Multi-Aspect Question Reformulation

• Multi-Aspect Question Reformulation is a set of methods that recast natural language queries along distinct dimensions to improve clarity and answer preservation.
• It systematically manipulates questions via surface, semantic, and complexity adjustments, ensuring that the original answer remains consistent.
• Applications include conversational AI, mathematical reasoning, and non-factoid question answering, with measurable improvements in performance metrics.

Multi-Aspect Question Reformulation (MQR) is a family of methodologies and algorithmic frameworks that recast, rewrite, or augment natural language questions along multiple orthogonal dimensions to facilitate more robust understanding, retrieval, or reasoning. MQR techniques span applications in conversational AI, mathematical reasoning, non-factoid question answering, information retrieval, and question rewriting. They systematically manipulate questions to address linguistic ambiguity, incomplete context, or insufficient complexity, often guided by explicit aspect typologies and with quantifiable impacts on downstream task performance.

1. Formal Definitions and Dimensional Taxonomy

MQR refers to the process of generating question variants—each reformulated along distinct axes or “aspects”—with the aim of preserving intent, answer equivalence, or gold label, while improving properties such as fluency, explicitness, coverage, or intrinsic difficulty. Precisely, if $q$ is an input question and $\mathcal{A}$ the set of considered aspects (e.g., background, terminology, sub-problems, syntactic completeness), then MQR produces $\{q^{(\alpha)}: \alpha \in \mathcal{A}\}$ where each $q^{(\alpha)}$ is a reformulation under aspect $\alpha$.

Common reformulation axes include:

• Surface/Linguistic: Grammar correction, typo repair, explicitness (e.g., question vs. fragment) (Chu et al., 2019).
• Semantic/Pragmatic: Co-reference and ellipsis resolution in conversation (Ye et al., 2021).
• Complexity: Injecting irrelevant background, introducing symbolic definitions, or splitting conditions into sub-problems to systematically increase difficulty while preserving the answer (Dai et al., 28 Jan 2026).
• Aspect Decomposition: Parsing multi-faceted or non-factoid questions into single-aspect sub-queries (e.g., pro/con, compare/contrast, procedural steps) (Lee et al., 20 Mar 2025).
• Facet Expansion: Generating diverse related questions to cover heterogeneous information needs in retrieval (Seo et al., 12 Feb 2025).

A defining property in many MQR settings is the answer-preserving constraint:

$$\text{Solve}(q^{(\alpha)}) = a \ \text{whenever} \ \text{Solve}(q) = a.$$

This ensures that augmentation, editing, or decomposition does not alter gold truth.

2. Instance Methodologies and Architectures

Several distinct MQR instantiations have been proposed, exhibiting both task-specific and generalizable design principles:

• Text-to-Text Sequence Transduction: Chu et al. (2019) construct a large dataset of ill-formed vs. well-formed questions and train LSTM and Transformer encoder-decoder networks to map from poorly-formed to grammatical, explicit, typo-free queries (Chu et al., 2019). Supervised cross-entropy loss is minimized:

$$\mathcal{L}_{\text{CE}}(\theta)=- \sum_{t=1}^T \log P(y_t|y_{<t},x;\theta).$$

• Action-Based Conversational Frameworks: ActNet (2021) models conversational reformulation as a two-stage process: sequence tagging to identify spans requiring replacement or insertion (co-reference/ellipsis), and attention-based retrieval to supply content from prior utterances (Ye et al., 2021). Reformulation is concretely an alternating sequence of “replace” and “insert” operations guided by predicted span labels.
• Multi-Aspect Difficulty Augmentation: In MathForge, each mathematical question is reformulated up to three ways—by augmenting background context, introducing abstract terms, or embedding sub-problems—by prompting an LLM with aspect-specific instructions (Dai et al., 28 Jan 2026). Only reformulated questions that preserve the original answer (validated by a checker) are retained for augmentation.
• Decomposition for Non-Factoid QA: Typed-RAG first predicts a semantic type for the input question (e.g., evidence, comparison, instruction) using a RoBERTa classifier, then decomposes the question into single-aspect sub-queries appropriate for that type. Each sub-query directs retrieval and answer generation, and responses are aggregated for coverage across all aspects (Lee et al., 20 Mar 2025).
• Feedback-Driven Multi-Question Expansion: QA-Expand leverages LLMs to generate multiple aspect questions from a query, produces pseudo-answers as surrogate documents, and filters/rewrites these answers via an LLM-based feedback module. This pipeline increases retrieval diversity and informativeness in IR (Seo et al., 12 Feb 2025).
• RL-Driven Multi-Signal Reformulation: In QRT5, policy-gradient RL tunes a T5 model with rewards both for downstream answer F1 and for fluency (well-formedness), yielding reformulations that balance task fidelity with natural language quality (Chen et al., 2020).

3. Algorithmic Procedures and Formal Properties

The characteristic MQR workflow includes the following general stages, with task-dependent variants:

1. Aspect Tagging or Classification: Either directly, as in ActNet’s span labeling (Ye et al., 2021), or by global question type classifiers (Typed-RAG, (Lee et al., 20 Mar 2025)).
2. Reformulation/Decomposition:
   • Prompt or rule-based LLM generation for augmentation (MathForge, (Dai et al., 28 Jan 2026); QA-Expand, (Seo et al., 12 Feb 2025)).
   • Neural sequence transduction for surface reformulation (Chu et al., (Chu et al., 2019)).
   • Algorithmic decomposition into aspect-isolating sub-queries (Typed-RAG, (Lee et al., 20 Mar 2025)).
3. Answer or Feedback Verification: Answer-preservation checkers (MathForge, (Dai et al., 28 Jan 2026)); downstream reward models (QRT5, (Chen et al., 2020)); relevance filtering (QA-Expand, (Seo et al., 12 Feb 2025)).
4. Aggregation or Synthesis: Multi-response composition, e.g., LLM-mediated answer aggregation in Typed-RAG (Lee et al., 20 Mar 2025).

A representative mathematical formalization of the aspect mapping in MQR (MathForge) is: $$f_{\alpha} : q \longmapsto q^{(\alpha)}, \quad \alpha \in \{\text{Background}, \text{Term}, \text{SubProblem}\},$$ with the constraint that each $q^{(\alpha)}$ admits the same solution as $q$.

For multi-aspect decomposition: $Q \xrightarrow{\text{decomp}} \{q_1, ..., q_k\}$ where each $q_i$ targets a unique reasoning aspect $A_i$ of the original question.

4. Performance Gains and Empirical Benchmarks

Systematic evaluations demonstrate consistent improvements when applying MQR strategies:

• Mathematical Reasoning: MathForge shows a gain from 39.79% to 41.04% in accuracy when switching from original to MQR-augmented data (GRPO setting), and further to 42.17% via DGPO integration. Gains are maximized when all three reformulation aspects are used (+2.27%) (Dai et al., 28 Jan 2026).
• Conversational Reformulation: ActNet improves exact match by +3.9% (49.3%→53.2%) and ROUGE-L by +1.0% (90.0→91.0) on the Restoration-200K benchmark (Ye et al., 2021).
• Non-Factoid QA: Typed-RAG achieves an absolute MRR gain of +0.1766 (0.5893 → 0.7659, ≈30% relative) vs. standard RAG in Wiki-NFQA (Lee et al., 20 Mar 2025).
• Retrieval Expansion: QA-Expand produces significant improvements on BEIR and TREC, e.g., BEIR nDCG@10 increases from 0.5202 (prior SOTA) to 0.5302 (+1.0 point, ≈13% relative gain) (Seo et al., 12 Feb 2025).
• Question Rewriting: Transformer models trained on the MQR dataset improve BLEU-4 from 5.9 (ill-formed input) to 22.1, and over GEC/Paraphrase baselines by 13.2 points (Chu et al., 2019).

A summary table highlights representative gains in different settings:

System	Metric	Baseline	With MQR	Δ	Reference
MathForge	Accuracy (%)	39.79	41.04	+1.25	(Dai et al., 28 Jan 2026)
ActNet	EM (%)	49.3	53.2	+3.9	(Ye et al., 2021)
Typed-RAG	MRR	0.5893	0.7659	+0.1766	(Lee et al., 20 Mar 2025)
QA-Expand	nDCG@10	0.5202	0.5302	+1.0	(Seo et al., 12 Feb 2025)
Transformer (MQR)	BLEU-4	5.9	22.1	+16.2	(Chu et al., 2019)

In each domain, these gains are confirmed via statistically significant tests or controlled ablation studies.

5. Limitations, Pitfalls, and Open Challenges

Empirical and theoretical results across works reveal several limitations:

• Semantic Drift and Equivalence Checking: Aggressive reformulation may introduce semantic drift. In MathForge, 2–3% of augmented questions alter the logical core and must be filtered (Dai et al., 28 Jan 2026). Similar drift is observed in neural rewriting models, where only 60–70% of outputs are judged semantically equivalent to originals by human annotators (Chu et al., 2019).
• Aspect Coverage: Most frameworks operate over a modest fixed set of aspects (e.g., background, term, sub-problem). Other dimensions (e.g., verbosity, symbolization, question intent) remain to be systematically exploited (Dai et al., 28 Jan 2026).
• Prompt Engineering and LLM Dependence: Performance often depends acutely on prompt design and LLM internals (external reformulation, hint extraction), and incurs API latency or unpredictability (Li et al., 24 Mar 2025).
• Reformulation Quality Dependency: If early reformulation or hint extraction produces off-track or vacuous summaries, downstream modules may misfire or oscillate (Li et al., 24 Mar 2025).
• RL Instability: In reinforcement learning-based MQR, fluency may degrade when optimizing solely for task rewards, and overfitting can rapidly set in (Chen et al., 2020).
• Resource Overhead: Multi-aspect approaches frequently require additional computation, both for aspect-specific generation and post hoc filtering or selection (e.g., feedback LLMs in QA-Expand) (Seo et al., 12 Feb 2025).

6. Synthesis and Generalization Across Domains

MQR has proven broadly effective and adaptable, with domain-specific tailoring:

• Situation Puzzles and Interactive Reasoning: External MQR rescues LLMs from stagnant dialog loops by distilling interaction history into multi-aspect hints, resetting context and promoting new lines of inquiry (Li et al., 24 Mar 2025).
• Conversation QA: Unified architectures resolve surface ablation via span tagging and reconstructive insertion/replacement actions, rather than simple paraphrasing (Ye et al., 2021).
• Mathematical RL: MQR-driven augmentation enables difficulty-controlled curriculum learning without the need for new solution generation, integrating seamlessly with advanced RL policy optimization algorithms (Dai et al., 28 Jan 2026).
• Non-Factoid QA and IR: Multi-aspect decomposition (and subsequent aggregation) meaningfully improves comprehensiveness and facet coverage in open-ended and retrieval tasks (Lee et al., 20 Mar 2025, Seo et al., 12 Feb 2025).
• Surface Quality Improvement: Large multi-domain rewriting datasets demonstrate that neural models can simultaneously improve multiple question quality axes (grammar, spelling, explicitness) given explicit aspect-based supervision (Chu et al., 2019).

A plausible implication is that the key ingredient for effective multi-turn or multi-stage reasoning is the distillation of accumulated interaction or evidence into a concise aspect-diversified summary—whether as prompts, hints, or sub-queries—so that downstream models are equipped to advance beyond contextual local minima.

7. Future Directions

Several extensions and ongoing challenges are prominent in contemporary MQR research:

• Aspect Discovery and Automation: Automated selection of the most beneficial aspects per question (e.g., via meta-controllers, difficulty predictors) (Dai et al., 28 Jan 2026).
• Compositional and Nested Reformulations: Composing multiple aspect transformations recursively or in sequence to amplify beneficial effects (Dai et al., 28 Jan 2026).
• Curriculum and Adversarial Ordering: Dynamic organization of reformulated questions by increasing hardness or contrast, potentially with adversarial selection (Dai et al., 28 Jan 2026).
• Integration with External Knowledge: Coupling MQR pipelines with fact-grounded retrieval or knowledge graphs to support answer-preserving reformulation (Li et al., 24 Mar 2025).
• Multilingual and Domain Transfer: Adapting aspect-based or decomposition strategies to less-resourced languages and non-English conversational contexts (Ye et al., 2021, Lee et al., 20 Mar 2025).
• Human-in-the-Loop Evaluation: Reducing self-evaluation bias from LLM scorers by systematic human evaluation on facet coverage and semantic equivalence (Lee et al., 20 Mar 2025).

In summary, Multi-Aspect Question Reformulation provides a principled, extensible set of methodologies for enhancing question quality, diversity, and complexity—exploiting diverse axes of reformulation and decomposition to drive measurable improvements in performance and robustness across a wide range of machine reasoning, retrieval, and dialog tasks.