In the context of geo-data infrastructures users may want to combine data from different sources and expect consistent data. If both datasets are maintained separately, different capturing methods and intervals leads to inconsistencies in geometry and semantic, even if the same reality has been modelled. Our project aims to automatically harmonize such datasets and to allow an efficient actualisation of the semantics. The application domain in our project is cadastral and topographic datasets. To resolve geometric conflicts between topographic and cadastral data a local nearest neighbour method was used to identify perpendicular distances between a node in the topographic and an edge in the cadastral dataset. The perpendicular distances are reduced iteratively in a constraint least squares adjustment (LSA) process moving the coordinates from node and edge towards each other. The adjustment result has to be checked for conflicts caused by the movement of the coordinates in the LSA. The correct choice of matching partners has a major influence on the result of the LSA. If wrong matching partners are linked a wrong adaptation is derived. Therefore we present an improved matching method, where we take distance, orientation and semantic similarity of the neighbouring objects into account. Using Machine Learning techniques we obtain corresponding land-use classes. From these a measurement for the semantic distance is derived. It is combined with the orientation difference to generate a matching probability for the two matching candidates. Examples show the benefit of the proposed similarity measure.