Coordination—the linking of linguistic units using conjunctions like and or or—is one of the most ubiquitous yet deceptively complex phenomena in human language. As noted by Qing Wang et al. in CoRec: An Easy Approach for Coordination Recognition, coordination structures appear in nearly half of all standard sentences. Yet, identifying where one item in a list ends and another begins remains a notorious challenge for syntactic parsers.
Consider the phrase organic apples and oranges. Does organic modify only apples, or does its scope extend to oranges as well? While humans resolve this instantly using context, automated systems face a profound structural ambiguity.
The Two Dimensions of Coordination Ambiguity
To understand why coordination is so difficult to parse, we must separate the problem into two distinct structural dimensions: boundary identification and scope resolution.
1. Boundary Ambiguity
Boundary ambiguity occurs when a parser cannot determine the precise start and end points of the phrases being joined. In A Neural Network for Coordination Boundary Prediction, the authors demonstrate that the lack of explicit structural markers often leads to multiple valid-looking syntactic interpretations.
This issue is particularly acute in complex sentences or conjoined fields. Without deep semantic or lexical knowledge, shallow parsing models frequently struggle to correctly assign phrases to the appropriate clause within a complex sentence, as shown by Markus C. Becker et al. in A Stochastic Topological Parser for German.
2. Scope and Modifier Ambiguity
Scope ambiguity arises when a modifier precedes or follows a coordinated group, leaving its attachment unclear. Deirdre Hogan frames this in Coordinate Noun Phrase Disambiguation in a Generative Parsing Model as a fundamental bracketing problem: does the modifier apply to a single conjunct or the entire coordinated phrase?
When shared modifiers lack explicit structural annotation, parsers routinely fail to determine the correct argument scope, a problem addressed directly by Jessica Ficler in Coordination Annotation Extension in the Penn Tree Bank.
Why Traditional Parsers Fail
Standard parsing architectures fail to resolve coordination because their core mathematical and structural assumptions conflict with the nature of coordinate syntax.
The Dependency Grammar Dilemma
Dependency parsing relies on establishing hierarchical, asymmetric head-dependent relationships (where one word governs another). However, coordinate structures like A and B are inherently flat and symmetric.
As Adam Przepiórkowski et al. explain in Conjunct Lengths in English, Dependency Length Minimization, and Dependency Structure of Coordination, linguistic theories conflict on how to assign a primary head in these constructions—whether it should be the conjunction, the first conjunct, or both. This lack of theoretical consensus leads to highly inconsistent annotation schemes across corpora (Who Did What to Whom? A Contrastive Study of Syntacto-Semantic Dependencies), making coordination a primary failure mode for unsupervised dependency models (Unsupervised Induction of Tree Substitution Grammars for Dependency Parsing).
The Limitations of Linear Processing
Sequence labeling models process text token-by-token in a linear fashion. In Exploring the boundaries: gene and protein identification in biomedical text, Jenny Rose Finkel et al. show that linear processing frequently misses the structural relationships between shared head nouns and distant modifiers. When multiple descriptors share a single head noun, sequence models fail to capture the long-distance dependencies required to reconstruct the full entity.
Mechanisms for Disambiguation
To move past these limitations, modern approaches leverage specific linguistic insights to resolve coordination structurally and semantically.
Leveraging Structural Symmetry
Coordinated elements often exhibit strong structural and semantic parallelism. Parsers can exploit this symmetry to narrow down valid interpretations. For instance, Deirdre Hogan integrated structural symmetry features directly into a generative framework to reduce coordination errors.
Alternatively, systems can use language models to score the "naturalness" of potential sentence rephrasings. As demonstrated by Philip V. Ogren in Improving Syntactic Coordination Resolution using Language Modeling, evaluating whether a rephrased conjunct sounds fluent allows the system to validate or reject ambiguous boundaries.
Integrating Semantic Frames and Lexical Preferences
Purely lexicalized or syntactic models often produce multiple valid semantic structures for a single sequence of categories (A* CCG Parsing with a Supertag-factored Model). To break this tie, parsers must incorporate semantic knowledge.
Daisuke Kawahara et al. developed a model in Probabilistic Coordination Disambiguation in a Fully-Lexicalized Japanese Parser that jointly considers syntactic similarity and lexical preferences from large-scale case frames. By understanding the real-world meaning of the conjuncts, the model can rule out logically impossible or structurally weak interpretations that a purely syntactic parser might accept (Boeing's NLP System and the Challenges of Semantic Representation).
Downstream Impact
When coordination resolution fails, the errors cascade into downstream applications, corrupting the extraction of structured knowledge.
- Information Extraction: In specialized domains like biology, failing to resolve coordinate structures leads to incomplete candidate generation and broken entity boundaries (Enhancing automatic term recognition through recognition of variation). It also causes incomplete relation extraction; for example, Massimiliano Ciaramita et al. note in Unsupervised Learning of Semantic Relations between Concepts of a Molecular Biology Ontology that parsers often associate only the first element of a coordinated list with a shared verb, completely missing subsequent relations.
- Text-to-SQL Conversion: Modern semantic parsers that translate natural language questions into database queries struggle to group multiple attributes or columns within a single clause (such as a
SELECTstatement). Without explicit syntactic coordination recognition, as argued by Binyuan Hui et al. in S²SQL: Injecting Syntax to Question-Schema Interaction Graph Encoder for Text-to-SQL Parsers, the system fails to realize that two nouns are part of the same logical operation.
Key Foundational Work
Dealing With Conjunctions in a Machine Translation Environment (1984) — contributed 2 concept elements to this direction.
Bundles and Syntactic Analysis (1982) — contributed 1 concept elements to this direction.
Zum Ambiguitätenproblem Bei Der Analyse Koordinativer Verbindungen (1982) — contributed 1 concept elements to this direction.