Abstract
Traditional multi-hop Reading Comprehension (RC) models often rely on extractive explanations—highlighting existing sentences to justify an answer. However, these extracts are frequently verbose and contain irrelevant information. We introduce SuQA (Summarizer-augmented QA), a framework that shifts toward abstractive explanations. By employing a semi-supervised reinforcement learning pipeline, SuQA generates concise summaries that resolve ambiguities and strip away "fluff" while remaining sufficient for a model to answer complex questions. SuQA achieves a 2.9x higher compression rate than extractive baselines and improves human-judged sufficiency.
Introduction
In multi-hop RC, a model must connect disparate pieces of information across multiple paragraphs. To ensure transparency, these models are typically designed to output "supporting facts." The prevailing extractive approach, however, suffers from two main issues: a lack of conciseness and poor comprehensibility. Extractive sentences often include superfluous details (e.g., dates or adjectives) unnecessary for the specific query. Furthermore, they may contain unresolved pronouns (e.g., "he" or "it") that make the explanation difficult to understand in isolation.
SuQA addresses these limitations by generating explanations that are minimal (no extra info), sufficient (containing all necessary facts), and comprehensible (self-contained and readable).
Methodology
The SuQA framework utilizes a "faithful-by-construction" pipeline consisting of two primary components:
- Abstractive Explainer (AX): A DistilBART-based model that takes the question and relevant paragraphs as input to generate a concise abstractive summary.
- QA Module (QAM): A UnifiedQA-based model that attempts to answer the question using only the generated explanation. This ensures the explanation is the sole source of reasoning for the final answer.
Reinforcement Learning with Multi-faceted Rewards
To overcome the scarcity of human-annotated abstractive explanations, SuQA employs a semi-supervised reinforcement learning (RL) approach. The AX is first pre-trained on standard summarization datasets and fine-tuned on a small subset of the R4C dataset. Subsequently, it is optimized using Self-Critical Sequence Training with a reward function that balances three objectives:
- Summarization: Incentivizes high compression and relevance.
- Sufficiency: Rewards the model when the QAM correctly answers the question based on the AX's output.
- Comprehensibility: Uses a CoLA (Corpus of Linguistic Acceptability) scorer to ensure grammatical soundness.
Experimental Setup and Results
The framework was evaluated on the HotpotQA benchmark, using the R4C dataset for limited supervision (only 2,000 instances, ~3% of the total).
Key Findings
- Superior Compression: SuQA achieved a compression rate 2.9x higher than extractive baselines, significantly reducing the cognitive load for human readers.
- Enhanced Sufficiency: Human evaluations indicated that SuQA’s explanations were judged as "sufficient" for answering the question more frequently than extractive ones (+2.5 points).
- XF1 Metric: The authors introduced XF1, a metric combining sufficiency and F1 score. SuQA outperformed the extractive baseline (63.7 vs 60.5), demonstrating its ability to provide useful, answer-conducive explanations.
- Data Efficiency: The RL approach proved highly effective in low-resource settings. Incorporating just 298 human-annotated examples improved human-judged sufficiency by 11.0 points compared to an unsupervised setup.
Discussion and Limitations
A notable observation was the "shortcut" problem: extractive models often achieve high F1 scores by exploiting lexical cues in the original text rather than following a logical reasoning path. By forcing an abstractive distillation, SuQA filters out these shortcuts. While this leads to a slight decrease in raw F1 scores, it results in more "honest" and human-readable explanations.
Future work may involve joint optimization of the Explainer and QA Module to improve alignment. Additionally, while SuQA explains what facts are used, expanding the framework to describe the inference process (the "how") of the multi-hop jump remains an open challenge.
Conclusion
SuQA demonstrates that abstractive summarization is a viable and superior alternative to extractive methods for explainable multi-hop RC. By prioritizing conciseness and comprehensibility through reinforcement learning, it provides a path toward more transparent and user-friendly AI reasoning systems.
Summary updated