Learning to Collectively Link Entities

Recently Kulkarni et al. [20] proposed an approach for collective disambiguation of entity mentions occurring in natural language text. Their model achieves disambiguation by efficiently computing exact MAP inference in a binary labeled Markov Random Field. Here, we build on their disambiguation model and propose an approach to jointly learn the node and edge parameters of such a model. We use a max margin framework, which is efficiently implemented using projected subgradient, for collective learning. We leverage this in an online and interactive annotation system which incrementally trains the model as data gets curated progressively. We demonstrate the usefulness of our system by manually completing annotations for a subset of the Wikipedia collection. We have made this data publicly available. Evaluation shows that learning helps and our system performs better than several other systems including that of Kulkarni et al.