Prior Probability

A specific type of prior probability calculated from large-scale hyperlink statistics. Commonness quantifies how frequently a surface form is used as the primary anchor text for a particular entity in a corpus like Wikipedia.

• Derived from counting (mention, entity) pairs in hyperlink graphs
• Serves as a powerful baseline for candidate ranking
• Example: The surface form "Paris" links to the French capital ~90% of the time in Wikipedia, giving it high commonness for that entity

A dynamic measure of semantic relatedness between the text surrounding a mention and the descriptive text of a candidate entity. Unlike prior probability, contextual similarity is computed at inference time for each specific occurrence.

• Uses dense entity embeddings to compare mention context with entity descriptions
• Resolves cases where commonness alone would fail
• Example: "Paris" in a sentence about Greek mythology requires contextual similarity to correctly link to the Trojan prince rather than the city

The core process of resolving the correct identity of an ambiguous entity mention by combining prior probability with contextual signals. Modern systems use a two-stage architecture:

• Candidate Generation: Prior probability filters the knowledge base to a manageable candidate set
• Candidate Ranking: Contextual similarity and coherence features select the final entity
• The prior acts as a strong Bayesian prior that contextual evidence must overcome

The initial retrieval step that uses prior probability to narrow billions of knowledge base entities down to a shortlist of plausible candidates for a given surface form.

• Builds a surface form dictionary mapping mentions to top-N entities by prior
• Dramatically reduces the search space for downstream ranking
• Typical candidate sets contain 5-20 entities per mention
• Systems like BLINK precompute these dictionaries from Wikipedia and Crosswikis data

A dense, low-dimensional vector representation of a knowledge base entity that captures its semantic properties for efficient similarity computation. Prior probability complements embeddings by providing a frequency-based signal that embeddings alone may miss.

• Learned through contrastive representation learning on entity descriptions
• Enables fast dot-product scoring in Bi-Encoder architectures
• Prior probability can be incorporated as an additional feature in the final scoring layer

A numerical value between 0 and 1 representing the system's certainty in a specific prediction. This score typically integrates prior probability with contextual and coherence signals.

• Used for threshold tuning to balance precision and recall
• Critical for NIL prediction: mentions scoring below threshold are classified as unlinkable
• Calibrated using Platt scaling or isotonic regression on validation data
• Example: A mention with high prior but low contextual similarity may receive a moderate confidence score