In a step toward creating a robust natural language understanding system which detects and avoids miscommunication, this artificial intelligence research report provides a taxonomy of miscommunication problems that arise in expert-apprentice dialogues (including misunderstandings, wrong communication, and bad analogies), and proposes a flexible extension of the succeed/fail paradigm to handle reference mistakes. Extended examples of these reference failures are provided. A theory of relaxation (similar to human referent identification processes) for recovering from reference failures is then developed using the representational language "KL-One," which represents real-world objects hierarchically (in the tradition of semantic networks and frames). Rule-based relaxation is described as an integral part of the process whereby people who are asked to identify objects behave in a particular way: by finding candidates, re-trying, and if necessary, giving up and asking for help. The study models the relaxation process and provides a computational model for experimenting with the different parameters of the relaxation process. Extensive examples show how the model handles problems with imprecision and over-specification in a speaker's description using rules in a hierarchical knowledge base. Finally, the relaxation model is shown to be preferable to "closest match" models, which fail because looking for the closest match cannot determine the most salient features of a possible referent. The report indicates that the relaxation model avoids this problem by determining salient features using the kinds of knowledge humans have about language and the physical world. Many computer-language examples and diagrams supplement the discussion. A 6-page reference list is appended. (SKC)