TL;DR
This paper introduces a novel type-logical grammar framework that accounts for "polarity sensitivity" in language, explaining why certain words (like "any" or "ever") can only appear in specific grammatical contexts. By integrating programming language concepts like delimited continuations and evaluation order, the model achieves greater empirical coverage than previous approaches, correctly predicting sentence acceptability based on both semantic and syntactic constraints.
What Problem Does This Paper Address?
Linguists have long been fascinated by "polarity sensitivity," a phenomenon where certain words, known as polarity items (PIs), are restricted to specific environments. For instance, negative polarity items (NPIs) like "any" and "ever" typically appear in negative or downward-entailing contexts (e.g., "Nobody saw anybody"), but not in positive ones (e.g., "Everybody saw anybody"). This restriction is not purely semantic; word order also plays a crucial role. For example, while "Nobody saw anybody" is grammatical, "Anybody saw nobody" is not, even though both sentences involve a negative quantifier and an NPI.
Existing linguistic theories often struggle to capture both the semantic licensing conditions and the syntactic constraints, particularly the interaction between linear order and scope. This paper aims to bridge this gap by proposing a unified, type-logical grammar that accounts for both aspects of polarity sensitivity.
Methods
The authors propose a novel analysis within the framework of multimodal categorial grammar, a type-logical grammar formalism. Their key innovation is the integration of concepts from programming language theory:
- Delimited Continuations: This allows the grammar to model how subexpressions are embedded within specific contexts, akin to how code is evaluated within a delimited scope.
- Evaluation Order: By explicitly considering the order in which linguistic constituents are processed, the model can capture the linear order constraints that affect polarity item licensing.
They introduce three unary modes:
- Value Mode (blank): Represents expressions devoid of in-situ quantification.
- Unquotation Mode (u): Allows for the cancellation of embedded expressions, analogous to
evalin programming. - Polarity Mode (p): This is the core mode for handling polarity sensitivity. They assign specific types to positive and negative clauses (e.g.,
s⁺ands⁻) that are "subtypes" of a neutral clause (s°), allowing for flexible conversion between these polarities.
The framework uses natural deduction rules (Figure 1) and structural postulates (Figure 2) to derive grammatical sentences. The interaction between these modes and rules is designed to capture the complex interplay of semantic scope and linear order.
Datasets
Not discussed in the paper. The paper focuses on theoretical linguistic analysis and uses example sentences rather than formal datasets.
Key Insights and Ideas
- Unified Account of Syntax and Semantics: The framework successfully integrates semantic licensing conditions (e.g., downward entailment) with syntactic constraints (e.g., linear order and c-command) under a single, formal system.
- Programming Language Concepts Illuminate Linguistics: The application of delimited continuations and evaluation order from programming languages provides a powerful new lens for understanding linguistic phenomena like polarity sensitivity and scope.
- Linear Order is Crucial: The model explicitly shows how the linear order of constituents, mediated by the evaluation order of subexpressions, is essential for correctly predicting sentence acceptability, a feature often overlooked in purely scope-based analyses.
- Structural Postulates as Key: The structural postulates, particularly the requirement for certain constituents to be embedded under diamonds (
◇), are shown to be critical in ruling out ungrammatical sentences like "*Anybody saw nobody." - Polarity as a Type Property: Polarity items are characterized by their input and output polarities, which can be visualized as transitions in a finite-state machine, offering a systematic way to assign types and predict interactions.
Key Results
The paper demonstrates that its type-logical framework can:
- Correctly derive grammatical sentences like "Nobody saw anybody."
- Accurately rule out ungrammatical sentences like "*Anybody saw nobody" by leveraging the evaluation order and structural postulates.
- Predict the correct scope relations for quantifiers, accounting for both linear and non-linear scope possibilities.
- Explain why sentences like "Nobody introduced everybody to somebody" have a specific reading, while "Nobody introduced Alice to somebody" does not, based on the interaction of polarity and quantifier types.
Limitations and Future Work
- Decidability: While the authors mention that the parsing problem for their fragment (NL◇r) is decidable using proof-net technology, the full implications for computational tractability of the entire system are not detailed.
- Scope Ambiguity: The paper notes that sentences like "Somebody saw everybody" are correctly predicted to have ambiguous scope, but a deeper exploration of how the framework handles complex scope ambiguities could be beneficial.
- Further Empirical Coverage: The authors suggest that this framework could be extended to analyze other polarity-sensitive phenomena and potentially other linguistic puzzles that involve complex interactions between syntax, semantics, and linear order.
Summary updated