OntoNotes Corpus

OntoNotes provides a unique, integrated annotation stack on the same text, enabling research into the interaction of linguistic phenomena. Layers include:

• Syntactic parsing: Phrase structure trees in the Penn Treebank style.
• Semantic Role Labeling: Predicate-argument structures using PropBank verb frames and VerbNet thematic roles.
• Coreference Resolution: Entity and event coreference chains.
• Named Entity Recognition: 18 entity types, including geopolitical entities, organizations, and facilities.
• Word Sense Disambiguation: Linking open-class words to their WordNet senses.

The corpus spans diverse linguistic domains to ensure models generalize beyond newswire text. Genres include:

• Newswire: Standard Wall Street Journal articles.
• Broadcast News: Transcribed television and radio news.
• Broadcast Conversation: Informal, spontaneous dialogue from talk shows.
• Web Text: Blogs and newsgroups representing informal written language.
• Telephone Speech: Transcribed conversational telephone speech.
• Pivot Text: Old and New Testament text, providing a parallel translation layer.

The English portion of the corpus contains over 1.5 million words of annotated text. The scale provides sufficient training data for deep neural models like BERT-based SRL systems. Key statistics include:

• ~1.5M words of English text.
• ~174k coreference chains.
• ~300k predicate instances for SRL.
• Annotations for Arabic and Chinese are also included, supporting multilingual research.

Unlike earlier corpora, OntoNotes enforces a single, consistent entity ontology across all genres and languages. This unified scheme ensures that a PERSON entity in newswire is defined identically to a PERSON in telephone speech. The 18 types are hierarchical, allowing for coarse-grained (e.g., PERSON) or fine-grained (e.g., PERSON_Artist) tagging, which is critical for training robust Named Entity Recognition systems.

The annotation of coreference chains links mentions across sentences, providing the discourse-level connectivity required for document understanding. This allows models to learn that 'she' in one sentence and 'the CEO' in another refer to the same entity. This layer is foundational for training systems that perform multi-document summarization and question answering over long contexts.

The landmark benchmarking challenge that established OntoNotes as the definitive evaluation standard for end-to-end systems. The task required models to jointly perform coreference resolution, named entity recognition, and semantic role labeling on the same corpus, driving the development of multi-task architectures. The shared task's blind test sets and strict scoring metrics (MUC, B³, CEAF) remain the primary citation benchmark for state-of-the-art coreference systems.

The NLP task of identifying all expressions that refer to the same real-world entity, for which OntoNotes provides the gold-standard training data. The corpus annotates identity chains linking pronouns, definite descriptions, and named mentions across sentence boundaries. Modern neural coreference models (e2e-coref, s2e-coref) are trained and evaluated almost exclusively on OntoNotes, making it the single most critical resource for this task.