Abstract: Here presented is a new approach to NP phrase structure rule acquisition based on multi\|feature similarity estimation from corpora without bracketed and nonterminal labels. By computing the distance between a rule and all feature rules based on their local contextual information, the system could sort all rules by their distances. The smaller the distance, the larger the similarity. Experiments using Berlitz corpus show that the approach presented achieves a relatively high accuracy: 80% in the first 50 rules. This result demonstrates that training data acquisition based on NA assumption is effective for rule acquisition and parsing.