Graph-to-Sequence Model for Question Generation via Reinforcement Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current natural question generation systems fail to consider global interactions between answer and context, neglect rich hidden structural information in text, and rely solely on cross-entropy based objectives, leading to issues like exposure bias and inconsistency between training and testing.
Innovation Solution
A reinforcement learning based graph-to-sequence model is introduced, utilizing a deep alignment network for word-level and hidden representation-level alignments, combined with a bidirectional gated graph neural network to construct a passage graph and decode questions, and a hybrid evaluator that combines cross-entropy and reinforcement learning evaluations to ensure syntactic and semantic coherence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If cross-entropy based objectives are used for training, then the model can be trained efficiently with straightforward optimization, but the model suffers from exposure bias and inconsistency between training and testing
Solution Approach 1:
The patent introduces reinforcement learning with a hybrid evaluator that provides feedback signals during training. The evaluator combines cross-entropy evaluation with reinforcement learning evaluation, allowing the model to receive corrective feedback that reduces the gap between training and testing behaviors, thereby mitigating exposure bias while maintaining training efficiency.
2Measurement precision
If global interactions between answer and context are considered, then the model generates more accurate and relevant questions, but the computational complexity increases
Solution Approach 1:
The patent segments the question generation process into distinct components: a deep alignment network for contextual understanding, a graph encoder for structural representation, and a decoder for question generation. This segmentation allows global interactions to be modeled through coordinated components rather than a monolithic complex system, managing computational complexity while maintaining accuracy.
Solution Approach 2:
The patent transforms the input passage into a graph structure that captures global interactions in a different dimensional representation. By encoding the passage as a graph with nodes and edges representing semantic relationships, the model can efficiently capture global context without the quadratic complexity of pairwise interactions in the original sequence space.
3Loss of information
If hidden structural information of the word sequence is exploited, then the model captures richer semantic relationships, but the model complexity and processing requirements increase
Solution Approach 1:
The patent transforms the linear word sequence into a graph structure that preserves and emphasizes hidden structural information. The graph encoding approach represents semantic relationships in a different dimensional space where structural patterns become more explicit and computationally accessible, allowing the model to capture rich relationships without proportionally increasing complexity.
Data Source
AI summary
For a passage text and a corresponding answer text, perform a word-level soft alignment to obtain contextualized passage embeddings and contextualized answer embeddings, and a hidden level soft alignment on the contextualized passage embeddings and the contextualized answer embeddings to obtain a passage embedding matrix. Construct a passage graph of the passage text based on the passage embedding matrix, and apply a bidirectional gated graph neural network to the passage graph until a final state embedding is determined, during which intermediate node embeddings are fused from both incoming and outgoing edges. Obtain a graph-level embedding from the final state embedding, and decode the final state embedding to generate an output sequence word-by-word. Train a machine learning model to generate at least one question corresponding to the passage text and the answer text, by evaluating the output sequence with a hybrid evaluator combining cross-entropy evaluation and reinforcement learning evaluation.