Improved error recovery in generated LR parsers