Gradable Adjective Embedding for Commonsense Knowledge

Abstract

Adjective understanding is crucial for answering qualitative or subjective questions, such as “is New York a big city”, yet not as sufficiently studied as answering factoid questions. Our goal is to project adjectives in the continuous distributional space, which enables to answer not only the qualitative question example above, but also comparative ones, such as “is New York bigger than San Francisco?”. As a basis, we build on the probability P(New York—big city) and P(Boston—big city) observed in Hearst patterns from a large Web corpus (as captured in a probabilistic knowledge base such as Probase). From this base model, we observe that this probability well predicts the graded score of adjective, but only for “head entities” with sufficient observations. However, the observation of a city is scattered to many adjectives – Cities are described with 194 adjectives in Probase, and, on average, only 2% of cities are sufficiently observed in adjective-modified concepts. Our goal is to train a distributional model such that any entity can be associated to any adjective by its distance from the vector of ‘big city’ concept. To overcome sparsity, we learn highly synonymous adjectives, such as big and huge cities, to improve prediction accuracy. We validate our finding with real-word knowledge bases.