Expectation-Driven Treatment of Difficult Referring Expressions

Like many difficult linguistic phenomena, so-called “broad referring expressions” (BREs) – such as pronominal this, that and it – have been excluded from the purview of most natural language processing systems, being tacitly deemed unmanageably difficult. However, when building cognitively-inspired intelligent agents that are meant to have real-world utility in human-agent teams, the wholesale exclusion of difficult phenomena is neither practical nor necessary. I suggest the following strategy for incorporating the treatment of difficult language phenomena into an agent’s repertoire over time. Agents are configured to automatically determine which instances of a linguistic phenomenon they can and cannot confidently treat. For the high-confidence cases, the agents carry out the language understanding (i.e., “perception”) and move on to decision making and action; for the low-confidence cases, they seek clarification from their human collaborators. This paper details some strategies for resolving BREs that appear to offer high confidence solutions within the current state of the art. The analysis of BREs is distributed across language processing modules in a way inspired by principles of cognitive modeling. The data analysis and modeling strategy show that a natural language processing problem that seems impenetrable when viewed from the current mainstream perspective of supervised machine learning becomes more manageable when modeled according to human-like reasoning.