HOLM: Hallucinating Objects with Language Models for Referring Expression Recognition in Partially-Observed Scenes