This research is aimed at developing a hierarchical alternation-based lexical architecture for machine translation. The proposed architecture makes extensive use of information sharing in describing valency frames through derivational links from base frames, rather than as independent entities. This has advantages in descriptive efficiency, robustness and maintainability. The lexicon being developed is built up automatically from the Japanese component of an existing Japanese-English machine translation lexicon. The reconstruction process consists of analysing consistencies in selectional restrictions between valency frames, and postulating alternations where selectional restrictions are preserved on matching case slots; this was found to perform at 60.9% accuracy. All alternation candidates are incorporated into the final-version lexicon as derivational links, and expanded out at run time.