PaperlayerView exact revision on arXiv

Source-linked AI summary

CLAMBER: A Benchmark of Identifying and Clarifying Ambiguous Information Needs in Large Language Models

Tong Zhang, Peixin Qin, Yang Deng, Chen Huang, Wenqiang Lei, Junhong Liu, Dingnan Jin, Hongru Liang, Tat-Seng Chua

arXiv:2405.12063v2cs.CL

TL;DR

The effectiveness of LLMs on ambiguous user queries remains uncertain, although such ambiguity can put user trust and satisfaction at risk. CLAMBER benchmarks ambiguity identification and clarification using a taxonomy and ~12K data, and finds that current LLMs still struggle with both tasks.

  • Problem

    The effectiveness of LLMs on ambiguous user queries remains uncertain, even though ambiguity can risk user trust and satisfaction.

    Source: §1 · p. 1

  • Method

    CLAMBER evaluates LLMs on identifying and clarifying query ambiguity using a taxonomy with three dimensions, eight categories, and ~12K data.

    Source: §1 · pp. 1–2

  • Results

    Current LLMs still struggle to achieve optimal performance in identifying and clarifying ambiguity, limiting their practical utility in advanced information search applications.

    Source: §7 · p. 9

  • Takeaways & Limitations

    CLAMBER provides guidance for evaluating and addressing ambiguous information needs and a foundation for enhancing LLMs’ proactive capabilities.

    Source: §1 · p. 2, §7 · p. 9

    Evaluation results may be prompt-sensitive, and the benchmark covers only five LLMs due to computational constraints.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large language models (LLMs) are increasingly used to meet user information needs, but their effectiveness in dealing with user queries that contain various types of ambiguity remains unknown, ultimately risking user trust and satisfaction. To this end, we introduce CLAMBER, a benchmark for evaluating LLMs using a well-organized taxonomy. Building upon the taxonomy, we construct ~12K high-quality data to assess the strengths, weaknesses, and potential risks of various off-the-shelf LLMs. Our findings indicate the limited practical utility of current LLMs in identifying and clarifying ambiguous user queries, even enhanced by chain-of-thought (CoT) and few-shot prompting. These techniques may result in overconfidence in LLMs and yield only marginal enhancements in identifying ambiguity. Furthermore, current LLMs fall short in generating high-quality clarifying questions due to a lack of conflict resolution and inaccurate utilization of inherent knowledge. In this paper, CLAMBER presents a guidance and promotes further research on proactive and trustworthy LLMs. Our dataset is available at https://github.com/zt991211/CLAMBER

1 Introduction

CLAMBER addresses fragmented understanding of ambiguity in LLM queries with a taxonomy-based benchmark of approximately 12K examples. Its evaluation finds that current models still struggle to identify ambiguity and ask effective clarifying questions.

1 Introduction

  • Ambiguous wording can point to different user needs, so answering without identifying or clarifying it risks misalignment with users’ needs, trust, and satisfaction.The example query “what are the strategies for saving?” could mean saving money or saving from sins.Source: §1 · p. 1
  • Earlier investigations lacked a comprehensive taxonomy and often focused on contextual ambiguity, leaving it unclear which ambiguities LLMs can identify and clarify effectively.The authors describe prior work as fragmented, with incomplete and inconsistent coverage of ambiguity distributions.Source: §1 · p. 1
  • CLAMBER organizes query ambiguity across three dimensions and eight categories, then uses approximately 12K data points to assess off-the-shelf LLMs.The taxonomy combines input-understanding and task-completion perspectives.Source: §1 · pp. 1–2
  • ChatGPT outperforms other small-scale LLMs, especially on multifaceted queries, but current models still struggle with ambiguity; CoT and few-shot prompting yield only marginal identification gains and may increase overconfidence.Models also struggle to use contextual cues to disambiguate pronouns.Source: §1 · p. 2
  • Even when models recognize ambiguity, poor conflict resolution and inaccurate use of inherent knowledge undermine their ability to ask high-quality clarifying questions.These difficulties leave models uncertain about which ambiguity to clarify.Source: §1 · p. 2
  • The authors present CLAMBER as a resource for guiding evaluation and future work on addressing ambiguous information needs.They say their findings shed light on why current models struggle to identify and clarify ambiguities.Source: §1 · p. 2

2 Related Works

The related work describes gaps in ambiguity taxonomies and in studies of LLMs’ ability to resolve ambiguous queries. CLAMBER uses its taxonomy to evaluate prompting strategies across a broader range of ambiguity types.

2 Related Works

  • Earlier taxonomies were fragmented or underdeveloped, while later ones used limited factual questions and lacked precise category definitions [Min et al., 2020].The paper also notes that existing taxonomies predated LLMs and did not account for ambiguity from conflicting interpretations based on inherent knowledge.Source: §2 · p. 2
  • Prior CoT and few-shot studies showed some gains on specific ambiguity types, such as lexical ambiguity, and CLAMBER evaluates these strategies across a broader range [Deng et al., 2023b; Kuhn et al., 2023; Cole et al., 2023].Source: §2 · p. 2
  • CLAMBER’s taxonomy combines input-understanding and task-completion perspectives to categorize query ambiguity and support analysis of approximately 12K examples.The paper presents this as a response to the lack of a well-organized taxonomy for information retrieval ambiguity.Source: §2 · p. 2

3 CLAMBER Benchmark

CLAMBER organizes query ambiguity across three dimensions and eight categories, covering both input understanding and task completion. It pairs this taxonomy with roughly 12K examples and expert validation.

3 CLAMBER Benchmark

  • The benchmark defines Epistemic Misalignment, Linguistic Ambiguity, and Aleatoric Output as three dimensions divided into eight categories.Its entries contain a query, a binary ambiguity label, and a clarifying question for ambiguous queries.Source: §3 · p. 3

3.1 Epistemic Misalignment (EM)

  • Epistemic Misalignment covers unfamiliar entities or facts and contradictions between queries and examples, using fabricated-entity and AmbiTask data.CLAMBER generates clarifying questions for unfamiliar entities and uses templates for contradiction cases.Source: §3.1 · pp. 3–4

3.2 Linguistic Ambiguity (LA)

  • Linguistic Ambiguity distinguishes multiple meanings of terms from context-poor interpretations; syntactic and pragmatic ambiguities are omitted as uncommon in information retrieval.Lexical examples draw on AmbER and AmbiPun, while referent ambiguity is addressed using AmbiCoref.Source: §3.2 · p. 4

3.3 Aleatoric Output (AO)

  • Aleatoric Output covers well-formed queries missing essential details, grouped by missing personal, spatial, temporal, or other information.Its categories are constructed from AmbigQA and Dolly-16K queries, with ambiguity labels and category assignments checked as described.Source: §3.3 · pp. 4–5

3.4 Validation and Revision

  • Five linguistic experts validate and revise ambiguity labels and clarifying questions, with a fifth expert resolving discrepancies among four reviewers.The final expert reviews feedback and makes necessary revisions.Source: §3.4 · p. 5

4 Experimental Design

The experimental design evaluates off-the-shelf LLMs on ambiguity identification and clarifying-question tasks. It compares four prompting schemes using a category-balanced random subsample and averages results across three prompts.

Test Dataset

  • The evaluation covers identifying ambiguities and asking clarifying questions on a randomly selected 3,600-instance subsample with equal category representation.There are 200 positive and negative examples per category, except Aleatoric Output has 800 negative examples.Source: §4 · p. 5

Usage of LLMs

  • Five models are evaluated: Llama2-13B-Chat, Llama2-13B-Instruct, Vicuna-13B, Llama2-70B-Chat, and GPT-3.5-Turbo-16k.The passage identifies GPT-3.5-Turbo-16k as ChatGPT.Source: §4 · p. 5

Prompting Schemes

  • Four prompting schemes cross zero-shot or few-shot examples with CoT or no CoT, and results are averaged across three prompts.Few-shot prompts use two randomly selected examples, one ambiguous and one unambiguous.Source: §4 · p. 5

5 Task 1: Identifying Ambiguity

The task evaluates LLMs’ ambiguity identification overall and by category using Accuracy and F1. Current models struggle despite prompting enhancements; ChatGPT leads overall, but performance varies across ambiguity categories.

5 Task 1: Identifying Ambiguity

  • The task measures ambiguity identification both overall and by category, using Accuracy and F1 score.Source: §5 · p. 5

5.1 Overall Evaluation

  • 54.25% accuracy and 52.77% F1 are ChatGPT’s overall scores, making it the strongest model while leaving considerable room for improvement.The evaluation reports results under varying prompting schemes.Source: Table 3 · p. 6, §5.1 · p. 6
  • In zero-shot without CoT, small-scale models struggle to distinguish ambiguous from unambiguous queries and tend to label most queries ambiguous.The paper reports a 50.37 accuracy and 34.27 F1 score for Llama2-70B in this setting.Source: Figure 1 · p. 6, §5.1 · p. 6
  • CoT and few-shot prompting do not consistently improve identification and may leave small models overconfident and less accurate.ChatGPT’s few-shot gains are minimal and often below zero-shot; 12 shots are required to outperform zero-shot, increasing input length.Source: §5.1 · pp. 6–7, Table 4 · p. 6, Table 5 · p. 6

5.2 Fine-Grained Evaluation

  • Under few-shot without CoT, Table 6 reports ChatGPT performs excellently across all Aleatoric Output categories, but does not effectively address semantic and contradiction categories in this evaluation.Under few-shot with CoT, ChatGPT attains average increases of 5% in accuracy and 8% in F1 score across Aleatoric Output categories compared with small-scale LLMs.Source: Table 6 · p. 7, §5.2 · p. 7
  • On contradiction queries, ChatGPT has limited accuracy and low F1; 81.97% of its errors are false negatives.The authors suggest its training approach may contribute, as it encourages responses even when queries contain contradictions.Source: §5.2 · pp. 7–8, Table 6 · p. 7

6 Task 2: Asking Clarifying Questions

The section evaluates LLMs’ ability to ask clarifying questions: ChatGPT outperforms smaller models, but analyses find persistent failures tied to recognizing knowledge boundaries.

6.1 Overall Evaluation

  • ChatGPT’s clarifying-question performance improved by 10.29 on average over Vicuna-13B, the top-performing small-scale LLM.The authors interpret this as stronger performance in generating natural and useful clarifying questions.Source: §6.1 · p. 8
  • Evaluation uses BERT semantic similarity and human Help scores to assess whether generated questions match references and resolve ambiguity.Help assigns each generated question a binary score of 0 or 1.Source: §6.1 · p. 8

6.2 Fine-Grained Evaluation

  • The error analysis examines 400 failed questions: 50 from each category, classified as Wrong Aspect, Under-specified, Over-specified, or Generation error.These cases come from ChatGPT + Few-shot w/ CoT, identified as the most effective model.Source: §6.2 · p. 8
  • Inability to recognize knowledge gaps is the main reason for ineffective clarifying questions; error patterns vary across ambiguity categories.Epistemic Misalignment and Linguistic Ambiguity errors concentrate on under-specified and over-specified questions, while Aleatoric Output shows an average Wrong Aspect error rate of 52.25%.Source: §6.2 · p. 8, Figure 3 · p. 9

7 Conclusion

The conclusion presents CLAMBER as a taxonomy-based benchmark of ambiguous-query identification and clarification. Results show persistent difficulty for current LLMs, while future work will expand the dataset.

7 Conclusion

  • CLAMBER is a benchmark with approximately 12K high-quality examples spanning ambiguity categories, used to assess off-the-shelf LLMs.The paper frames the benchmark as evaluating model strengths, weaknesses, and potential risks.Source: §7 · p. 9
  • Current LLMs still struggle to identify and clarify ambiguous queries, limiting their practical utility in advanced information search applications.Source: §7 · p. 9
  • The authors plan to add more challenging and comprehensive datasets to CLAMBER, guided by its taxonomy.Source: §7 · p. 9

Limitations

The paper identifies prompt sensitivity and limited model coverage as limitations, then documents its prompt design and implementation settings.

Limitations

  • Evaluation results may depend on prompt choice; the authors average three prompts but cannot establish that these are optimal for the task.Prompt sensitivity and prompt optimality are identified as areas for future research.Source: Limitations · p. 9

B Implementation Details

  • The study evaluates only five LLMs because of computational constraints, leaving other models unevaluated.ChatGPT uses temperature 0; open-source models use temperature 0.5 and top-p 0.8, with a 128-token generation limit.Source: Limitations · p. 9, B Implementation Details · p. 11

A Prompt Design

  • Few-shot prompts use two randomly selected CLAMBER examples, while human annotators wrote the chain-of-thought demonstrations as ambiguity analyses.Source: A Prompt Design · p. 11

C Details of Data Collection

The data-collection appendix describes how several source datasets were adapted or used to build CLAMBER’s ambiguity categories and clarifying-question examples.

C.1 ALCUNA Dataset

  • For ALCUNA, new-entity questions are treated as ambiguous and existing-entity questions as unambiguous; GPT-4 generates clarifications for 1,500 sampled ambiguous queries.Source: §C.1 · pp. 11–12

C.2 AmbiTask Dataset

  • AmbiTask contributes 1,200 instances from three classification tasks; rule-based templates create ambiguous queries and clarifications, while manual instruction edits create unambiguous queries.Source: §C.2 · p. 12

C.3 AmbER Dataset

  • AmbER supplies the 500 most frequent non-human entities; GPT-4 generates ambiguous queries, clarifications, and clear versions using entity meanings and related factual questions.Source: §C.3 · p. 12

C.4 AmbiPun Dataset

  • AmbiPun contributes 500 pun-word instances with multiple meanings, following the AmbER data-collection process.Source: §C.4 · p. 12

C.5 AmbiCoref Dataset

  • AmbiCoref’s ambiguous and unambiguous referents are used as queries, with a template generating clarifying questions for ambiguous cases.Source: §C.5 · p. 12

C.6 AmbigQA Dataset

  • AmbigQA questions with multiple answers are labeled ambiguous and single-answer questions unambiguous; annotated clarifications are adopted and ambiguous items manually assigned to four categories.Source: §C.6 · p. 12

C.7 Dolly Dataset

  • For Dolly open-QA instructions, GPT-4 distinguishes ambiguous from unambiguous queries, generates clarifying questions for ambiguous ones, and assigns predefined categories.Source: §C.7 · p. 12

D Human Evaluation Details

Annotators assess whether generated clarifying questions follow the required format and resolve query ambiguity, with reliability and error categories also reported.

D Human Evaluation Details

E More Task Results

Additional prompting-setting results retain the paper’s earlier overall conclusions, while human validation reports quality checks and the appendix supplies prompts, templates, and dataset examples.

E More Task Results

  • Across three prompting settings, LLM performance varies in exact values, but the overall results and analysis conclusions remain largely consistent with Section 5.2.Source: E More Task Results · p. 13

F Human Validation and Revision

  • 92.4% average label accuracy and a 73.2 average BLEU score were reported for 400 samples reviewed by the authors; five experts passed qualification to validate and revise the data.Source: F Human Validation and Revision · p. 13

ALCUNA dataset; AmbiCoref dataset

AmbiTask dataset

  • AmbiTask materials include a prompt, clarifying-question template, and data examples.Source: Table 16 · p. 18

AmbER dataset

AmbiPun dataset

AmbigQA dataset

Dolly dataset

Related papers

Loading 2405.12063v2…