Graph-Based Lexicon Expansion with Sparsity-Inducing Penalties

摘要: We present novel methods to construct compact natural language lexicons within a graph-based semi-supervised learning framework, an attractive platform suited for propagating soft labels onto new types from seed data. To achieve compactness, we induce sparse measures at graph vertices by incorporating sparsity-inducing penalties in Gaussian and entropic pairwise Markov networks constructed labeled unlabeled Sparse are desirable high-dimensional multi-class problems such as the induction of on types, which typically associate with only few labels. Compared standard methods, two lexicon expansion problems, our approach produces significantly smaller obtains better predictive performance.