural constraints extend to complex discourse environments, such as relative clauses, zero pronouns, and text generation:

  • Relative Clauses and Focus: Jose Abracos et al. in Extending DRT With a Focusing Mechanism for Pronominal Anaphora and Ellipsis Resolution introduce specialized rules for handling referent accessibility within restrictive relative clauses, showing that pronouns in agent positions behave differently from non-agent positions. Even focus constructions like clefts ("It was X that Y") are bound by syntax; J. L. Delin in A Multi-Level Account of Cleft Constructions in Discourse shows how syntactic constraints limit which constituents can serve as the head of a cleft.
  • Zero Pronouns: In languages where pronouns are frequently omitted (pro-drop), structural constraints are vital. Hiromi Nakaiwa et al. in Anaphora Resolution of Japanese Zero Pronouns with Deictic Reference combine verbal semantics and conjunction types to resolve omitted subjects and objects, showing how clause-linking structures dictate pronoun case and reference.
  • Natural Language Generation: Constraints are equally critical when generating text to avoid creating ambiguous or ungrammatical sentences. In Italian text generation, Fiammetta Namer in Subject Erasing and Pronominalization in Italian Text Generation demonstrates a "no-coreferentiality" rule: a system cannot freely erase a subject pronoun if a complement pronoun is coreferential with it, because doing so would necessitate a reflexive pronoun to remain grammatically correct.

Ultimately, as Hoifung Poon et al. note in Joint Unsupervised Coreference Resolution with Markov Logic, coreference models often falter when they lack deep linguistic knowledge of structural constraints. By leveraging syntax as a strict, geometric filter, systems can bypass the noise of purely statistical or semantic matching, navigating the boundaries of reference exactly as human grammar dictates.

Key Foundational Work

Towards Better Understanding of Anaphora (1983) — contributed 3 concept elements to this direction.
A PROLOG Implementation of Government-Binding Theory (1986) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_constraint_based_reference_resolution_266afd3c,
  title = {constraint-based reference resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/constraint-based-reference-resolution/rdr_266afd3c16b94b37a3cdee5e72287484},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_constraint_based_reference_resolution_266afd3c_v1,
  title = {constraint-based reference resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/constraint-based-reference-resolution/rdr_266afd3c16b94b37a3cdee5e72287484/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake