In the frame of using models to promote dialogue among stakeholders, we are interested in modelling representations as a mean to share viewpoints and opinions. A first trend for modelling representations uses socio-cognitive theories as frameworks for modelling. This article proposes a different approach combining Knowledge Engineering elicitation techniques and Agent Based Modelling. The use of elicitation techniques provides a methodological framework to capture representations directly from stakeholders' discourse. Agent Based Modelling serves as a modelling structure which enables managing representations' heterogeneity and conflicting opinions. The proposed methodology is detailed and applied to farmers' representations of runoff and erosion in south of France. To achieve the construction of a running model of the interviewees' shared representation, it was necessary to implement theoretical data supplementing the elicited information. We argue that companion modelling approach can help sharing and discussing those non-elicited information with stakeholders.