Exploring graph topology via matrix factorization to improve wikification

Abstract

Challenges faced by authors and reviewers of encyclopedic content include the identification of which candidate concepts in one article should be linked to other articles, since concepts should be selected if they are able to provide a deeper understanding of the article topic, and ambiguity resolution. This task is called wikification. Wikification is normally addressed as a classification problem. Many supervised and semi-supervised techniques have been proposed to deal with wikification: their aim is to learn, from examples of concepts, which instances should be wikified. We observed that, although one encyclopedia can be viewed as a concept graph, supervised techniques seldom explore the richness of information provided by the graph topology: they normally only take advantage of statistical features related to concepts and the association among them. In this work, we address wikification as the problem of predicting the edges of a graph. Unlike previous approaches, our model fully explores the graph topology in the prediction task by using a matrix factorization technique. We also include in our model features related to concepts and its associations. Upon experimenting our approach with a sample of Wikipedia, we achieved gains up to 20% over a state-of-the-art baseline.

A wikification prediction model based on the combination of latent, dyadic, and monadic features

“…In our previous work (Ferreira, Pimentel, & Cristo, ) we proposed a prediction model which uses latent features to predict links in Wikipedia. In that work we reported the gains obtained over a state‐of‐the‐art wikification (Milne & Witten, ) approach by incorporating latent features derived from the Wikipedia underlying graph topology.…”

A wikification prediction model based on the combination of latent, dyadic and monadic features

“…Considering the important results by Menon and Elkan (2011) in terms of, for instance, scalability and appropriateness for imbalanced supervised tasks, in our previous work we extended their model for wikification (Ferreira et al, 2015). One novelty of our proposed model is the handling of directed graphs which we achieve by including a latent predictor composed of two latent components: one to capture the undirected aspect of the link, and another to capture the residual directional aspect of the link.…”

“…In this section, we summarize the link prediction model detailed in our previous work (Ferreira et al, 2015).…”

“…Thus, the relationship between i and j may be modelled by associating latent features of each node taking the dot-product such that W ij % U > i KU j ; U i and U j 2 R k ; K 2 R k3k . However, as we pointed in (Ferreira et al, 2015), a predictor based on simple combination of such matrix component and the features associated with (i, j), and i and j need to be extended to support the directed aspect of Wikipedia graph because the matrix component uses the same matrix of latent features (U i and U j ) to capture inlink and outlink behavior.…”

“…In our previous work (Ferreira, Pimentel, & Cristo, 2015) we proposed a prediction model which uses latent features to predict links in Wikipedia. In that work we reported the gains obtained over a state-of-the-art wikification (Milne & Witten, 2013) approach by incorporating latent features derived from the Wikipedia underlying graph topology.…”