R2DQG: A Quality Meets Diversity Framework for Question Generation over Knowledge Bases

1 Introduction

Generating natural language questions based on a set of formatted facts is the core objective of Knowledge-Based Question Generation (KBQG) [Guo et al., 2024a]. Over the last decade, KBQG has garnered substantial research interest across diverse fields, exemplified by its applications in academia and industry. In intelligent tutoring systems, KBQG plays a vital role, as educational questions or quizzes are useful for student assessment and coaching purposes [Agrawal et al., 2024]. Furthermore, in industry, KBQG can potentially be used for generating high-quality QA pairs, thereby enhancing the corpus of training data for KnowledgeBased Question Answering (KBQA) tasks [Li et al., 2024; Lyu et al., 2021].

∗Corresponding author.

Over the years, several methods have been developed to address the KBQG task, each focusing on specific aspects of question generation. Early rule-based approaches relied on predefined linguistic templates and graph traversal rules, generating interpretable but limited questions due to their rigidity and lack of scalability [Jia and Liang, 2016; Seyler et al., 2017]. Later, Seq2Seq models introduced encoder-decoder architectures to model relationships between subgraphs and natural language questions, improving flexibility [Chen et al., 2023; Elsahar et al., 2018]. However, these models often struggled to produce diverse syntactic structures. With the advent of pretrained language models (PLMs), such as BERT, T5, and GPT, KBQG has seen significant advancements in generating semantically relevant questions [Kim et al., 2019; Guo et al., 2022]. Nonetheless, these models frequently prioritize surface-level similarity to ground truth at the expense of linguistic diversity. Meanwhile, diversity-driven approaches employ mechanisms like external knowledge inject, content selectors, and diverse decoding strategies to introduce syntactic and lexical variety into the generated questions [Guo et al., 2024c; Bi et al., 2020; Narayan et al., 2022].

Despite these advancements, KBQG systems face several critical challenges. First, many approaches suffer from a lack of semantic guidance, as the templates that guide question structure are often unavailable until the generation process is complete. This makes it difficult to provide token-level guidance during generation, leading to contextually imprecise questions. Second, existing methods follow an inflexible generation process, where suboptimal subgraph-question pairs are discarded instead of refined. Refining "deficient" drafts is often more practical and efficient than generating well-formed questions from scratch, but current models lack the capability to perform such corrections effectively. Finally, KBQG systems face the challenge of conflicting objectives in balancing linguistic diversity and semantic relevance. These goals often conflict, requiring innovative strategies to resolve the trade-off and generate both high-quality and diverse questions simultaneously.

To address these challenges, we propose R2DQG 1 : RefineReinforced Diverse Question Generation, a novel two-phase framework that effectively balances linguistic diversity and semantic relevance in Knowledge-Based Question Generation. At its core, R2DQG follows a generation-thenrefinement paradigm inspired by human question-drafting processes, ensuring flexibility and robustness. Figure 1 illustrates the R2DQG workflow, comprising two key components: the Generator and the Corrector. The Generator first constructs a diverse set of expressive templates by leveraging dependency parse tree similarity, capturing varied syntactic structures and phrasing styles. These templates serve as guiding scaffolds to generate diverse yet semantically aligned question drafts. The Corrector then refines these drafts, mitigating semantic drift and enhancing coherence, resulting in well-formed, high-quality questions. By decoupling the generation and refinement processes, R2DQG effectively balances diversity and relevance while offering a scalable and practical solution. Our framework overcomes existing limitations with a novel drift compensation mechanism that refines imperfect drafts for enhanced expressiveness and accuracy. Extensive experiments and case studies on two datasets validate the effectiveness of R2DQG through benchmark evaluations and human assessments.

To sum up, the contributions of our work are as follows:

• We propose R2DQG, a novel post-hoc KBQG framework inspired by human drafting and revising processes, to generate diverse and semantically accurate questions using a two-phase generation-then-refinement paradigm.
• We design a template-aware Generator to enhance syntactic and stylistic diversity, coupled with a Corrector module to refine drafts, ensuring semantic alignment and linguistic richness.
• Extensive experiments on public datasets validate that R2DQG achieves superior performance on diversity, while its synthetic datasets significantly boost downstream QA performance.

2 RelatedWork

Knowledge Base Question Generation. The field of KBQG has witnessed significant advancements. Early rule-based methods relied on predefined linguistic templates and graph traversal rules, generating interpretable but limited questions due to their rigidity and lack of scalability [Jia and Liang, 2016; Seyler et al., 2017]. With the emergence of data-driven learning approaches, Seq2Seq model introduced encoderdecoder architectures to model relationships between subgraphs and natural language questions, improving flexibility [Chen et al., 2023; Elsahar et al., 2018]. With the advent of pretrained language models (PLMs), such as BERT, T5, and GPT, KBQG has seen significant advancements in generating semantically relevant questions [Kim et al., 2019; Guo et al., 2022; Chen et al., 2023; Fei et al., 2022]. LLMsbased methods [Wang et al., 2024], with their powerful generation capabilities, have further boosted KBQG. Techniques such as skeleton-guided prompting [Guo et al., 2024b], incontext demonstrations [Liang et al., 2023] and multi-agent collaborative frameworks [Zhao et al., 2024] inspired by memory mechanisms [Dang et al., 2024a; Dang et al., 2024b] have been used to guide the model generation.

1 Codes are here: https://github.com/Elaine-explorer/R2DQG.

Diversifying Question Generation. Diversity in question generation remains a critical challenge, with research efforts focused on two dimensions: leveraging internal knowledge and exploiting external patterns. The former make use of internal knowledge such as content selection [Wang et al., 2020] and apply various decoding algorithms to promote diversity, such as diverse beam search [Narayan et al., 2022], nucleus sampling [Holtzman et al., 2019], entmax transformation [Martins et al., 2020] and well-designed loss functions [Zhang and Zhu, 2021]. And those that exploit external patterns such as fact-infused [Deschamps et al., 2021], question type ontology [Cao and Wang, 2021], choose reliable pesudo pairs [Guo et al., 2024c].

Most similar to our work is [Guo et al., 2024c]. However, it fails to explicitly guide generation following syntactic structure and focuses on directly generating valid questions. Contrary to this, our focus is to adopt templates contain useful syntactic structures that help organize questions well. Besides, we refine specific expressions that align with quality-meets-diversity control conditions within semantic space, thereby guiding the generated questions toward highquality outputs on the premise of the diverse expression ways.

3 Methodology

We first formalize the KBQG problem statement and give an overview of our R2DQG. Then we elaborate on the details of two individual components in the following sections.

3.1 Task Definition andModel Architecture

Problem Formulation. The KBQG task aims to generate a natural language question ˆ qi = (w1, . . . , wm), which can be expressed as optimizing the model parameter Θ to maximize the conditional likelihood P . This can further be decomposed into a sequential word-by-word generation process:

where Gi = {k(i) 1 , . . . , k(i) n } represents the subgraph comprising a set of triples, ai signifies the target answer, and qi denotes the generated question.

Model Overview. The overall R2DQG framework, illustrated in Figure 2, comprises two key modules: the Generator and the Corrector. The Generator consists of two components: (a) Diversified Templates Acquisition, which selects expressive templates based on tree kernel similarity to enhance diversity, and (b) Template-guided Question Generation, which leverages these templates to generate diverse question drafts while maintaining syntactic variety. However, these drafts may experience semantic drift, necessitating further refinement. Hence, the Corrector introduces (c) Drift Compensation, a post-editing mechanism that refines initial drafts into high-quality, semantically accurate questions. This two-stage process ensures that the generated questions achieve both linguistic diversity and semantic precision.

3.2 Generator

Now we deal with the first challenge of deriving diversified question drafts with the guidance of syntax templates. To achieve that, we propose a simple but effective templateguided Generator, which collect and filter diversified template by hierarchically expanding constituents in syntax contexts throughout the syntax tree. The resulting templates, containing rich semantic information, can serve as reliable grammatical guidance to assist in the generation of linguistically diverse question drafts.

Diversified Templates Acquisition. The templates play an important part in guiding question generation. As depicted in Figure 2(a), we propose a generate-then-filter strategy to extract diverse template training samples from existing QG datasets. For each input D = (Gi, ai, qi), we first linearize it into a textual prompt, and then feed it into the closed-source LLMs to generate diverse candidate templates. Then, we select the top-K results of the beam search as candidate templates. However, solely relying on LLMs' capability makes it hard to guarantee the effectiveness of the generated templates and may introduce bias. To mitigate errors made by LLMs, we impose constraints grounded in Dependency Parse Tree Similarity (DPTS) to filter templates by jointly computing their semantic and syntactic similarity. We use a neural parser [Qi et al., 2020] to generate parse trees from sentences, which are then transformed into graph representations. These dependency parse trees (DPT) capture syntactic relationships, where nodes represent words and edges denote dependencies. Unlike n-gram-based metrics, DPT-based representations enable the generation of templates that are both diverse and semantically valid. To compute similarity, we adopt the efficient Attention Constituency Vector Tree (ACVTree) kernel [Quan et al., 2019], which quantifies similarity based on shared substructures between parse trees. The tree kernel similarity between two trees T1 and T2 is calculated as follows:

In this expression, NT1 and NT2 are the sets of nodes in T1 and T2, respectively. The function ∆(n1, n2), which is simplified from ∆(·), measures the similarity between two nodes n1 and n2, capturing both semantic and structural features, and is computed as:

Let vec1 and Attw1 (resp., vec2 and Attw2 ) denote the word vector and attention weights of node n1 (resp., n2). The cosine similarity between the vectors of the leaf nodes is computed by the function SIM(·). Let cn1 (resp., cn2 ) represent the list of child nodes of n1 (resp., n2). The term ∆p(.) counts the number of common subsequences of length p. The parameters λ and µ are decay factors controlling the influence of the child sequences' lengths and the tree height, respectively.

Finally, we select exemplars of input-template with high DPTS scores to form the training dataset, denoted by Dtrain = {(Gi, ai, ti)}Ni=1 .

Template-guided Question Generation. Afterwards, we focus on generating questions according to the input subgraph, the expected answer, and the template, in which the input sub-graph offers contextual information to determine what to ask, the answer reflects the asking direction to indicate what is the target, and the template guides how to ask in a reasonable manner. Specifically, we fine-tune the LLaMAbased Template Generator to automatically produce question templates using the constructed dataset Dtrain. To achieve efficient adaptation, we freeze the parameters Φ of LLaMA and incorporate low-rank adaptation (LoRA) [Hu et al., 2021] adapters. Formally, we generate template ti conditioned on the given input (Gi, ai), where the objective L′ is defined as:

Building upon the question templates, we utilize LLMs to seamlessly integrate them into the test input, steering the model to generate the initial question drafts qd , as shown in Figure 2(b).

The diversified initial question drafts may suffer from "semantic drift", where the generated questions deviate from the provided input subgraph and answer, requiring further adjustments for alignment. Rather than preventing drift, our R2DQG framework introduces a semantic drift compensation method, called Corrector, which learns the corrective residuals between the desired question and the imperfect draft to effectively mitigate the issue of semantic drift.

Similar to how residual blocks enhance an architecture without altering its foundation, the Corrector refines initial question drafts through a flexible copy-and-refine mechanism. To train the Corrector, we first create example pairs that map deficient questions to well-formed ones. As shown in Figure 2(c), the input subgraph Gi, target answer ai, prior question draft qd i , and ground-truth question qg i are collected to construct the data pool as follows:

The semantic alignment pool DC is applied to fine-tune LLaMA, denoted as µϕ(qg i |Gi, ai, qd i ), where the parameters ϕ are trained to align the generated questions with the groundtruth qg i . The probability distribution for generating the refined question can be represented as:

where qk is a possible draft question generated by the template-guided Generator πθ.

By calculating the empirical loss over the dataset DC, we get the following constraint:

Since the second term does not involve ϕ, the training objective can be derived as:

Corrector's Training Strategy - Residual Refinement. Now that the generated question drafts qd i may vary significantly in quality. Some drafts are close to the target questions qg i and do not require further correction. Other drafts may exhibit semantic drift or deviate significantly from the desired target, requiring substantial adjustments. Treating all drafts uniformly during correction not only increases training complexity but also risks over-correcting high-quality drafts. To address this, we first construct a G-A-QG -QG dataset using a subset of the training data to pre-train the R2DQG model, a process we refer to as warm-up. This phase helps the model establish a foundational ability to handle question drafts with minor semantic drift. Subsequently, we further train the model using the G-A-Q-QG dataset to enhance its correction capabilities. This training strategy aligns with techniques applied beyond the question generation domain. In particular, ResNet employs residual connections to mitigate the vanishing gradient problem in deep neural networks, conceptually resembling our drift compensation strategy.

During inference, a question qj is sampled from the conditional probability distribution µϕ, and the model is then tasked with predicting the corrected question qj .

The Corrector is a versatile, pluggable component that operates independently of the preceding Generator's parameters, πθ. Instead of generating newquestions directly from the input, it focuses on learning the alignment between the distribution of initial question drafts and the desired questions, effectively mitigating semantic drift.

4 Experiment

4.1 Experiment Setup

Datasets. Two public benchmark datasets are used to evaluate R2DQG: WebQuestions (WQ) [Kumar et al., 2019] and PathQuestions (PQ) [Zhou et al., 2018]. Specifically, WQ combines instances from WebQuestionsSP and ComplexWebQuestions, both of which are benchmarks for KBQA. Specifically, WQ consists of train/dev/test set with 18,989/2,000/2,000 instances, while PQ contains 9,793/1,000/1,000 instances.

Metrics. We introduce a diversity metric with semantic constraints, termed Diverse@k (D@K) [Guo et al., 2024c], to assess the diversity of the top-k generated questions while maintaining their relevance to the ground truth. We also adopt Distinct-n (D-n) [Song et al., 2019], which measures the diversity of generated text by calculating the proportion of unique n-grams within the output. We adopt 2 traditional metrics, namely BLEU 1 (B-1)[Papineni et al., 2002] and METEOR (ME) [Banerjee and Lavie, 2005], which compute the ratios of the common n-grams between the generated question and the ground truth. For QA Evaluation, we use Hits@1 to evaluate whether the top-1 predicted answer is accurate and report the F1 score. For Human Evaluation, three welleducated annotators are employed to measure the Fluency, Relevance and Diversity of the generated questions.

Baselines. we choose three categories of methods as baselines. (i) Graph2Seqs methods, including G2S+AE and G2S+AE+RL, aim to capture the complex relationship between entities, thereby improving the knowledge consistency of the generated question. (ii) PLMs methods, including MHQG, MHQG+AE, JointGT, DSM, DiversifyQG, T5 and Bart, adopt PLMs to generate question by fine-tuning them on specific tasks or datasets. and (iii) LLMs methods, including GPT-4-Turbo, GPT-3.5-Turbo, KQG-CoT, RoleAgentQG, SGSH and SGSH(B+S), design specific natural language prompts to directly influence question generation.

Implementations. All experiments are implemented using PyTorch on a server equipped with 8 NVIDIA GeForce RTX 3090Ti GPU. We select GPT-3.5-turbo-1106 as the default base LLM for the zero-shot setting and GPT-3.5-turbo-16k for the few-shot setting. For the Generator and Corrector that need to be fine-tuned, we employ LLaMA as the backbone and fine-tune them using LoRA. The Generator and Corrector are trained for 3 epochs with a learning rate of 1e-4. Empirically, 20% warm-up samples yield the best performance.

4.2 Performance Comparison

The question generation performances on WQ and PQ are presented in Table 1 and Table 2. The observations from these comparisons are as follows: (1) Regarding the average score, R2DQG achieves the superior results on two datasets, demonstrating the overall high relevance and diversity of the generated questions. (2) Compared with other baselines, our R2DQG method also demonstrates significant improvements regarding the diversity aspect. Notably, considering the top3 question generation, R2DQG reports up to a 16% increase in Div@3 on the WQ dataset and a 14% increase on the PQ dataset. This can be attributed to the well-designed template generator, which effectively guides the approximate syntactic structure of the generated questions. (3) Notably, the improvements in diversity by R2DQG do not extend equally to quality aspect. This is acceptable, as the R2DQGmodel often introduces varied forms of expression in pursuit of greater diversity. By design, R2DQG aims to strike an optimal balance between two interrelated yet conflicting objectives-ensuring diverse question expressions while minimizing quality loss. (4) Directly prompting LLM methods such as GPT-4-Turbo and GPT-3.5-Turbo, are proficient at producing a wide range of questions, they face challenges in maintaining high-quality outputs in terms of BLUE-1 and METEOR. (5) Except for DiversifyQG [Guo et al., 2024c], all other baselines exhibit lower diversity, as they tend to generate questions closely resembling the gold standard. This highlights that the absence of effective diversification guidance can significantly degrade performance in diversity metrics. Our results align with those of DiversifyQG.

4.3 Corrector Evaluation

We evaluate the impact of the key Corrector module in our R2DQG model by removing it while keeping all other components unchanged. To assess whether the refined questions achieve the desired relevance and diversity, we analyze the distribution shift before and after the correction process. For this evaluation, we randomly selected 300 samples from different datasets and reported the averaged results in Figure 3.

Subgraph: <north atlantic treaty organization (nato),member,bulgaria>, <bulgaria,appointed by,rosen plevneliev>

Answer: Bulgaria

Ground Truth:what country that is part of north atlantic treaty organization ( nato ) appointed rosen plevneliev to a governmental position?

Q1: what country that is part of north atlantic treaty organization ( nato ) had rosen plevneliev serving in an appointed capacity?

Q2: what country that is part of north atlantic treaty organization ( nato ) had rosen plevneliev serving in an appointed role?

Q3: what country that is part of north atlantic treaty organization ( nato ) had rosen plevneliev serving in an appointed position?

Q1: which nation in the north atlantic treaty organization (nato) appointed rosen plevneliev to a government role?

Q2: rosen plevneliev was appointed to a government post by what country, a member of north atlantic treaty organization (nato)?

Q3: which north atlantic treaty organization (nato) member country appointed rosen plevneliev to a governmental position?

Q1: which _ country that appointed _ to a position?

Q2: which nation that appointed _ to and is a part of _?

Q3: what country that appointed _ to a governmental position and is also a member of _?

Q1: which member country of the north atlantic treaty organization (nato) that appointed rosen plevneliev to a governmental position?

Q2: which nation that appointed rosen plevneliev to office and is a part of the north atlantic treaty organization (nato)?

Q3: what country that appointed rosen plevneliev to a governmental position and is also a member of the north atlantic treaty organization (nato)?

Subgraph: <tennessee river,origin,french broad river>,<tennessee river,mouth,ohio river> Answer: Ohio River

Ground Truth: where does the river that begins as french broad river flow into?

Q1: which north atlantic treaty organization (nato) country appointed rosen plevneliev to a governmental position?

Q2: which nation is a part of north atlantic treaty organization (nato) that appointed rosen plevneliev to govermental position?

Q3: what country that appointed rosen plevneliev to a governmental position and is a part of north atlantic treaty organization (nato)?

Q1: where does the river that has its origins in the french broad river merge with waterway?

Q2: where does the river that has its origins as the french broad river empty into?

Q3: where does the river that has its begins with the confluence of the french broad river empty into?

Q1: what is the final destination of the river that starts as the french broad?

Q2: where does the river, known as the french broad, merge into?

Q3: can you tell me where the french broad river converges?

Q1: where does the river originating from the french broad river and flowing through the tennessee river end?

Q2: what is the mouth of the river that originates from the french broad river?

Q3: the river that begins at the french broad river eventually joins which major river?

Q1: where does the river originating from the french broad river end?

Q2: what is the mouth of the river that originates from the french broad river and flows into?

Q3: the river that begins at the french broad river and flows into which river?

Figure 4: Comparison of top-3 generated questions on WQ. Distinct colors highlight different surface forms for each instance (indicating diversity), and underlined words denote expressions that match the ground truth (indicating quality).

Based on our comprehensive analysis of the experimental results, we have identified several important findings. (1) The correction process leads to higher BLEU-1 scores, demonstrating that the refined questions effectively mitigate the semantic drift problem and align more closely with the target distribution. (2) While the refinement process enhances question quality, it leads to a slight decrease in linguistic diversity, suggesting a necessary trade-off where the Corrector ensures relevance while maintaining creative variation in question generation. (3) Our R2DQGmethod exhibits amore concentrated distribution across multiple experiments, reflecting improved stability and consistency in performance. This further validates the effectiveness of R2DQG in achieving a balanced trade-off between stability and diversity. These analytical findings underscore the importance of the drift compensation strategy in addressing distribution shifts, highlighting its crucial role in optimizing KBQG systems.

4.4 Case Study

To intuitively demonstrate the effectiveness of R2DQG, Figure 4 presents a comparison of the top-3 generated questions for two examples from the WQ dataset using three methods. Key observations include: (1) We find that questions generated from SGSH are highly consistent in both structure and content, e.g. "what country" and "serving in an appointed" frequently appear across different questions. (2) By contrast, DiversityQG excels in generating express diverse questions, marked with two or three distinct colors. By overly prioritizing linguistic diversity, the model sometimes generates unrelated phrase expressions such as "can you tell me", which may detract from the question's relevance. (3) Notably in R2DQG, the corrected questions demonstrate superior relevance and diversity compared to its draft and other baseline approach results, highlighting that the drift compensation process not only improves the relevance of the generated

Q1: where does the river from _ end?

Q2: what is the mouth of the river that _?

Q3: the river that begins at _ eventually joins？

question but also maintains diverse expressions. The ability to maintain diversity can be attributed to the strong foundation provided by the initial question drafts. Thus, findings from the case study further reinforce our conclusions.

4.5 DownstreamQA Evaluation

To address the issue of insufficient manual annotation corpus during the model training process, KBQG systems are developed to generate synthetic QA pairs and augment QA datasets [Sultan et al., 2020]. Along this line, we evaluate the performance of two typical QA model (i.e., GRAFT-Net [Sun et al., 2018] and NSM [He et al., 2021]) on WebQSP test data with 2,848 (question, answer) training instances. Since 1,409 (question, answer) pairs in the training data of WebQSP overlap with those in WQ, we can just take their corresponding subgraphs from WQ. With these subgraphs, KBQG models can produce their corresponding questions. To evaluate the quality of the augmented questions, we conducted experiments using data augmented by both baselines and R2DQG, as well as the original WebQSP data (referred to as ORI). We denote the new datasets as "ORI + baseline" and "ORI + R2DQG," respectively. On these augmented datasets, we train GRAFT-Net and NSM separately on "ORI + baseline," "ORI + R2DQG," and the original WebQSP data (ORI) to compare their performance. The observations from Table 3 are as follows: (1) The generated additional QA data pairs can be viewed as an approach to data augmentation for KBQA, as both GRAFT-Net and NSM trained on datasets augmented by different KBQG models can enhance their QA performance. (2) We notice that QA models trained on datasets augmented by R2DQG obtain the best results, indicating that the question generation synthetic datasets significantly enhance downstream QA tasks.

4.6 Human Evaluation

To systemically assess our R2DQG, human evaluation is recruited to investigate the generated question. We followed the existing work to select three evaluation dimensions: Fluency (whether a question has no grammatical errors and is fluent in expression), Relevance (whether a question clearly describes the given subgraphs) and Diversity (whether a question express the same semantics with ground truth, but in different expression forms). We randomly sample 100 instances from the testing set of the WQ dataset, and collected the generated results from R2DQG and several competitive baselines for comparison. Three master's students were recruited as annotators to rate each question in a blind setup. Each question was rated independently on a Likert scale ranging from 1 to 5, with 5 indicating the highest quality. We calculated Kendall's tau coefficient to assess how closely the annotators agreed. The resulting value of 0.83 demonstrates a strong level of agreement among them. We average the scores from raters on each question. As shown in Table 4, our R2DQG consistently achieves the best performance and produces questions that are more fluent, diverse, and relevant. Particularly, we observe that R2DQG surpasses baselines on diversity evaluation, which shows that our model equipped with the template generator can produce more diversified questions. This is consistent with the observations from our main results.

5 Conclusion

Recent advancements in KBQG have primarily focused on improving question quality, while the challenge of achieving diversity remains largely unexplored. We argue that KBQG inherently follows a one-to-many mapping paradigm, making diverse yet relevant question generation essential. To address this, we introduced R2DQG, a framework that combines template-guided generation with a drift compensation mechanism to produce high-quality and diverse questions. Our approach utilizes diversified templates to guide question structure and employs drift compensation to refine biased drafts into coherent and contextually relevant questions without human intervention. Moreover, R2DQG serves as an effective data augmentation tool, enhancing the robustness of QA models. Extensive experiments on the WQ and PQ datasets demonstrate the effectiveness of our approach in balancing quality and diversity. To address the current limitations in handling large and complex subgraphs in KBQG, we leave this challenge to future work, where we intend to incorporate GNN-based techniques [Lu et al., 2024] for more effective subgraph representation and question generation. In addition, we will focus on extending the framework to more complex domains and exploring adaptive strategies to further enhance question generation.

Acknowledgements

The research was supported in part by the National Natural Science Foundation of China (No.U22B2019) and the China Scholarship Council program (Project ID: 202406470057). We thank the anonymous reviewers for their comments.