Disambiguating entity references within an ontological model

In our everyday conversations, entities (persons, companies, etc.) are referred to by natural language identifiers (NLIs). Humans employ personal experience and situational context to interpret such identifiers. However, due to ambiguity, even humans run the risk of misinterpretations. In our prior work, we presented a novel method to resolve entity references in texts under the aspect of ambiguity. We explore ontological background knowledge represented in an RDF(S) graph. The different interpretation possibilities lead to different subgraphs of the underlying ontology, each subgraph describing one consistent, non-ambiguous interpretation of the ambiguous NLIs within the ontological knowledge base. Our domain-independent approach is based on spreading activation and uses a semantic relational ranking. In this paper, we suggest three extensions to our original algorithm. First, we process in a two-step interpretation---instead of the whole original input text---at first hand smaller text windows in order to get more precise reference interpretations through a smaller local text context. Second, we extend the spreading-activation algorithm within the RDF(S) graph towards a bidirectional exploration of edges which shall speed-up the algorithm. Third we use reinforcement learning in order to take advantage of re-occurring information. We present first experimental results with these algorithmic extensions and derive directions for future work.