Embodiments of a system and method for identifying and correcting marginal false positives in machine learning models may include, based on reference data that includes pairs of information items and labels indicating whether pairs of information items have a specific relationship, generating a first machine learning model for determining whether pairs of information items have that relationship. Embodiments may include identifying one or more false positive pairs (e.g., a pair of information items that the first machine learning model indicates as having the specific relationship and which are labeled within the reference data as not having that relationship). Embodiments may include selecting identified false positive pairs as candidates for correction. Embodiments may include, subsequent to a correction of the reference data associated with the selected false positives, generating based on the corrected reference data a new machine learning model for determining whether pairs of information items have the specific relationship.