Source-linked AI summary
Clarify When Necessary: Resolving Ambiguity Through Interaction with LMs
Michael J. Q. Zhang, Eunsol Choi
TL;DR
Ambiguity can leave a speaker’s intent unclear, making clarification an important part of interaction. The paper proposes a three-task framework and an intent-entropy method for choosing when to clarify; INTENT-SIM improves identification of predictions that can benefit from clarification across tasks and language models.
Problem
The paper addresses how assistants can resolve ambiguous requests through interaction, including estimating whether clarification will improve performance rather than only whether a prediction is uncertain.
Source: §1 · p. 1, §2.1 · p. 3Method
Results
INTENT-SIM consistently outperforms uncertainty-estimation baselines at identifying predictions that can improve with clarification, with improvements across tasks and language models.
Source: §1 · p. 2Takeaways & Limitations
The framework exposes challenges in modeling clarifying interactions and motivates further study of separating uncertainty estimation from identifying uncertainty attributable to ambiguity.
Source: §8 · p. 9With only 10 samples, simulated-answer entropy can be coarse-grained and may underperform at large interaction budgets, which are less practical because fewer than half of examples benefit from clarification in both datasets.
Source: §6.3 · p. 9
Abstract
from arXiv · showhide
Resolving ambiguities through interaction is a hallmark of natural language, and modeling this behavior is a core challenge in crafting AI assistants. In this work, we study such behavior in LMs by proposing a task-agnostic framework for resolving ambiguity by asking users clarifying questions. Our framework breaks down this objective into three subtasks: (1) determining when clarification is needed, (2) determining what clarifying question to ask, and (3) responding accurately with the new information gathered through clarification. We evaluate systems across three NLP applications: question answering, machine translation and natural language inference. For the first subtask, we present a novel uncertainty estimation approach, intent-sim, that determines the utility of querying for clarification by estimating the entropy over user intents. Our method consistently outperforms existing uncertainty estimation approaches at identifying predictions that will benefit from clarification. When only allowed to ask for clarification on 10% of examples, our system is able to double the performance gains over randomly selecting examples to clarify. Furthermore, we find that intent-sim is robust, demonstrating improvements across a wide range of NLP tasks and LMs. Together, our work lays foundation for studying clarifying interactions with LMs.
1 Introduction
The paper frames ambiguity resolution as a three-stage interaction: decide when to clarify, ask a question, then use the answer to respond. It introduces INTENT-SIM to estimate when clarification will help, evaluating the framework across question answering, machine translation, and natural language inference.
1 Introduction
- The framework separates ambiguity resolution into deciding when to clarify, generating a clarifying question, and responding using the user’s answer.The clarifying QA pair corresponds to the gold interpretation and is generated from existing disambiguated input/output pairs.Source: §1 · p. 1, Figure 1 · p. 2
- The evaluation spans question answering, machine translation, and natural language inference, using multiple annotations to approximate distributions over user intents.The framework’s first task must identify cases that benefit from clarification while avoiding unnecessary questions when one interpretation dominates.Source: §1 · pp. 1–2
- The paper develops an oracle method for producing clarification questions and answers for different intents to evaluate the other subtasks.The authors note that performance on those tasks depends heavily on the quality of clarifying interactions.Source: §1 · p. 2
- INTENT-SIM estimates entropy over user intents and consistently improves identification of predictions that are incorrect but can benefit from clarification.Its improvements hold across tasks and language models, and it is best in four of six evaluated LLM-task settings for identifying when clarification improves expected performance.Source: §1 · p. 2
2 A Framework for Resolving Ambiguity through Interaction
The framework models ambiguity resolution as a sequence of deciding whether to clarify, generating a clarifying interaction, and responding to it. Its first task estimates expected benefit from clarification, while oracle-generated interactions support evaluation of the other tasks.
2 A Framework for Resolving Ambiguity through Interaction
- The framework represents an ambiguous request with feasible outputs and a user-intended gold output, and assumes a one-to-one matching between clarifying answers and feasible responses.Because the true intent distribution is difficult to annotate and can vary by time, place, and preference, examples sample request-output pairs from that distribution.Source: §2 · pp. 2–3
2.1 Task 1: Determining when Clarification is Necessary
- Systems estimate how much performance will improve with clarification, distinguishing ambiguity-driven aleatoric uncertainty from epistemic uncertainty due to missing knowledge.The evaluation selects the top b% of examples under an interaction budget and adapts AUROC to identify examples whose performance improves with clarification.Source: §2.1 · p. 3
2.2 Task 2: Generating Clarifying Questions and Answers
- Rather than proposing a new question-generation method, the paper uses an oracle to generate clarifying questions and answers for different user intents.The oracle provides a stable test bed because the quality of clarifying interactions affects both deciding when to ask and responding accurately.Source: §2.2 · pp. 3–4
2.3 Task 3: Responding to Clarifications
- The final task evaluates whether a system’s response, conditioned on the ambiguous input and clarifying QA pair, matches the gold output.The paper tests both sampled interpretations and a uniform setting that weights every interpretation equally to avoid overemphasizing common intents.Source: §2.3 · p. 4
3 Datasets and Applications
The paper applies its ambiguity-resolution framework to datasets for question answering, natural language inference, and machine translation. These datasets provide ambiguous inputs, alternate interpretations, and task-specific outputs or labels for evaluating clarification.
3 Datasets and Applications
- The applications use datasets with ambiguous inputs and multiple interpretations, paired with disambiguated rewrites or contexts and their corresponding outputs.The annotations also support generating oracle clarifying questions and answers; the datasets do not already label distributions over user intents.Source: §3 · p. 4
3.1 Question Answering
- For question answering, AmbigQA pairs ambiguous questions with disambiguated revisions and answers, while NaturalQuestions answers are sampled as intended outputs.Performance is measured by answer recall: whether the normalized generated output contains the gold answer string.Source: §3.1 · p. 4
3.2 Natural Language Inference
- For natural language inference, AmbiEnt provides ambiguous premise-hypothesis pairs and disambiguations, and the matching interpretation supplies the sampled intent and output label.Systems are evaluated with standard three-way entailment, contradiction, and neutral classification accuracy.Source: §3.2 · p. 4
3.3 Machine Translation
- For machine translation, DiscourseMT pairs ambiguous English-French test sentences with two possible context sentences and translations that depend on which context precedes.The evaluation uses contrastive accuracy, checking whether the intended translation receives greater likelihood than the alternative.Source: §3.3 · pp. 4–5
4 An Oracle for Generating Clarifying Questions
The oracle uses few-shot GPT-3.5 prompts and known interpretations to generate clarifying questions and intent-specific answers. Analysis of clarifying questions identifies recurring ambiguity causes that vary across tasks.
4 An Oracle for Generating Clarifying Questions
- The oracle generates a question that differentiates interpretations and a corresponding clarifying answer for each interpretation, using the ambiguous input and its feasible outputs.It is an oracle setting because it requires access to the input's feasible interpretations.Source: §4 · p. 5
- Word-sense ambiguity is most common across tasks; entity linking also surfaces in QA, while literal-versus-implied readings arise in QA and figurative-language interpretation in NLI.In MT, source-sentence ambiguities can usually be captured in the translation.Source: §4 · pp. 5–6
- Table 2 compares clarifying QA pairs, disambiguated inputs, and direct inputs; QA and NLI include uniform and sampled interpretations, with task-specific metrics.MT uses contrastive accuracy, QA uses EM accuracy, and NLI uses 3-way classification accuracy.Source: Table 2 · p. 6
- The analysis also identifies multiple valid outputs as an ambiguity type specific to QA, where reporting only one answer may mislead users.The paper does not find this type in MT or NLI.Source: §4 · p. 6
5 Experiments: Responsiveness to Clarification
Across tasks and systems, LLMs generally improve when given clarifying questions and answers, though LLaMA-2 variants are an exception on MT. Clarifications often outperform disambiguated inputs on QA and NLI, while chat fine-tuning does not consistently improve clarification use.
5 Experiments: Responsiveness to Clarification
- Across tasks and systems, LLMs generally improve with clarifying questions and answers, except LLaMA-2 variants on MT, which the authors attribute to insufficient multilingual pretraining.The reported comparison is in Table 2; the supplied passage gives no cell values.Source: §5 · p. 6
- Clarifying interactions tend to outperform disambiguated inputs, particularly on QA and NLI, whose minimal-edit rewrites are less natural than the interactions.The authors note that unconstrained clarifying interactions more closely resemble the systems' pretraining distributions.Source: §5 · p. 6
- Chat fine-tuning does not consistently improve LLaMA-2 variants' ability to use clarifications, and clarification gains remain consistent between equal-size systems.The authors conclude that existing chat fine-tuning does not adequately train this ability.Source: §5 · p. 6
6 Experiments: Determining When to Clarify
The section evaluates when systems should ask for clarification, comparing uncertainty estimates and prompting baselines. INTENT-SIM estimates ambiguity from simulated user intents rather than uncertainty over model outputs.
6.1 INTENT-SIM
- INTENT-SIM estimates ambiguity as entropy over simulated user intents, distinguishing it from uncertainty over model outputs.It samples responses to a clarifying question, groups semantically equivalent answers, and estimates entropy over the resulting intent distribution.Source: §6.1 · p. 7, Figure 2 · p. 7
6.2 Baselines
- Likelihood uses answer confidence, Self-Ask scores the likelihood of generating “No” to a follow-up question, and semantic entropy measures uncertainty over sampled outputs.For MT and NLI, semantic entropy groups sampled outputs using exact match.Source: §6.2 · pp. 7–8
6.3 Results
- Semantic entropy and simulated user interactions consistently outperform all baselines and random selection across interaction budgets.Table 4 reports AUROC and performance under fixed interaction budgets for MT, QA, and NLI.Source: Table 4 · p. 8, §6.3 · pp. 8–9
- Likelihood and Self-Ask generally outperform random selection but perform considerably worse under some budget settings.The authors attribute weaker AUROC gains in some cases to coarse entropy estimates from only 10 samples; large interaction budgets may also underperform other baselines.Source: §6.3 · p. 9
7 Related Work
Prior work studies clarification questions, uncertainty estimation, and ambiguity benchmarks in task-specific settings. The paper positions its focus on modeling ambiguity and resolving it through interaction across tasks.
Clarifying Questions
- Clarification-question research includes product FAQ reranking (Rao and Daumé III, 2018), supervised question generation (Shridhar et al., 2023), and RL-guided questions (Pyatkin et al., 2022).Yu et al. (2019) balance asking questions with classification, but their questions cover existing attributes rather than being open-ended.Source: §7 · p. 9
Uncertainty Estimation
- Uncertainty work addresses out-of-distribution prediction and semantic equivalence, while Cole et al. (2023) do not model ambiguity degree or resolve it through interaction.Other methods do not explicitly disentangle uncertainty sources, and supervised estimates generalize poorly to new domains.Source: §7 · p. 9
Ambiguity in NLP
- Ambiguity datasets span coreference, NLI, and MT; prior interactive MT work does not model input ambiguity or determine whether interaction helps.Ambiguous QA benchmarks have also been used to study reliance on harmful social biases (Parrish et al., 2021).Source: §7 · p. 9
8 Conclusion
The paper presents a unified framework for resolving ambiguity through clarifying questions across QA, MT, and NLI, alongside INTENT-SIM for deciding when clarification helps. It positions the framework as a basis for further research on interactive ambiguity resolution.
8 Conclusion
- The framework applies clarifying-question ambiguity resolution to QA, MT, and NLI, and motivates further study of when uncertainty reflects ambiguity.INTENT-SIM improves detection of when to clarify.Source: §8 · p. 9
- Future work could use the framework for task-agnostic question generation, multi-turn interactions, and learning objectives that encourage pragmatic clarification.The authors also propose examining what LLMs learn through chat fine-tuning.Source: §8 · p. 9
A Data Details
The data details describe filtering examples when matching NaturalQuestions answers to AmbigQA interpretations.
A Data Details
- Examples were excluded when a NaturalQuestions answer matched no AmbigQA interpretation or matched more than one.Source: A Data Details · p. 12
B Modeling Details
The oracle clarification-generation system uses prompts to generate follow-up questions. The MT prompt carries a potential answer-leakage risk, but generated questions and answers remain in the source language.
B.1 Oracle Question Generation System
- The oracle system generates clarification questions using prompts for QA, NLI, and MT.The prompts are presented in Tables 7, 8, and 9, respectively.Source: §B.1 · p. 12
- Although the MT oracle prompt includes the output translation and risks answer leakage, its generated follow-up questions and answers are always in the source language.Source: §B.1 · p. 12
B.2 Experimental Setups and Additional Results
The experimental setup uses task-specific prompts, prepares disambiguated inputs for QA, NLI, and MT, and selects low-variation examples as unambiguous for NLI.
B.2 Experimental Setups and Additional Results
- The prompts vary by task, and the NLI instruction and class-to-token mapping are based on Liu et al. (2023).The prompts cover clarification, no clarification, and SelfAsk.Source: §B.2 · p. 12
- For disambiguated QA and NLI, the setup substitutes the input with its disambiguated form; for MT, it prepends the disambiguation as context.Source: §B.2 · p. 12
- For NLI SelfAsk, examples interpreted identically by all nine annotators serve as unambiguous cases because they show the least variation in user intents.MT and QA use dataset-labeled unambiguous examples.Source: §B.2 · p. 12
- Table 5 presents raw ambiguous examples from QA, MT, and NLI datasets.Source: Table 5 · p. 13
- Table 6 reports task-level counts of ambiguous and unambiguous inputs, sampled interpretations for clarification evaluations, and total interpretations.Source: Table 6 · p. 13