Knowledge Graph Based Question Answering System: Architecture, Research Results, and Deep Learning Approaches

The talk, originally presented by Dr. Cui Wanyun of Fudan University at Ctrip Tech Center's Deep Learning Meetup, introduces a knowledge‑graph‑based question answering (QA) system developed at Fudan.

QA systems answer natural‑language queries and have achieved success in various domains such as IBM Watson, Apple Siri, Google Now, Amazon Alexa, and medical applications.

Based on answer corpora, QA can use pure text sources or structured knowledge graphs (RDF triples). Structured graphs enable more precise answers, and large‑scale graphs like WolframAlpha, Google Knowledge Graph, and Freebase make open‑domain QA feasible.

The proposed system follows a three‑layer architecture: entity layer (semantic community search, word sense disambiguation, co‑occurrence network), language layer (semantic templates for verbs and nouns), and application layer (integrated QA with question templates and deep‑learning modules).

Key components include semantic community search that groups words into communities to compute similarity, semantic disambiguation, and verb‑semantic templates that map verbs to concepts or objects, guided by minimum description length theory.

Research results show training of over 27 million question templates covering 2,782 intent groups, achieving a 59 % success rate in a knowledge‑graph QA competition and high accuracy on the QALD benchmark.

Deep‑learning techniques such as a simple CNN for entity extraction and a multi‑CNN model that predicts answer path, context, and type further improve performance.

The authors discuss current challenges: limited high‑quality training data, incomplete knowledge‑graph content, and the need to combine KB‑based and IR‑based QA to broaden coverage while maintaining answer precision.