Probabilistic grounding of situated speech using plan recognition and reference resolution