TL;DR
Prepositions are notoriously tricky in English due to their vast polysemy. This paper introduces a novel, hierarchical taxonomy of "supersenses" for prepositions, aiming to provide a coarse-grained, yet semantically rich, annotation scheme that can be applied broadly and facilitate NLP tasks like WSD.
What Problem Does This Paper Address?
English prepositions are incredibly frequent and exhibit a dizzying range of meanings, making them a significant challenge for natural language processing (NLP). They can indicate spatial, temporal, or causal relationships, but also function as grammatical markers tied to specific verbs or constructions. This polysemy makes tasks like Word Sense Disambiguation (WSD) for prepositions difficult. Existing resources often focus on specific prepositions or are too fine-grained, making them impractical for broad-coverage annotation. The authors aim to create a general-purpose, coarse-grained taxonomy of preposition functions (supersenses) that is both semantically informative and computationally tractable for annotation.
Methods
The core of this paper is the development of a hierarchical taxonomy of preposition supersenses. This taxonomy is built upon existing work, particularly the Srikumar and Roth (S&R) inventory, but is significantly revised and extended to improve descriptive power and compatibility with other semantic resources like VerbNet and AMR. The hierarchy allows for multiple inheritance, providing a fallback for annotators and a more nuanced representation of prepositional meanings. The authors also define clear criteria for when a preposition token should be annotated with a supersense, excluding those that are part of multiword expressions or have purely syntactic functions. They also refine specific areas of the hierarchy, such as temporal relations and paths, to better capture the complexities of prepositional semantics.
Datasets
The paper primarily discusses the use of the SemEval-2007 shared task data on preposition WSD. This dataset consists of over 25,000 sentences annotated with 34 frequent preposition types and their senses, originally derived from FrameNet and the Oxford Dictionary of English. The authors also mention their own resource, which includes extensive documentation, example sentences, and mappings to other lexical resources, intended for public release.
Key Insights and Ideas
- Hierarchy is Key: A hierarchical structure for preposition supersenses is crucial. It allows for both fine-grained distinctions and broader, more general categories, making the annotation process more manageable and robust.
- Beyond Spatiotemporal: While many preposition studies focus on spatial and temporal uses, this work acknowledges and categorizes the broader functional roles, including causal, instrumental, and abstract relationships.
- Bridging Lexicon and Grammar: Prepositions occupy a unique space between lexicon and grammar. The supersense taxonomy aims to capture this by providing semantic labels that are useful for both lexical and grammatical analyses.
- Practical Annotation: The supersense scheme is designed to be "coarse and unlexicalized" to facilitate efficient manual annotation, making it practical for large-scale corpus development.
- Cross-Resource Compatibility: By aligning with existing resources like VerbNet and AMR, the proposed supersense hierarchy can leverage and contribute to a broader semantic annotation ecosystem.
Key Results
The paper presents a detailed hierarchical taxonomy of 73 preposition supersenses, illustrated with a comprehensive figure. They report that 285 out of 332 fine-grained TPP senses have been mapped to one or more supersenses, suggesting broad coverage. Preliminary annotation efforts indicate that the scheme is stable and broadly applicable, with a pilot study showing promising results for annotator agreement. The paper also highlights specific refinements to temporal and path-related supersenses, demonstrating a more nuanced approach to these common prepositional uses.
Limitations and Future Work
The authors acknowledge that the proposed supersense set is larger than inventories for other parts of speech, which might initially seem daunting for annotators. However, they argue that the hierarchical structure mitigates this concern. They also note that while the scheme is intended to be general-purpose, some prepositions are excluded if they are part of multiword expressions or have purely syntactic functions. Future work includes completing the annotation of the full TPP dataset and further investigating the scheme's adaptability to other languages and adposition systems. They also suggest that if the hierarchy proves too unwieldy, finer-grained distinctions could be removed.
Summary updated