Summary: A Simple Approach to Case-Based Reasoning in Knowledge Bases

Authors: Rajarshi Das1 RAJARSHI, Ameya Godbole, Shehzaad Dhuliawala, Manzil Zaheer, Andrew McCallum

The summary below was automatically created by Flexudy. Feel free to download the app from your play store.

Have fun reading.

Our non-parametric approach derives crisp logical rules for each query by finding multiple graph path patterns that connect similar source entities through the given relation. We also demonstrate that our model is robust in low data settings, outperforming recently proposed meta-learning approaches1. Introduction Given a new problem, humans possess the innate ability to ‘retrieve’ and ‘adapt’ solutions to similar problems from the past.

For example, an automobile mechanic might fix a car engine by recalling previous experiences where cars exhibited similar symptoms of damage. At a high level, a case-based reasoning system is comprised of four steps [Aamodt and Plaza, 1994] A ‘case’ is usually associated with a problem description (used for matching it to a new problem) and its corresponding solution.

Even though a lot of facts about MELINDA is captured, it is missing the edge corresponding to works in city. However, the number of paths starting from an entity increases exponentially w.r.t the path length and therefore past work used parametric models to do approximate search using reinforcement learning (RL) , 2018]. Using RL-based methods have their own shortcomings, like hard to train and high computational requirements.

Moreover, these models try to encode all the rules for reasoning into the parameters of the model which makes learning even harder. The retrieved entities could be present anywhere in the KG and are not just restricted in the immediate neighborhood of eq. As noted before, in our formulation, a case is a fact augmented with a sample of KG paths that connect the entities of the fact.

Instead if we find an exact match for the sequence of A SIMPLE APPROACH TO CASE-BASED REASONING IN KNOWLEDGE BASES relations, we revise the rules by instantiating the variables with the entities which lie along the path in the neighborhood of e1q. As mentioned before, our CBR based approach needs no training and gathers reasoning patterns from few similar entities. [2009] propose a model that learn features of entities and the non-parametric approach allows learning unbounded number of dimensions.