Skip to content Skip to navigation
Qualifying Exam
12/14/2018 01:30 pm
Hill 482

Towards Selective Composition of Facts for Generating Open-domain Entity Descriptions

Rajarshi Bhowmik, Dept. of Computer Science

Examination Committee: Prof. Gerard de Melo, Prof. Matthew Stone, Prof. Sungjin Ahn, and Prof. Pranjal Awasthi

Abstract

Selective compositionality – the ability to select relevant concepts from the acquired knowledge and composing them into a coherent linguistic form – is one of the cornerstones of human language. In this work, we study a specific form of selective compositionality where the goal is to synthesize a succinct textual description of an entity given the available factual knowledge. Large-scale open-domain knowledge graphs such as Wikidata and Google Knowledge Graph sparsely provide a synoptic textual description that entails the most discerning facts of an entity (e.g. Roger Federer: Swiss tennis player). Such descriptions effectively provide for a near-instantaneous human understanding of an entity and are also useful in a number of linguistic tasks, including named entity disambiguation. Additionally, they can serve as fine-grained ontological types in question answering and reasoning-driven applications. Unfortunately, due to the rapid increase in the number of entities, such descriptions are missing for many entities in the existing knowledge graphs, necessitating an automated synthesis of succinct textual descriptions from underlying factual knowledge. To this end, we propose a neural architecture that can precisely generate synoptic descriptions for a diverse set of entities. Through extensive evaluation, we demonstrate the ability of our architecture to select relevant facts and composing them to coherent and succinct descriptions more accurately by pitting it against a number of competitive baselines.