Abstract
Prepositions and postpositions (adpositions) are among the most polysemous and frequent elements in language, yet they remain notoriously difficult to categorize in Natural Language Understanding (NLU) tasks. Traditional "supersense" approaches often fail when a preposition's literal meaning conflicts with its contextual role. This paper proposes a novel bipartite annotation framework that separates the semantic "Scene Role" required by a situation from the lexical "Function" contributed by the adposition itself. This decoupling allows for a more nuanced representation of linguistic construal and provides a robust, simplified foundation for cross-lingual semantic analysis.
Introduction: The Challenge of Prepositions
Prepositions are often considered the "red-headed stepchildren" of linguistics—highly frequent but difficult to pin down. For instance, the word at can indicate a location ("at the shop"), a time ("at 7pm"), or even a stimulus ("laughed at my acting"). Previous efforts in NLU attempted to assign a single "supersense" label to each preposition token. However, this approach struggles with linguistic "packaging" or construal. In a sentence like "I was scared by the bear," the bear is contextually the stimulus, but the word by carries an agentive or causal function. Single-label systems often force annotators to choose one at the expense of the other, losing vital semantic information.
The Bipartite Construal Framework
To address these limitations, the authors introduce Bipartite Construal Analysis, a framework rooted in Cognitive Linguistics and Construction Grammar. Instead of a single label, every adposition token receives two distinct categories:
- Scene Role: The semantic relation called for by the governing predicate or the overall situation (the "what").
- Function: The lexical meaning that the adposition itself codes for or brings to the table (the "how").
The relationship is represented using the notation ROLE $\leadsto$ FUNCTION. For example, in "works by Puccini," the Scene Role is CREATOR, while the Function is AGENT. Conversely, in "works of Puccini," the Scene Role remains CREATOR, but the Function is SOURCE. This distinction captures the subtle nuances speakers use to frame the same underlying relationship.
Refining the Supersense Taxonomy
By adopting this bipartite approach, the authors suggest that the existing hierarchy of semantic labels can be significantly streamlined.
As shown in Figure 1, the current taxonomy includes a wide array of specific categories. Previously, complex labels were often created specifically to handle cases where roles and functions overlapped. By decoupling these two layers, many of these hybrid labels can be removed, simplifying the hierarchy for human annotators while maintaining high expressive power.
Cross-Linguistic Application and Datasets
The framework was tested on a 55,000-word corpus of English online reviews containing 4,250 preposition tokens. Beyond English, the authors conducted pilot annotations in Hindi, Hebrew, and Korean. This cross-linguistic testing is crucial because many languages use case markers that naturally conflate multiple semantic roles. For example, the Hindi genitive marker kaa can represent both possession and experience. The bipartite system handles this "overloading" naturally without requiring the creation of hundreds of language-specific labels.
Key Insights and Results
- Construal as a Feature: Mismatches between role and function are not errors; they represent how language expresses perspective. "Scared by the bear" (causal) feels different from "scared about the bear" (topical), even if the bear's underlying role as a stimulus is identical.
- Resolving Annotator Ambiguity: The bipartite approach provides a principled way to resolve "annotator hand-wringing" in cases where a token seems to fit two categories (e.g.,
TOPICvs.STIMULUS). - Framework for NLU: This methodology provides a path for statistical NLU systems to learn how different languages realize semantic roles through specific adpositions, a major hurdle for machine translation and semantic parsing.
Limitations and Future Work
While promising, the bipartite scheme increases the cognitive load on human annotators, who must now provide two labels instead of one. Future work will focus on measuring inter-annotator agreement and exploring how to handle "fossilized" idiomatic uses (e.g., "interested in politics") where the preposition's function may be effectively empty. The ultimate goal is to produce large-scale, multi-lingual datasets to train models that truly understand the subtle dance of prepositions.
Summary updated