Inferring Metaphoric Structure from Financial Articles Using Bayesian Sparse Models

Drawing from a large corpus (17,000+ articles) of financial news, we perform a Bayesian sparse model analysis of the argument-distributions of the UP and DOWN-verbs, used to describe movements in indices, stocks and shares. Previous work, by Gerow and Keane (2011a, 2011b, 2011c), has shown, using measures of overlap and k-means clustering, that metaphor hierarchies and antonymic relations can be found in this data; for instance, UP verbs have rise as a superordinate organizing a distinct set of subordinate verbs (soar, jump, climb, surge, rebound, advance). This work empirically realizes theories about the structuring of our conceptual systems with metaphors (Lakoff, 1992; Lakoff & Johnson, 1980) but does so using a distributional approach to meaning; namely, that words that occur in similar contexts have similar meanings (see Wittgenstein, 1953). However, Gerow and Keane’s analysis does not show the overall structure of how these metaphors semantically relate to one another. In the present paper, we re-analyzed their data using a Bayesian sparse model (Lake & Tenenbaum, 2010) in order to infer this metaphor space as a uniform representation, based on the argument distributions. Therefore, we treated arguments as features of metaphors. Our model learned three dimensional graphs in an unsupervised manner as sparse representations of the metaphoric structure over all argument distributions, in parallel. Doing so, it also successfully indicates the metaphoric hierarchies and antonymy relations, that were found by the previous models. In conclusion, we discuss the benefits of this approach.