[Paper Review] Variational Reasoning for Question Answering with Knowledge Graph

Yuyu Zhang, Hanjun Dai+3 more|arXiv (Cornell University)|Sep 12, 2017

Topic ModelingComputer Science181 citations

TL;DR

The paper presents VRN, an end-to-end variational framework that jointly learns topic-entity recognition and multi-hop reasoning over a knowledge graph, enabling robust QA on noisy text and audio inputs.

ABSTRACT

Knowledge graph (KG) is known to be helpful for the task of question answering (QA), since it provides well-structured relational information between entities, and allows one to further infer indirect facts. However, it is challenging to build QA systems which can learn to reason over knowledge graphs based on question-answer pairs alone. First, when people ask questions, their expressions are noisy (for example, typos in texts, or variations in pronunciations), which is non-trivial for the QA system to match those mentioned entities to the knowledge graph. Second, many questions require multi-hop logic reasoning over the knowledge graph to retrieve the answers. To address these challenges, we propose a novel and unified deep learning architecture, and an end-to-end variational learning algorithm which can handle noise in questions, and learn multi-hop reasoning simultaneously. Our method achieves state-of-the-art performance on a recent benchmark dataset in the literature. We also derive a series of new benchmark datasets, including questions for multi-hop reasoning, questions paraphrased by neural translation model, and questions in human voice. Our method yields very promising results on all these challenging datasets.

Motivation & Objective

• Address noisy question inputs and uncertain topic entity localization.
• Enable multi-hop logical reasoning over large-scale knowledge graphs in an end-to-end trainable model.
• Learn topic entity recognition and reasoning embeddings jointly from question-answer pairs without strong annotations.
• Provide a scalable architecture that can handle text and speech inputs for QA over KG.

Proposed method

• Introduce VRN, a two-module probabilistic framework: P(y|q) for topic entity recognition and P(a|y,q) for KG-based reasoning to locate answers.
• Represent topic entities via a softmax over KG entities with a neural question encoder f_ent(q) (text or audio) as in P_theta1(y|q) = softmax(W_y^T f_ent(q)).
• Model reasoning over KG as embeddings of reasoning subgraphs; compatibility P_theta2(a|y,q) computed via softmax of question-type embedding f_qt(q) against graph-embedding g(G_{y→a}).
• Define reasoning graphs G_{y} within T hops of the topic entity, and G_{y→a} as the minimal subgraph containing all paths from y to a; embed G_{y→a} recursively with a forward-graph architecture (Equation 6).
• Train end-to-end with latent y treated via variational inference; objective is the variational lower bound using Q_ψ(y|q,a) and REINFORCE with variance reduction (Equations 7–10).
• Inference uses beam search over top-k y from P_theta1(y|q) to compute max_a P_theta2(a|y,q) (approx. (11)).

Experimental results

Research questions

• RQ1 Can an end-to-end model jointly learn topic entity recognition and multi-hop reasoning over a KG from question-answer pairs without strong annotations?
• RQ2 Does a forward-graph KG embedding approach enable scalable and effective multi-hop reasoning over large knowledge graphs?
• RQ3 To what extent can end-to-end variational training with REINFORCE improve entity recognition and reasoning under noisy text or audio inputs?
• RQ4 How does the proposed VRN perform on single-hop vs multi-hop questions and on text vs audio QA tasks?
• RQ5 Is the MetaQA benchmark effective for evaluating KG-based QA under realistic noise and paraphrasing conditions?

Key findings

• VRN achieves state-of-the-art performance on Vanilla (1-hop), 2-hop, and 3-hop datasets.
• In the Vanilla-EU setting (entity unlabeled), performance drops across methods but VRN remains strongest, highlighting the benefit of joint training.
• VRN substantially outperforms baselines on 2-hop and 3-hop questions, indicating strong multi-hop reasoning capability.
• On NTM-EU and Audio-EU datasets, VRN outperforms baselines, with audio QA showing promise despite higher difficulty.
• Variational training with REINFORCE and variance reduction improves entity recognition and overall QA performance, especially with limited labeled topic entities.
• A forward-graph embedding architecture with shared computation enables efficient scoring of many possible reasoning paths (G_{y→a}).