Question-Based Paraphrasing (QBP)

• Question-Based Paraphrasing (QBP) is a semantic reformulation framework that transforms questions to preserve their intended answers and underlying event structures.
• It is applied across tasks like VideoQA narrative synthesis, conversational QA rewriting, and duplicate question retrieval for improved model robustness.
• Empirical studies show that QBP enriches supervision by consolidating QA pairs into denser narratives, leading to faster convergence and higher accuracy.

Searching arXiv for the cited papers to ground the article and verify metadata. Question-Based Paraphrasing (QBP) denotes a family of paraphrastic operations in which questions, question–answer pairs, or question groups serve as the organizing signal for semantic reformulation. In the recent literature represented here, the label spans two closely related uses: rewriting a question into another question with equivalent meaning but different surface form, often under an answer-equivalence criterion; and, in VideoQA, paraphrasing an entire group of question–answer pairs for one video into a single coherent narrative paragraph that reconstructs event structure (Liang et al., 29 Sep 2025). Across these uses, QBP is motivated by the same basic problem: systems trained on isolated lexical forms or fragmented supervision remain brittle under reformulation, whereas paraphrastic supervision can expose causal, temporal, social, and discourse structure that is otherwise latent (Brabant et al., 2022).

1. Conceptual scope and formal definitions

A concise way to situate the main formulations is to distinguish between narrative synthesis, question rewriting, and paraphrase recognition.

Formulation	Input → output	Representative sources
Narrative QBP	grouped QA pairs for one video → one coherent narrative paragraph	(Liang et al., 29 Sep 2025)
Question rewriting / question paraphrasing	in-context or out-of-context question → equivalent question with different surface form	(Brabant et al., 2022, Hosking et al., 2021)
Retrieval / identification	question pair or query question → duplicate label or retrieved equivalent questions	(Chandra et al., 2020, Bonadiman et al., 2019)

In the VideoQA setting, QBP is defined as a data-synthesis strategy that uses a LLM to paraphrase an entire group of QA pairs for a video into a single narrative paragraph. Formally, given a question group $\mathcal{G}_i = \{(Q_{i,k}, A_{i,k})\}_{k=1}^{K_i}$, a text-only LLM $\Phi_{\text{QBP}}$ generates one narrative paragraph $$\tilde{d}_i^{\text{narrative}} = \Phi_{\text{QBP}}(\mathcal{G}_i)$$, yielding a synthesized dataset $$\mathcal{D}^{\text{QBP}} = \{(v_i,\tilde{d}_i^{\text{narrative}})\}_{i=1}^N$$ (Liang et al., 29 Sep 2025). The central claim is that standard VideoQA supervision treats each $(v,q,a)$ triple independently and therefore reduces supervision to a “bag of facts,” even when all questions for a video describe one coherent event.

In the broader NLP literature, paraphrase generation is defined as “generating a phrase which is semantically as close as possible to the original while being rewritten in new words and phrases,” and question paraphrasing is treated as a direct special case of that conditional language-modeling setup (Witteveen et al., 2019). CoQAR defines question paraphrasing as transforming a source question into a question with equivalent meaning but different surface form, and in that paper both source and target are out-of-context questions (Brabant et al., 2022). “Factorising Meaning and Form for Intent-Preserving Paraphrasing” defines the criterion even more explicitly for questions: paraphrases should preserve the original intent, and the equivalence condition is that different phrasings should lead to the same answer (Hosking et al., 2021). In retrieval-oriented work, paraphrases are operationalized as “equivalent questions that result in the same answer as the original question,” which makes answer-equivalence the core retrieval criterion (Bonadiman et al., 2019).

This suggests that QBP is best understood not as a single algorithm, but as a paraphrase-centered supervision paradigm whose invariant is answer-preserving or event-preserving reformulation.

2. Narrative QBP in VideoQA

The most specialized use of the term appears in “Beyond Isolated Facts: Synthesizing Narrative and Grounded Supervision for VideoQA,” where QBP addresses the limitations of VideoQA benchmarks such as NExT-QA and STAR. In that setting, each video $v_i$ is a sequence of $T$ frames, and each video is paired with a group of human-annotated QA pairs $\mathcal{G}_i$. Standard training treats those pairs independently, which encourages memorization of local correlations and does not explicitly represent causal, temporal, or social relations (Liang et al., 29 Sep 2025).

QBP changes the supervision unit. Instead of many isolated QA targets, the model receives one dense narrative target per video. The synthesis prompt instructs the LLM to transform the QA list into a single logically coherent paragraph in present tense, to use only facts from the QA pairs, to preserve implied chronological order without inventing temporal or causal links, and to group related facts by theme. The implementation uses text-only LLMs such as DeepSeek or GPT-4o. The downstream multimodal model is then fine-tuned with a unified next-token prediction objective, with loss computed on the narrative tokens only. The paper uses existing multimodal LLM backbones—Qwen2.5-VL (3B and 7B) and MiMo-VL-SFT (7B)—with 16 uniformly sampled frames, AdamW, learning rate $10^{-6}$, cosine schedule, batch size 8, and 1–2 epochs (Liang et al., 29 Sep 2025).

QBP is paired with Question-Based Captioning (QBC), which produces per-question visual rationales. The two are described as orthogonal yet synergistic: QBP supplies holistic, narrative-level supervision, while QBC supplies instance-level grounding. The paper’s human evaluation reports high-quality synthesized narratives, with factual consistency $4.21 \pm 0.55$, logical coherence $\Phi_{\text{QBP}}$0, and fluency $\Phi_{\text{QBP}}$1. Reported failure modes are occasional wrong temporal ordering when source QA pairs have ambiguous cues, and entity confusion due to inconsistent pronouns; severe hallucinations are described as “extremely rare” (Liang et al., 29 Sep 2025).

The empirical effect is not limited to in-domain accuracy. With QBP+QBC, a Qwen2.5-VL-3B model improves on STAR from 67.5% under raw QA supervision to 72.5%, and a Qwen2.5-VL-7B model reaches 80.8% on NExT-QA. More specifically attributable to QBP, narrative supervision improves both in-domain and cross-dataset performance and accelerates convergence by more than 2.5×: on NExT-QA, raw QA training needs about 600 steps to plateau, whereas QBP reaches similar performance by about 220 steps (Liang et al., 29 Sep 2025). The paper’s interpretation is that each QBP paragraph is semantically denser than an individual QA pair and therefore provides richer information per optimization step.

3. Question rewriting and intent-preserving question generation

Conversational QA gives a more classical formulation of QBP. CoQAR defines Question Rewriting (QR) as rewriting each original in-context question into an out-of-context question that is understandable by itself and answerable without conversation history. It also defines Question Paraphrasing (QP) as transforming a source question into a question with equivalent meaning but different surface form, with both source and target out-of-context (Brabant et al., 2022). The corpus contains 4.5K conversations from CoQA, totaling 53K follow-up question–answer pairs; each original question has at least 2 and at most 3 out-of-context rewritings. Original questions average 5.5 words, while out-of-context rewrites average 8.8 words.

The annotation scheme in CoQAR makes explicit what conversational QBP must do: resolve coreference, resolve ellipsis, expand contextually underspecified phrases, and still paraphrase lexically and syntactically. Examples in the paper include “What were they doing?” rewritten as “What were the girl and her dog up to?” or “What was the activity of the girl and the dog for the day?”, and a bare “where?” rewritten as “Where did the girl and her dog go on a trip?” (Brabant et al., 2022). On the modeling side, CoQAR uses BART-base for QP and T5 v1.1-small for QR. On CoQAR test data, BART trained on CoQAR paraphrase pairs achieves BLEU 0.705 and METEOR 0.537 for question paraphrasing, while T5 trained on CoQAR+CANARD reaches BLEU 0.39 and METEOR 0.59 for rewriting. Human evaluation shows that generated paraphrases can match or exceed reference rewrites on meaning preservation and linguistic correctness in some settings, but QR models still trail human rewrites in meaning preservation.

A complementary line of work addresses intent-preserving paraphrasing by architectural factorization. SEPARATOR represents semantics with a continuous VAE latent $\Phi_{\text{QBP}}$2 and surface form with a discrete VQ-VAE latent $\Phi_{\text{QBP}}$3, training the decoder to reconstruct a target question from a semantic input with the same meaning and a syntactic exemplar with the same template but different meaning (Hosking et al., 2021). This enforces a separation between “what is being asked” and “how it is asked.” On Paralex and clustered QQP, SEPARATOR reports the highest iBLEU among compared systems without oracle exemplars, with 14.84 on Paralex and 5.84 on QQP; with oracle exemplars, those rise to 29.99 and 12.34. The paper’s human evaluation finds that VAE baselines preserve meaning better but copy more, whereas SEPARATOR yields the best tradeoff between semantic preservation and dissimilarity to the original question.

At a more general level, “Paraphrasing with LLMs” shows that a simple conditional language-modeling format—original text, a separator token, and paraphrase—already supports sentence- and paragraph-level paraphrasing with GPT-2, and that semantic similarity and lexical overlap can be managed post hoc with USE similarity and ROUGE-L thresholds (Witteveen et al., 2019). Although that paper does not specialize to questions, its formulation directly transfers to question paraphrasing and helped establish the pattern-conditioning approach later used by question-centered systems.

4. Retrieval, recognition, and QA-centered robustness

A large part of QBP research is not generative but discriminative. On Quora Question Pairs, paraphrase identification is treated as a binary classification problem: given two questions, predict whether they are duplicates. In that setting, simple unigram Count Vectorizer and TF–IDF features with XGBoost or CatBoost reach about 68–75% accuracy, LSTM with WordPiece reaches 80.32%, and BERT-Base-Cased reaches 97.07–97.08% accuracy (Chandra et al., 2020). That result is often read as evidence that contextual semantics, not lexical overlap alone, is decisive for recognizing question paraphrases.

At retrieval scale, “Large Scale Question Paraphrase Retrieval with Smoothed Deep Metric Learning” defines Question Paraphrase Retrieval as retrieving equivalent questions that result in the same answer as the original question. The system uses a CNN sentence encoder and a FAISS approximate kNN index; in the open-domain QA setting the index contains about 10 million questions and average query time is under 10 ms with 10 probes (Bonadiman et al., 2019). Its main contribution is Smoothed Deep Metric Loss (SDML), which replaces hard triplet-style separation with a smoothed distribution over in-batch candidates. On Quora test data, SDML with squared Euclidean distance improves retrieval over triplet-loss baselines, reaching 0.6043 / 0.8179 / 0.6789 on the reported retrieval metrics; on the open-domain QA test set it reaches 0.6718 / 0.8830 / 0.7480. The paper argues that standard triplet loss, especially with hard negatives, is brittle in the presence of noisy labels.

Another direct integration of QBP into QA appears in “Learning to Paraphrase for Question Answering,” which models answer probability by marginalizing over the original question and its paraphrases,

$\Phi_{\text{QBP}}$4

Here $\Phi_{\text{QBP}}$5 is a learned paraphrase distribution, while $\Phi_{\text{QBP}}$6 is task-specific QA likelihood. On GraphQuestions, this Para4QA framework raises average F1 from 15.9 for the SimpleGraph baseline to 20.4, and on WebQuestions from 48.5 to 50.7; on WikiQA it improves MAP/MRR from 0.6456/0.6608 to 0.6759/0.6918 (Dong et al., 2017). The paper’s central claim is that paraphrases should be judged operationally by whether they improve answer prediction, not only by semantic similarity.

More recent robustness work reframes QBP as invariance under paraphrase. RoParQ constructs a benchmark for closed-book multiple-choice QA using original questions plus paraphrases from Gemini 2.5 Flash Lite and Claude 3.5 Sonnet, and retains only examples that elicit “inconsistent confidence” from a judge model (Choi, 26 Nov 2025). It introduces XParaCon,

$\Phi_{\text{QBP}}$7

which measures cross-paraphrase consistency. The accompanying paraphrase-aware SFT improves both robustness and, often, accuracy: for example, Llama-3.1-8B on general knowledge moves from 0.781 accuracy and 2.186 XParaCon to 0.798 and 2.629, while Qwen3-4B on math reasoning moves from 0.942 and 4.489 to 0.951 and 4.856.

5. Structured, domain-specific, and passage-centered variants

Not all QBP work paraphrases questions directly. PIE-QG argues that directly paraphrasing question–answer pairs can cause semantic drift, and therefore paraphrases passages instead, uses OpenIE triples from the paraphrased text, and generates synthetic questions from those triples (Nagumothu et al., 2023). The key device is to maximize lexical difference between original sentence and paraphrase using Jensen–Shannon Divergence, then ask a BERT QA model questions derived from paraphrased facts but answer them on the original passage. On SQuAD v1.1 dev, adding paraphrasing to OpenIE-based question generation improves EM from 22.8 to 37.7 and F1 from 36.5 to 53.6, and the full PIE-QG pipeline reaches EM 48.6 and F1 58.7 with BERT-base. This is not direct question paraphrasing, but it shows that paraphrastic diversification can be fact-anchored rather than question-anchored.

“Dense Paraphrasing for Textual Enrichment” pushes in a different direction by rewriting passage text so that hidden arguments, tools, habitats, and event consequences become explicit (Tu et al., 2022). In that work, human-readable dense paraphrases enrich recipe instructions such as “Chop onions, saute until browned” into explicit chains involving cutting boards, knives, pans, spatulas, and result states. The QA benefit is substantial: on DP-generated recipe questions, Base+DP improves over Base from 33.04 EM / 55.27 F1 to 44.99 EM / 64.24 F1 overall, with particularly large gains on Implicit questions, where EM rises from 47.51 to 70.88. Although the paraphrasing target is a passage rather than a question, the method is explicitly justified as support for question answering over implicit content.

Two strongly structured task formulations also show how domain constraints reshape QBP. In Korean dialog, “Extracting Arguments from Korean Question and Command” defines “structured paraphrasing” as mapping a conversational question or command into a canonical argument phrase, such as turning a yes/no question into a “whether” nominal or a wh-question into a phrase headed by 사람, 의미, 위치, 시간, 이유, or 방법 (Cho et al., 2018). The resulting corpus contains 30,837 utterances: 17,869 questions and 12,968 commands. In maritime NL2SQL, paraphrasing is used to expand synthetic question–SQL training data from a synchronous grammar over the maritime schema. There, canonical utterances are paraphrased with backtranslation, GPT-3, BART, and Quillbot, then filtered by a parser-in-the-loop criterion that retains only paraphrases mapping back to the original SQL (Shiri et al., 2022). On the real-world maritime test set, RoBERTa-base trained on original plus Quillbot paraphrases reaches exact match 50.65 and component-match F1 80.58, compared with 43.07 and 77.26 on original data alone.

These domain-specific formulations show that QBP can target very different outputs—narratives, self-contained questions, canonical arguments, retrieved duplicates, SQL-friendly reformulations, or enriched passages—while still preserving the same central invariant: the reformulated text must remain aligned with the original question’s intended answer or event semantics.

6. Limitations, misconceptions, and open problems

A recurring misconception is that paraphrase quality can be reduced to generic semantic similarity. The algebraic word-problem study “’John ate 5 apples’ != ‘John ate some apples’” argues that this is false in mathematically sensitive domains, because paraphrases must preserve numbers, units, relationships, and solvability, not only topical similarity (Gupta et al., 2022). ParaQD therefore trains a self-supervised paraphrase quality detector using positive augmentations such as Num2Words and UnitExpansion and negative augmentations such as numerical deletion, unit replacement, and last-sentence deletion. On AquaRAT, its weighted F1 reaches 0.687 with separation $\Phi_{\text{QBP}}$8, outperforming self-supervised baselines; on human-written PAWP paraphrases it reaches weighted F1 0.685 with $\Phi_{\text{QBP}}$9.

Another misconception is that more paraphrastic diversity is always better. The maritime NL2SQL study shows the opposite tradeoff: Chinese backtranslation has the lowest BLEU-4 and therefore the most lexical diversity, but GPT-3 and Quillbot retain far more paraphrases after SQL-consistency filtering and yield stronger downstream parsing gains (Shiri et al., 2022). Likewise, the FRANK QA study finds that semantically adequate paraphrases are not sufficient when the downstream parser is rigid: on a challenge set of 20 human-created unparsable questions, English–French backtranslation produces adequate paraphrases, but only 1 yields the original alist under FRANK’s parser (Ferguson et al., 2022). That paper concludes that cleaning LC-QuAD 2.0 is necessary because dataset errors distort evaluation, and that paraphrase generation is not a reliable way to increase FRANK’s question coverage under current parser limitations.

For narrative VideoQA QBP, the limitations are different. The method depends on powerful LLMs such as GPT-4o or DeepSeek; the paper notes occasional wrong temporal ordering, entity confusion, and the absence of any formal guarantee that all facts are covered or that logical structure is exact (Liang et al., 29 Sep 2025). More generally, the 2019 GPT-2 paraphrasing work has no explicit faithfulness loss and relies on post hoc filtering rather than controlled semantic constraints (Witteveen et al., 2019). CoQAR also shows that rewriting is not universally beneficial: for a strong RoBERTa conversational QA model, human or T5-generated rewrites hurt performance relative to using the original question plus context, even though the same rewriting helps a non-conversational DistilBERT SQuAD model on CoQAR by about +6.3 F1 (Brabant et al., 2022).

The open research directions follow directly from these limits. The VideoQA work suggests richer structural supervision, including event graphs or causal graphs instead of only unstructured narratives, as well as automatic checks for hallucination and logical inconsistency (Liang et al., 29 Sep 2025). CoQAR suggests jointly learning rewriting and answering rather than keeping them modular (Brabant et al., 2022). RoParQ points toward paraphrase-aware alignment and semantic invariance as explicit training targets rather than post hoc diagnostics (Choi, 26 Nov 2025). Taken together, these results indicate that QBP is most effective when paraphrastic reformulation is tightly coupled to the downstream semantic task—answering, retrieval, grounding, or event reconstruction—rather than treated as an isolated text-generation problem.