Reference and Reference Failures. Technical Report No. 398
In order to build robust natural language processing systems that can detect and recover from miscommunication, the investigation of how people communicate and how they recover from problems in communication described in this artificial intelligence report focused on reference problems which a listener may have in determining what or whom a speaker is talking about. The paper first details the investigation, in which protocols of subjects communicating about a task were analyzed, and knowledge about linguistic and physical context that listeners used to recover from reference miscommunication was isolated. The report then discusses the algorithms designed to apply that knowledge to identify the most likely places for the speaker's error. The paper describes how computer programs were then written (1) to represent a spatially complex physical world, (2) to manipulate that structure to reflect the changes caused by the listener's interpretation of the speaker's utterances, (3) to identify referent noun phrases, and, when that fails, (4) to search the physical world for reasonable candidates for the referent. The report proceeds with the development of an algorithm–FWIM (Find What I Mean)–that uses knowledge sources to guide relaxation techniques that delete or replace potentially misleading portions of the speaker's description. Finally, the paper presents a case study, including protocols of two subjects assembling a “toy water pump” to highlight the complexity of reference identification in a task-oriented domain. Throughout, the paper asserts that the new computational model was designed to allow a speaker leeway in forming an utterance about a task, in determining how to deliver it, and also in promoting a new view for extensional reference. A comprehensive list of references is appended. (NKA)
Goodman, B.A. Reference and Reference Failures. Technical Report No. 398. Retrieved March 20, 2019 from https://www.learntechlib.org/p/138067/.