Variational Reasoning for Question Answering with Knowledge Graph (1709.04071v5)

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.

Variational Reasoning for Question Answering with Knowledge Graph

The paper proposes a novel deep learning framework, Variational Reasoning Network (VRN), designed for question answering (QA) using knowledge graphs (KGs). The framework addresses significant challenges in QA tasks, specifically the need for multi-hop reasoning and the handling of noisy question inputs. Traditional approaches using semantic parsers are inadequate as they often rely on predefined grammars and lack the ability to handle noise and support end-to-end training. In contrast, VRN aims to perform end-to-end learning and reasoning over a KG, leveraging the structure and relationships inherent in the graph.

VRN consists of two important probabilistic modules: one for recognizing the topic entity within the question and another for reasoning over the KG to derive the answer. The first module ($P_{\theta_1}(y|q)$) determines the probability of a given entity being a part of the question, while the second ($P_{\theta_2}(a|y,q)$) models the logic reasoning over the KG.

A crucial aspect of VRN is its embedding-based approach for reasoning, which focuses on the concept of a "reasoning graph." The reasoning graph consists of subgraphs within a specified hop of entities from the topic entity, capturing potential reasoning paths to derive an answer. The authors introduce innovative architectures that embed reasoning-graphs using recursive techniques inspired by forward filtering in Bayesian networks.

The training of VRN involves a variational inference method that parameterizes a posterior distribution over the possible topic entities given the question and answer. This approach utilizes REINFORCE with variance reduction to manage the non-differentiability inherent in sampling discrete variables. Such an architecture enables the handling of noisy data input, including paraphrased text and audio questions, by training a system that improves upon entity recognition jointly while reasoning over KGs.

The authors introduce MetaQA, a comprehensive benchmark designed to evaluate QA systems more effectively across scenarios requiring different reasoning hops. This includes datasets with paraphrased questions generated by dual learning models and audio questions derived from text-to-speech systems.

Experimental results reveal VRN's state-of-the-art performance across multiple test scenarios for both single-hop and multi-hop questions, notably outperforming existing methods such as Key-Value Memory Networks (KV-MemNN) and a QA system proposed by Bordes et al., especially in the more challenging environments with unlabeled topic entities and audio question formats.

The implications of this research are substantial for advancing QA systems capable of operating in realistic settings. VRN's ability to learn entity recognition and complex reasoning simultaneously suggests potential applicability in expanding AI-driven virtual assistant capabilities, moving towards natural and resilient interaction models. Looking forward, further exploration may focus on improving audio QA performance further and extending the framework to environments with even larger-scale KGs and diverse query types. This research marks a step towards more versatile AI systems that can handle intrinsic noise and leverage the full relational complexity within knowledge graphs.