Anaphora resolution (also known as coreference resolution) is the NLP task of determining which earlier noun or entity a pronoun or referring expression points back to in a text. It is essential for understanding natural language where speakers constantly use pronouns and references to avoid repetition.

Examples

• "TSMC announced new capacity. They will invest $40B." → "They" = TSMC
• "The wafer was processed, but it had defects." → "it" = the wafer
• "Jensen Huang said NVIDIA will release a new chip. He also mentioned the company is expanding." → "He" = Jensen Huang, "the company" = NVIDIA

Types of Anaphora

• Pronominal: Pronouns like he, she, it, they, them referring back to previously mentioned entities.
• Definite Noun Phrases: "the company," "the chip," "the process" referring to a specific previously mentioned entity.
• Demonstratives: "this approach," "that technology," "these results" pointing to prior concepts.
• Zero Anaphora: Implicit references where the referent is omitted entirely (common in some languages and informal text).

Modern Approaches

• Neural Coreference: End-to-end models (like Lee et al., 2017) that score all possible mention spans and their pairwise links, selecting the best coreference clusters.
• SpanBERT-based: Fine-tuning pretrained transformers on coreference data achieves strong results on benchmarks like OntoNotes.
• LLM In-Context: Large language models can perform coreference resolution through prompting, though dedicated models remain more reliable for structured outputs.

Why It Matters

Anaphora resolution is critical for dialogue systems, information extraction, machine translation, text summarization, and any NLP task where understanding who or what is being discussed depends on resolving references across sentences.