Political science scholarship increasingly explores innovative data and methods to capture positional heterogeneity within political parties and parliamentary groups. Legislators’ social media profiles constitute a source of unfiltered communication, allowing legislators to strategically deviate from their party line. This strategic diversification of policy supply is particularly important against the background of increasing geo-political polarization: In SMD or mixed member systems, MPs and candidates face ever greater incentives to selectively appeal to the specific needs of their local constituents. Capturing these specific appeals, however, requires that we accurately identify mentions and positional statements on specific policy domains from a large stream of textual data. In this paper, we therefore present a new methodological approach for estimating issue-specific legislator communication from Tweets with respect to both salience and positional strategies. Toward this end, we introduce a two-step approach to context-sensitive policy classification via the machine learning model BERT. Via contextualized and ordered representations of text, BERT is capable of capturing nuanced semantic characteristics. Based on parsimonious manual annotations, this approach sequentially handles the classification of tweets as belonging to specific policy domains (salience) and the classification of positional connotations within these domains. We then illustrate the merits of focusing on the diversification of salience and positional strategies at the legislator-level by focusing on the issue domain of housing in Germany -- a country with pronounced socio-economic geographic heterogeneity and a mixed-member PR system in which most MPs compete for personal votes in specific electoral districts.