Knowledge Completion for Generics using Guided Tensor Factorization

Knowledge Completion for Generics using Guided Tensor Factorization Hanie Sedghi∗ Google Brain Mountain View, CA, U.S.A. hsedghi@google.com Ashish Sabharwal Allen Institute for Artificial Intelligence (AI2) Seattle, WA, U.S.A. AshishS@allenai.org Abstract Given a knowledge base or KB containing (noisy) facts about common nouns or gener- ics, such as “all trees produce oxygen” or “some animals live in forests”, we consider the problem of inferring additional such facts at a precision similar to that of the starting KB. Such KBs capture general knowledge about the world, and are crucial for various appli- cations such as question answering. Differ- ent from commonly studied named entity KBs such as Freebase, generics KBs involve quan- tification, have more complex underlying reg- ularities, tend to be more incomplete, and vio- late the commonly used locally closed world assumption (LCWA). We show that existing KB completion methods struggle with this new task, and present the first approach that is successful. Our results demonstrate that ex- ternal information, such as relation schemas and entity taxonomies, if used appropriately, can be a surprisingly powerful tool in this set- ting. First, our simple yet effective knowledge guided tensor factorization approach achieves state-of-the-art results on two generics KBs (80% precise) for science, doubling their size at 74%-86% precision. Second, our novel tax- onomy guided, submodular, active learning method for collecting annotations about rare entities (e.g., oriole, a bird) is 6x more effec- tive at inferring further new facts about them than multiple active learning baselines. 1 Introduction We consider the problem of completing a partial knowledge base (KB) containing facts about gener- ∗ This work was done while the author was affiliated with the Allen Institute for Artificial Intelligence. ics or common nouns, represented as a third-order tensor of (source, relation, target) triples, such as (butterfly, pollinate, flower) and (thermometer, mea- sure, temperature). Such facts capture common knowledge that humans have about the world. They are arguably essential for intelligent agents with human-like conversational abilities as well as for specific applications such as question answering. We demonstrate that state-of-the-art KB completion methods perform poorly when faced with gener- ics, while our strategies for incorporating external knowledge as well as obtaining additional annota- tions for rare entities provide the first successful so- lution to this challenging new task. Since generics represent classes of similar indi- viduals, the truth value yi of a generics triple xi = (s, r, t) depends on the quantification semantics one associates with s and t. Indeed, the semantics of generics statements can be ambiguous, even self- contradictory, due to cultural norms. As Leslie (2008) points out, ‘ducks lay eggs’ is generally con- sidered true while ‘ducks are female’, which is true for a broader set of ducks than the former statement, is generally considered false. To avoid deep philosophical issues, we fix a par- ticular mathematical semantics that is especially rel- evant for noisy facts derived automatically from text: associate s with a categorical quantification from {all, some, none} and associate t (implicitly) with some. For instance, …

Content truncated to 500 words