Automatic symbol recognition can currently be applied to a wide range of technical documents (music scores, electrical diagrams, etc.). Automatizing this kind of task can help some professionals save time (Lu et al., 2005), but it can also generate a loss of situation awareness among operators (Endsley, 1995). Furthermore, as these systems sometimes interpret symbols incorrectly (Lu, Tai, Su, & Cai, 2005; Orbay & Kara, 2011), humans need to be able to intervene in the process to correct any errors as and when they arise. The objective of this thesis was to improve our understanding of the human-machine interactions elicited by this kind of task, in order to determine how best to enhance system performance. We conducted a series of studies during the development of a user-centred design (ISO 9241-210, 2010) of automatic interpretation software for architectural plans. These studies were set within a heterogeneous theoretical framework bringing together cognitive psychology, human-machine interactions and the ergonomics of the multimedia documents. The results they yielded led to several recommendations for designing recognition systems. For instance, superimposing the interpretation on the interpreted document makes it easier to cross-check the information. Furthermore, allowing users to monitor the interpretation process in real time and interrupt it at any point increases performances. Finally, while the highlighting of potential errors can also be beneficial, it can also make users over-confident and thus less watchful
