This paper concerns knowledge extraction for applications concerning the automated filling of templates from an input of semi-structured textual documents. The template filling task can be viewed as a collaboration between a number of agents, including NE-Agents that are specialised to detect occurrences of specific features in the text and TE-Agents that specialise at combining the results from multiple NE-Agents in order to create a template instance. This paper presents an automated learning approach for the generation of a TE-Agent that extracts spatial relationships between the various features of a template. It is shown that this TE-Agent can compensate for imprecise performance on the part of the NE-Agents.