Source-linked AI summary
Variational Reasoning for Question Answering with Knowledge Graph
Yuyu Zhang, Hanjun Dai, Zornitsa Kozareva, Alexander J. Smola, Le Song
TL;DR
KG question answering must identify noisy entity mentions and perform multi-hop reasoning from question-answer pairs with limited fine-grained supervision. The paper introduces VRN, a unified end-to-end variational framework, and reports state-of-the-art results on single- and multi-hop datasets alongside promising results on noisy audio QA.
Problem
KG-QA systems must handle noisy entity expressions and unknown multi-hop reasoning paths when trained mainly from question-answer pairs without fine-grained annotations.
Method
VRN jointly models probabilistic topic-entity recognition and logic reasoning over the knowledge graph in an end-to-end variational framework.
Results
The method achieves state-of-the-art performance on single- and multi-hop datasets and promising results on challenging audio QA datasets.
Takeaways & Limitations
The results support VRN as a step toward QA systems that handle noisy text and speech questions while learning multi-hop reasoning from examples.
Takeaways & Limitations
The approach assumes the maximum reasoning hop count T is known and uses labeled topic entities for some experimental settings.
Abstract
from arXiv · showhide
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.
1 Introduction
KG-based QA must handle noisy entity mentions and infer answers through multi-hop reasoning from question-answer pairs without fine-grained annotations. The paper proposes VRN, an end-to-end probabilistic framework that jointly addresses these challenges and evaluates it on new benchmarks.
- Challenges: KG-powered QA is difficult because questions contain noisy entity expressions and often require multi-hop reasoning.Noise includes typos, ambiguity, and pronunciation variation, while multi-hop relations grow explosively with hop count.
- Challenges: Question-answer training pairs usually lack annotations for the mentioned entity, question type, and exact reasoning path.This makes simultaneous entity locating and reasoning difficult to learn from question-answer pairs alone.
- Approach: VRN provides a probabilistic end-to-end framework that simultaneously handles uncertain topic entities and multi-hop reasoning.The framework is introduced specifically to address both key challenges in a unified model.
- Approach: The architecture uses a propagation-like deep learning model for logic inference and REINFORCE with variance reduction for training.The variance-reduction technique addresses the non-differentiable sampling involved in the probabilistic model.
- Benchmarks: MetaQA contains over 400K single- and multi-hop questions, plus neural-translation paraphrases and text-to-speech audio datasets.These datasets target more realistic and difficult question inputs.
- Results: Experiments report state-of-the-art performance on single- and multi-hop datasets and promising results on challenging audio QA datasets.The authors connect these results with multi-hop reasoning capability and end-to-end learning for noisy text and speech input.
2 Related Work
Prior KG-QA approaches commonly separate question parsing or entity locating from answer reasoning, limiting end-to-end learning and robustness to noisy inputs. VRN instead presents an end-to-end architecture with probabilistic topic-entity recognition and KG reasoning.
- Semantic parsing: Semantic-parser approaches map questions to logical forms or executable programs before querying the knowledge graph.They require domain-specific grammars, rules, or fine-grained annotations.
- Semantic parsing: Traditional semantic-parser systems are not designed for noisy questions and do not support end-to-end training across parsing and reasoning.Their question parsing and downstream processing use separate stages.
- Entity locating: Existing entity-locating methods include keyword matching, frequency-based methods, and domain-specific approaches that are not jointly trained with reasoning.Errors in locating entities can cascade into the downstream QA system, particularly for speech recognition outputs.
- VRN: VRN’s architecture combines probabilistic topic-entity recognition with probabilistic logic reasoning over the knowledge graph.Figure 1 depicts the two modules and reasoning graphs used to score candidate answers.
3 Model
VRN models topic-entity recognition and multi-hop KG reasoning in one probabilistic, end-to-end framework. Its reasoning-graph embeddings recursively represent possible paths while sharing intermediate computations.
- 3.1 Problem definition: A knowledge graph represents entities as nodes and relations as directed edges, and QA outputs an entity that answers the question.
- 3.1 Problem definition: Question-answer pairs lack annotations for the mentioned entity, question type, and reasoning steps, requiring direct learning from noisy questions and answers.
- 3.2 Overall formulation: VRN jointly models uncertain topic-entity recognition and multi-hop reasoning through two probabilistic components, Pθ1(y|q) and Pθ2(a|y,q).
- 3.3 Probabilistic module for topic entity recognition: The topic-entity module estimates which KG entity y is mentioned in question q using question representations suited to text or audio inputs.
- 3.4 Probabilistic module for logic reasoning over knowledge graph: The reasoning module embeds the subgraph containing all paths from topic entity y to candidate answer a, with a known maximum hop count T.
- 3.4 Probabilistic module for logic reasoning over knowledge graph: Forward graph embedding recursively combines parent embeddings and relation encodings, computing all scope embeddings in O(|V(Gy)| + |E(Gy)|) time.
4 End-to-end Learning
VRN learns its latent topic entities with variational inference because exact posterior computation is intractable. REINFORCE, normalization, and a learned baseline provide a practical gradient estimator.
- 4.1 Variational method with inverse reasoning-graph embedding: Exact EM updates are intractable because the posterior lacks a closed-form expression, so VRN optimizes a variational free-energy objective.
- 4.1 Variational method with inverse reasoning-graph embedding: The variational posterior Qψ(y|q,a) uses the answer together with question information to improve latent topic-entity inference through logic matching.
- 4.2 REINFORCE with variance reduction: REINFORCE estimates gradients for the discrete latent variable y, with variance reduction applied to the policy-gradient learning signal.
- 4.2 REINFORCE with variance reduction: The learning signal is centered and normalized, while a baseline network is subtracted to reduce gradient variance before Monte Carlo approximation.
- 4.2 REINFORCE with variance reduction: Parameters in Pθ1(y|q) and Pθ2(a|y,q) receive ordinary gradients, while the posterior parameters ψ use the variance-reduced estimator.
5 Inference
Inference approximates joint topic-entity and answer selection with beam search rather than exhaustive scoring. In experiments, greedy inference with k = 1 already performed well.
- Exact inference over all entity pairs is quadratic in the number of entities, so the model uses beam search as an approximation.
- k = 1, equivalent to greedy inference, already achieves good performance in the experiments.
6 Experiments
Experiments evaluate MetaQA across single- and multi-hop reasoning, paraphrased text, and audio questions, comparing VRN with several baselines. VRN performs best on reasoning tasks and improves entity recognition through variational joint training, including a 10% gain for audio 1-hop questions.
- The MetaQA benchmark: MetaQA extends WikiMovies with over 400K single- and multi-hop questions plus paraphrased and audio variants.The benchmark includes Vanilla, neural-translation-model paraphrases, and text-to-speech audio datasets.
- Results and discussions: VRN outperforms the competitor methods, especially on the harder 2-hop and 3-hop Vanilla questions.All methods lose accuracy as hop count increases, while VRN maintains a large margin over competitors.
- Results and discussions: In the entity-unlabeled setting, reasoning-based methods degrade on multi-hop questions, whereas supervised embedding benefits from memorizing question-answer entity pairs.The reported training statistics explain the differing behavior of supervised embedding.
- Model ablation: VRN's joint training improves the entity recognizer without corresponding topic-entity labels, with a 10% improvement on audio 1-hop questions.Improvements are substantial for 1-hop and 2-hop questions but marginal for 3-hop and audio multi-hop questions.
- Inspection of learning and inference: Variance reduction in REINFORCE significantly helps convergence, while simpler reasoning tasks converge better.The paper also inspects a highest-scoring inference path that reaches German through EuroTrip and the in language relation.
A Details of the MetaQA benchmark
MetaQA constructs benchmark subsets that vary reasoning depth and question formulation. Vanilla supplies templated multi-hop questions, while NTM introduces paraphrased wording through round-trip translation.
- Vanilla datasets: Vanilla 3-hop questions use meaningful relation chains and exclude awkward actor-or-director-to-movie constructions.The design includes chains such as Movie to Actor, Writer, or Director to Movie and then to another attribute or role.
- NTM datasets: NTM paraphrases Vanilla questions by translating English to French and back to English with beam search.The resulting questions vary in wording while preserving meaning and include 1-hop, 2-hop, and 3-hop categories.
B Question samples
This section presents MetaQA question categories, dataset statistics, and examples of multi-hop inference and question variation. It also identifies the benchmark’s audio-generation interface.
- Question categories: MetaQA includes 21 types of 2-hop questions and 15 types of 3-hop questions.The exact question types and examples are listed in Tables 5 and 6.
- Question variation: The NTM example changes the wording of a question about films directed by David Mandel while preserving its meaning.
- Inference example: Figure 5 illustrates an example of learned 2-hop inference.
- Dataset statistics: The benchmark’s dataset statistics and ratios of new entities and entity pairs are summarized in Tables 3 and 4.
C Visualization of audio data
The audio-data visualization shows that questions sharing an entity have similar but non-identical MFCC entity regions, motivating separate prediction objectives for entity and question type.
- MFCC visualization: MFCC visualizations show similar but non-identical entity regions for two questions sharing the same entity.The entity portions are highlighted with red dotted lines.
- MFCC visualization: The differing waveforms demonstrate the difficulty of handling audio questions.
- Audio model: Two CNNs use the same MFCC input to predict the topic entity and the question type.Their second-last-layer activations serve as audio-question embeddings.
E More experiment results
The convergence experiment reports that variance reduction substantially accelerates training, while simpler inference tasks converge to better solutions.
- Convergence: Variance reduction makes training converge very fast.The comparison is based on training-loss curves shown in Figure 4.
- Convergence: Tasks involving fewer inference steps converge to a better solution.
E.2 Visualization of the learned reasoning rule
The learned reasoning rule is visualized through inference paths in a reasoning graph, including a 2-hop example that connects a movie to its language through a learned relation.
- Visualization of the learned reasoning rule: For 1-hop answers, the model checks compatibility between the edge type and the question embedding.
- Visualization of the learned reasoning rule: For multi-hop answers, the model traverses from the answer toward the topic entity and selects the most compatible edge at each step.
- Visualization of the learned reasoning rule: The 2-hop example reaches German by finding EuroTrip through directed or wrote, then following the in language relation.