Referring expression generation: what can we learn from human data?

In this paper, we reflect on what we can learn about the processes involved in the generation of referring expressions by looking at a corpus of human-produced data. We find that the data vastly underspecifies what might be involved algorithmically, but it does rule out a number of popular algorithms for referring expression generation as candidates for models of what people do. We posit an alternative algorithmic schema which forces us to focus more clearly on the questions that need to be answered experimentally if we are to develop algorithms that emulate human behaviour on this task.