Argument mining is a field in natural language processing that studies the automatic extraction of arguments and the classification of their structure from free text. This is a promising research area with numerous applications, like fact checking, qualitative assessments of online debates or analysis of legal documents, but with a set of non trivial of challenges to overcome, the main one being the lack of large language resources for the development of this kind of models. In this work we propose a research path on a subarea of argument mining that has not been as explored as some other ones: semi-supervised argument mining. We will explore the adaptation of DISPUTool, a model originally trained on data from the US Presidential Elections Political Debates, to deliberative democracy debates within the ORBIS Project for ``augmenting participation, co-creation, trust and transparency in deliberative democracy at all scales''. Our discussion highlights the technical and non-technical challenges of this task and our planned path of action to overcome them.