This paper presents an algorithm, RAP (Resolution of Anaphora Procedure), for identifying the noun phrase antecedents of third-person pronouns and lexical anaphors (reflexives and reciprocals). RAP is implemented in Prolog and relies on salience measures derived from syntactic structure and a dynamic model of attentional state. The algorithm has been extensively tested on computer manual texts and a blind test was conducted on a corpus containing 360 pronoun occurrences. RAP successfully identified the antecedent for 86% of these occurrences. The paper examines the relative contributions of the algorithm's components to its overall success rate and discusses an enhancement that incorporates statistically modeled semantic and real-world relations. This enhancement marginally improves the algorithm's performance by 2%. RAP is compared with other anaphora resolution approaches, including Hobbs' algorithm and centering approaches, showing higher success rates than Hobbs' algorithm. The paper also explores the relationship between RAP and models that invoke various informational factors in ranking antecedent candidates.This paper presents an algorithm, RAP (Resolution of Anaphora Procedure), for identifying the noun phrase antecedents of third-person pronouns and lexical anaphors (reflexives and reciprocals). RAP is implemented in Prolog and relies on salience measures derived from syntactic structure and a dynamic model of attentional state. The algorithm has been extensively tested on computer manual texts and a blind test was conducted on a corpus containing 360 pronoun occurrences. RAP successfully identified the antecedent for 86% of these occurrences. The paper examines the relative contributions of the algorithm's components to its overall success rate and discusses an enhancement that incorporates statistically modeled semantic and real-world relations. This enhancement marginally improves the algorithm's performance by 2%. RAP is compared with other anaphora resolution approaches, including Hobbs' algorithm and centering approaches, showing higher success rates than Hobbs' algorithm. The paper also explores the relationship between RAP and models that invoke various informational factors in ranking antecedent candidates.