Locally least-cost error recovery is a technique for recovering from syntax errors by editing the input string at the point of error detection. An informal description of a parser generator which implements the technique is given. The generator takes as input an extended BNF description of a language together with a set of primitive edit costs and outputs a recursive descent syntax analyser including error recovery. Criteria for assessment of the technique are offered. Using these criteria the technique is assessed with respect to a database of over 100 example programs, and compared with an alternative local error recovery technique, that of follow set error recovery. The conclusion is that locally least-cost error recovery is more effective than follow set error recovery but much less economical in its use of storage space. The least-cost parser also runs between 15 and 20% slower than the follow set parser.

~While-least-cost~ error correction is fundamentally important to context-free language processing, it is inefficient w'Qhemdone globally.A locally least-cost repair method has been devised to model error recovery in conventional-~o-mpile]ts.The principles of this recovery technique have inspired a practical LL(1) method and should be of value m ~ntmg-error~rJel~a~lr m other parsing algorithms.-At each point in the syntax analysis, a locally optimal repair of't~e-next-sy-mBffl is defined to be a string w such that w can be parsed without error and such that editing the symbol following w can be achieved at least cost, costs being defined by the Wagner-Fischer model of string-to-string correction.In this paper we describe how error recovery can be achieved in Earley's algorithm by simulating locally optimal repairs.The main result is to show that the complexity of Earley's algorithm is not affected by this process; that is, for any input string of length n, the work involved is O(n) if the given grammar is deterministic, O(n 2) if the grammar is unambiguous, and O(n 3) in the worst case.Key Words and Phrases: error repair,