Senseannotation

Senseannotation refers to the labeling of words or multiword expressions with a specific sense from a defined sense inventory in a given context. It is a core task in semantic annotation and underpins many natural language processing applications by making meaning explicit for disambiguation purposes.

Sense inventories provide the reference for labeling. WordNet synsets, BabelNet, FrameNet, OntoNotes senses, and domain-specific lexicons are commonly used, though the granularity and scope of senses can vary between inventories. Because inventories differ in coverage and detail, senseannotation projects often map annotations to a chosen inventory or allow cross-inventory alignment.

Annotation can be performed manually or automatically. Manual annotation relies on guidelines and training to ensure consistency across annotators, with inter-annotator agreement (often measured by kappa) used to assess reliability. Automatic approaches range from knowledge-based methods, such as the Lesk algorithm that uses dictionary definitions and overlaps, to supervised and semi-supervised learning that leverage sense-tagged corpora. More recently, neural methods and sense embeddings aim to learn representations for senses directly from data, sometimes integrating multilingual or multimodal information.

Data for senseannotation typically consist of corpora annotated at word-level or multiword expressions with sense identifiers. Such corpora enable evaluation of word sense disambiguation systems and related tasks, and they support downstream applications including information retrieval, machine translation, question answering, and text understanding.

Common challenges include varying sense granularity, data sparsity for rare senses, handling polysemy and idioms, and cross-domain or cross-linguistic transfer. Ongoing work seeks better sense inventories, scalable annotation frameworks, and unified evaluation standards to improve cross-domain applicability and multilingual coverage.