Question Interpretation Diversity

• Question interpretation diversity is the multiplicity of valid semantic readings and framings derived from a single text, underlining its role in generating varied questions.
• It employs structured methods such as explicit ontologies, latent variables, and discourse analysis to precisely control semantic focus and maintain answer relevance.
• Evaluation metrics balance set-level precision and coverage, ensuring that generated questions remain answerable, contextually appropriate, and aligned with intended interpretations.

Question interpretation diversity denotes the multiplicity of valid readings, question targets, and semantic framings that can arise from the same source text, answer, or user query. In question generation over text, the task is inherently one-to-many: a single context may support questions that differ in answer focus, semantic aspect, and wording (Schlichtkrull et al., 2020). In open-ended generation, this diversity is not adequately characterized by WH-words alone, because the same source can license questions about cause, consequence, procedure, comparison, example, or judgment (Cao et al., 2021). In information-seeking question answering, the issue reappears on the response side: an answer may address one interpretation of a question, several interpretations, or explicitly reject one (Srikanth et al., 16 Apr 2026). Across NLP more broadly, diversity is often invoked in an ad hoc way and with inconsistent terminology, which makes question interpretation diversity both a conceptual and an evaluation problem (Estève et al., 28 Jul 2025).

1. Conceptual scope and boundaries

Question interpretation diversity is not reducible to lexical variation. Several papers distinguish between surface diversity and semantic or cognitive diversity. In text-based question generation, multiple valid questions from the same context may differ because they ask about different entities, events, times, causes, or consequences rather than because they paraphrase one another (Schlichtkrull et al., 2020). In open-ended generation, the same input answer or text can support distinct types such as Verification, Concept, Example, Cause, Consequence, Procedural, and Judgmental, which shifts the notion of diversity from wording to interpretation (Cao et al., 2021). In knowledge-base question generation, the relevant criterion is that diverse questions should convey the same underlying meaning through varied expressions, not merely exhibit unique n-grams (Guo et al., 2023).

This distinction matters because standard accounts of diversity often collapse several phenomena into one label. The survey of diversity quantification in NLP organizes the field by asking why diversity is measured, what is diversified, where in the pipeline it is measured, and how it is quantified, and casts diversity in terms of variety, balance, and disparity (Estève et al., 28 Jul 2025). Within that framing, question interpretation diversity spans at least two settings. First, it appears as in-text diversity when multiple questions or answers differ in semantic content or rhetorical function. Second, it appears as diversity of processing when systems explicitly vary question types, prompting conditions, or answering strategies.

Ambiguity is one important source of interpretation diversity, but it is not the only one. In binary QA, semantically underspecified questions are common enough that a manual sample of BoolQ found 33% of questions ambiguous (Rosales et al., 25 Jul 2025). Yet even when a source is not ambiguous in that narrow sense, it may still support multiple legitimate question formulations that probe different aspects of the same material. This suggests that question interpretation diversity includes both ambiguity resolution and systematic exploration of alternative valid foci.

A common misconception is that more diverse questions are simply “more different” questions. The literature instead treats validity as central. Generated questions must remain answerable from the source, relevant to the intended semantics, or aligned with a specified question type; otherwise, diversity degenerates into off-target variation (Yoon et al., 2023).

2. Ontologies and representational frameworks

A major line of work formalizes question interpretation diversity through explicit ontologies. The most direct example is the 10-category question type ontology for open-ended question generation: Verification, Disjunctive, Concept, Extent, Example, Comparison, Cause, Consequence, Procedural, and Judgmental (Cao et al., 2021). These categories are presented as more fine-grained than question words, and they are sorted by levels of cognition from lower to higher. The ontology therefore treats interpretation depth as part of the representational problem rather than as a downstream by-product of wording.

This ontology supports a form of type-level control. Because a single WH-word can correspond to multiple types, and a single type can be expressed with multiple WH-words, type conditioning provides a more precise handle on semantic intent than interrogative form alone. The resulting perspective recasts diversity measurement: diversity can be evaluated over types rather than only over lexical or syntactic variation.

A complementary framework appears in discourse-level analysis of answering behavior. DiscoTrace represents an answer as a sequence of tuples

$DT=\{(a_1,i_1),(a_2,i_2),\ldots,(a_n,i_n)\},$

where each $a_j$ is a question-related discourse act and each $i_j$ is an interpretation of the original question when applicable (Srikanth et al., 16 Apr 2026). The framework uses a 21-act ontology grouped into five families: Answer the Question, Comment on the Question, Seek more Information, Redirect the Question, and No-op / non-substantive. Here, interpretation diversity is not only which question type could be asked, but which reading an answer segment chooses to address.

DiscoTrace also introduces a structural segmentation procedure built on rhetorical structure theory. Answers are parsed into RST trees, split at boundary relations such as Contrast, Comparison, Topic-Change, Evaluation, Summary, and Background, and then labeled with discourse acts and, where eligible, question interpretations (Srikanth et al., 16 Apr 2026). The significance of this design is that RST alone does not encode what part of a question an answer is responding to. Pairing rhetorical moves with interpretations adds that missing pragmatic layer.

Taken together, these frameworks define two complementary representational axes. One axis models prospective interpretation space at question-generation time through question types. The other models realized interpretation choice at answer time through discourse acts linked to readings of the original question. This suggests that question interpretation diversity is simultaneously a generation problem, a representation problem, and a pragmatics problem.

3. Generative modeling of diverse interpretations

Generative approaches operationalize question interpretation diversity by introducing latent semantic choices, explicit type controls, recursive conditioning, or external sources of varied expression. In early work on text-based question generation, a conditional variational autoencoder adds a latent variable $z$ to the decoder so that different samples correspond to different valid semantic choices of what question to ask (Schlichtkrull et al., 2020). The stated goal is not simply to perturb wording, but to model the one-to-many structure of the task. Empirically, the latent-variable model improves diversity-sensitive evaluation while maintaining similar fluency and relevancy in human judgments.

Type-aware open-ended question generation pushes control further. A BART-based framework conditions on a predicted question type $t$, builds a semantic graph from dependency relations and semantic role relations, predicts question focuses with graph attention, and jointly optimizes focus prediction and question generation with equal weights (Cao et al., 2021). The semantic graph is constructed using Stanford CoreNLP and AllenNLP outputs, with node merging for same-surface-form or coreference mentions. This architecture is designed to discover what to ask about in long, multi-sentence inputs, which directly affects interpretation diversity. Template-guided variants then increase controllability: ExplGen uses exemplar templates, while TplGen first generates a new template and then realizes it into a question. The paper reports that TplGen achieves the highest controllability and the highest diversity in type and content.

Narrative question generation over storybooks uses a different mechanism. The mQG model is trained to generate a question from context $C$, question type $QT$, and a history of previously generated questions $H_i$, and it recursively feeds earlier generated questions back into the model at inference time (Yoon et al., 2023). Its objective combines cross-entropy loss with a Maximum Question Similarity term,

$L=L_{CE}+L_{MQS},$

or, in the appendix, a weighted form $L=L_{CE}+\beta\cdot L_{MQS}$. The model is intended to produce multiple, diverse, and answerable questions from the same passage. A DeBERTa-base answerability model, first fine-tuned on SQuAD2.0 and then trained on FairytaleQA, is used to classify generated questions as Explicit, Implicit, or Unanswerable, thereby filtering semantic diversity from unsupported variation.

Knowledge-base question generation treats diversity primarily as semantically faithful multi-output realization. The dual-model framework couples a forward model $a_j$0 mapping KB subgraphs to questions with a backward model $a_j$1 mapping questions to triplet-based sequences, and iteratively refines both using pseudo pairs derived from external natural questions (Guo et al., 2023). Two selection strategies are used: relevance filtering with SimCSE at threshold $a_j$2, and diversity-aware pseudo-pair selection using the proposed diversity score. The authors explicitly distinguish this from ordinary paraphrasing, arguing that external natural questions provide more semantic patterns and expressions than minor lexical rewrites of a gold target.

Across these systems, the common design principle is that diversity is induced by structured choice. Latent variables choose semantic focus, type ontologies choose interpretation class, recursive histories discourage repetition, and external-question corpora inject broader realizational patterns. This suggests that question interpretation diversity is best modeled by mechanisms that expose alternative semantic decisions rather than by stochastic decoding alone.

4. Evaluation and quantification

Evaluation is a central difficulty because standard overlap metrics were designed for tasks in which multiple references are close paraphrases. For question generation, that assumption often fails. One context may license a “who” question, a “when” question, and a “what event” question, and these are not paraphrases of one another (Schlichtkrull et al., 2020). To address this, a set-to-set evaluation scheme generalizes precision and recall with a base similarity function $a_j$3:

$a_j$4

$a_j$5

$a_j$6

Here, $a_j$7 behaves like generalized precision, $a_j$8 like generalized recall, and $a_j$9 rewards both per-question quality and coverage of multiple valid interpretations (Schlichtkrull et al., 2020). The paper instantiates this with BLEU, ROUGE, and METEOR to obtain $i_j$0-, $i_j$1-, and $i_j$2-style variants.

A different critique targets diversity metrics themselves. In KBQG, Distinct-n is rejected as an inadequate proxy because it mainly measures duplication within generated text rather than whether a model can produce several relevant questions expressing the same semantics (Guo et al., 2023). The proposed Diverse@k instead sums pairwise token-set differences among the top-$i_j$3 generated questions, subject to a relevance constraint with respect to the ground-truth question:

$i_j$4

with $i_j$5 and relevance computed using SimCSE. Human validation reports Pearson correlation of 0.935 for Diverse@3 and 0.949 for Diverse@5 on WebQuestions (Guo et al., 2023).

Narrative QG uses Self-BLEU as a direct within-set redundancy measure, where lower Self-BLEU indicates higher diversity among questions generated from the same context (Yoon et al., 2023). That work also supplements automatic metrics with human ratings on Type, Syntax, and Content diversity, alongside Appropriateness and Answerability. The content axis is particularly important because it approximates whether different generated questions actually require different answers or address different narrative aspects.

The broader survey perspective shows why no single metric has become canonical. Across NLP, 150 distinct diversity measures were found, most of them specialized and often inconsistently named (Estève et al., 28 Jul 2025). The dominant families are counts of categories, distance-based measures, type-token-style ratios, and entropy-based measures. The survey also notes that NLP emphasizes variety and disparity far more than balance. For question interpretation diversity, this means that many current metrics capture either how many different forms appear or how dissimilar they are, but not whether the distribution across interpretation classes is well calibrated. This suggests a persistent measurement gap between “many interpretations exist” and “interpretations are covered in an appropriate proportion.”

5. Answering strategies, communities, and LLM behavior

Question interpretation diversity is not confined to question generation; it also structures how answers are built. DiscoTrace shows that nine Reddit communities—AskHistorians, AskEconomics, asklinguistics, history, NoStupidQuestions, OutOfTheLoop, explainlikeimfive, ScienceBasedParenting, and beyondthebump—have different preferences for answer construction (Srikanth et al., 16 Apr 2026). Expert communities such as AskHistorians, AskEconomics, and asklinguistics are more similar to one another than to generic QA forums. Shared topic does not guarantee shared strategy: history and AskHistorians differ substantially despite subject overlap, and ScienceBasedParenting differs from beyondthebump despite both being about parenting. The conclusion is that answer strategy depends on community norms, expertise expectations, and pragmatic goals, not only topic.

The same framework exposes a marked human–LLM divergence. LLM-generated answers are less community-specific in discourse-act sequence models and remain much less aligned with human self-perplexity than actual human answers, even when prompted with community guidelines (Srikanth et al., 16 Apr 2026). At the interpretation level, LLMs systematically address more readings of a question than humans do. When fewer than 10% of humans address a given interpretation, Claude Haiku still addresses it over 20% of the time, and Qwen over 40% of the time. Interpretation dedication shows the complementary pattern: humans devote more eligible segments to each addressed interpretation, with mean dedication 0.62 versus 0.46 for Haiku and 0.40 for Qwen. In the paper’s terminology, LLMs are broader but shallower, whereas humans are narrower but deeper.

This behavior has a direct algorithmic analogue in ensembling. For binary QA, one study compares model diversity—multiple LLMs answering the same interpretation—with question interpretation diversity—the same LLM answering several semantically distinct interpretations and combining the results by majority vote (Rosales et al., 25 Jul 2025). The interpretation-generation step uses GPT-3.5 to produce at least three distinct semantic interpretations, after which each interpretation is answered separately and summarized with a forced yes/no output. Across BoolQ, StrategyQA, and PubMedQA, question interpretation diversity consistently leads to better ensemble accuracy than model diversity. The reported mean delta from ensemble to best member is -3.73 for ChatGPT model-diverse ensembles and -10.53 for LLaMA model-diverse ensembles, versus -1.35 for ChatGPT interpretation-diverse ensembles and 11.12 for LLaMA interpretation-diverse ensembles. The paper’s conclusion is that model diversity often yields results between the best and worst members without clear improvement, whereas interpretation diversity more directly targets ambiguity in the input.

A recurring controversy concerns comprehensiveness. It is easy to assume that answering more interpretations is always better. The discourse evidence contradicts that assumption: overanswering can function as hedging or deluge rather than pragmatic adequacy (Srikanth et al., 16 Apr 2026). For QA systems, this implies that selecting the right interpretation may be more important than maximizing breadth.

6. Data design, generalization, and unresolved issues

Question interpretation diversity also appears in dataset construction and training curricula. In visual question answering, D3: Data Diversity Design studies diversity in the compositional structure of questions rather than in image style or vocabulary, distinguishing attribute composition diversity from length diversity measured by hop count (Rahimi et al., 2023). The core result is that diverse simple tasks are especially valuable for systematic generalization. On attribute-composition tasks, adding D3(0-Hop) to a biased training set yields large gains, including MAC from 46.56 to 71.09 and FiLM from 33.57 to 64.21 in the Two attributes setting. On GQA, replacing 30% of training questions with shorter ones improves long-test accuracy by +1.94 for MAC and +1.35 for a GPT-2-style transformer. The paper argues that it may not be essential to gather a large and varied number of complex tasks; strategically chosen simple but diverse questions can suffice or do better.

These findings broaden the notion of question interpretation diversity from output multiplicity to training support. A model may generalize better not because it has seen more difficult questions, but because it has seen more varied ways in which questions combine attributes, types, and reasoning structures. This suggests that interpretation diversity is partly a coverage problem over compositional spaces.

Several limitations remain stable across the literature. In type-controlled open-ended generation, some categories are intrinsically ambiguous, and the authors explicitly suggest future multi-class or multi-label type identification because some questions genuinely fit multiple types; answer-to-type prediction is also relatively weak, with macro F1 of 0.48 on Yahoo-trained and 0.46 on Reddit-trained classifiers (Cao et al., 2021). In narrative generation, generating more questions improves coverage but lowers diversity, and the reported best balance is four questions per section per question type (Yoon et al., 2023). In interpretation-based QA ensembling, the evaluation is restricted to binary questions, inference cost scales roughly linearly with the number of interpretations, and majority voting is deliberately simple and unweighted (Rosales et al., 25 Jul 2025).

The survey perspective indicates why these limitations are persistent. Diversity in NLP is motivated by both ethical and practical concerns, can function either as a goal or as a means, and is not uniformly desirable in all tasks (Estève et al., 28 Jul 2025). For deterministic or factual settings, excessive diversity may be counterproductive. Applied to question interpretation diversity, this implies that the central problem is not maximizing variation indiscriminately, but deciding which semantic alternatives ought to be represented, measured, or answered.

The cumulative picture is that question interpretation diversity is a structured property of both questions and answers. It can be represented through type ontologies and discourse-act/interpretation pairings, induced through latent variables, semantic graphs, recursive decoding, or external-question corpora, evaluated through set-level coverage or diversity-with-relevance metrics, and exploited to improve systematic generalization and ensembling. The literature consistently rejects the idea that “diversity” is a unitary target. Instead, it treats question interpretation diversity as the controlled management of multiple valid meanings, with answerability, relevance, and pragmatic fit as the constraints that keep that multiplicity useful.