Clarify When Necessary: Resolving Ambiguity Through Interaction with LMs

In this work, we explore the challenges of developing interactive assistants that resolve ambiguity by asking their users clarifying questions. Specifically, we develop a task-agnostic framework for evaluating a system’s ability to determine when to ask for clarification. Determining when to ask for clarification is a challenging task that requires systems to consider the demands of the individual user (i.e., how much they prioritize speed and usability versus carefulness) and the distribution of interpretations for a given request (i.e., whether an ambiguous request has one dominant, inferable interpretation). Using this framework, we evaluate systems for determining when to clarify across three NLP applications: QA, MT, and NLI. Finally, we introduce present a novel uncertainty estimation approach, INTENT-SIM, that determines the utility of asking a clarifying question by estimating the entropy over user intents. Our method consistently outperforms existing uncertainty estimation approaches at identifying predictions that will benefit from clarification. 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 further studies on clarifying interactions with LM assistants.

Introduction. Ambiguity is embedded throughout natural language, and even simple utterances can have multiple interpretations when read in isolation. Ambiguity serves a key, communicative function in language, allowing speakers to omit details by relying on information that is inferable from the extralinguistic context of the conversation (e.g., temporal, social, and physical) (Piantadosi et al., 2012). At times, however, the speaker’s intent is still unclear despite the context. In such cases, further interaction is required to resolve the ambiguity, often by asking and answering clarifying questions.

With the recent progress in large language model (LLM) development, interactive AI assistants (e.g., ChatGPT, Claude, LLaMA-2) have risen to prominence in our daily lives; yet, these systems often fail to interact with users to resolve ambiguities in their requests. We identify that the inability to determine when to ask for clarification is one of core challenges limiting these systems’ ability to ask clarifying questions. To address these challenges, we develop an framework for evaluating a system’s ability to determine when to clarify and apply it to a variety of NLP tasks including question answering (QA), machine translation (MT), and natural language inference (NLI). While most previous works on modeling ambiguity in NLP (particularly in QA and MT) have treated ambiguity as a binary classification task (i.e., does this input have multiple, valid interpretations?), determining when to ask for clarification requires systems to consider a variety of more nuanced factors, like whether an ambiguous utterances have one, dominant interpretation that can be inferred (e.g., “She’s from Boston” typically does not mean “Boston, Georgia”) and how much the individual user prioritizes the LLM’s speed and usability versus its carefulness. Our evaluations capture these considerations by measuring a system’s ability to maximize end-task performance while minimizing the interaction cost from the user. To determine which examples will benefit from clarification, we evaluate this task as the first step in a three stage pipeline for resolving ambiguity, which we depict in Figure 1. Here, systems (1) determine if clarification is necessary before (2) asking a clarifying question and (3) responding given the user’s clarifying answer. We also support these evaluations by deriving datasets of sampled interpretations of ambiguous inputs from existing datasets focused on modeling ambiguity in NLI, MT, and QA (Min et al., 2020; Bawden et al., 2018; Liu et al., 2023).

Related work. Clarifying Questions Fried et al. (2022) notes that a relationship between pragmatic reasoning and ambiguity, where ambiguity can often be resolved via pragmatic reasoning. Prior works explore such goal-oriented dialogues have studied task-specific settings. Shridhar et al. (2023) studies generating clarifying questions as a supervised learning task, and then using them for knowledge distillation. (Pyatkin et al., 2022) uses RL to guide their question generation models for moral judgments of a situations. Prior work (Yu et al., 2019) studies balancing asking clarification questions and making the final classification prediction over multi turn interactions. Their clarifying questions only cover existing attributes, while ours are open ended. Deng et al. (2023b) also notes the tendency for current LLMs to not ask clarifying questions, and explores methods to trigger such responses via prompting. Clarifying questions have also been studied under the task of dialogue act prediction. While clarifying questions represent one class of dialogue acts, Deng et al. explores the task of predicting a variety of other dialogue acts that are relevant to other dialogue domains like negotiations. Deng et al. (2023a) surveys work exploring a wider range of ‘dialogue strategies’ and methods to predict when to use each. Our task of determining whether clarification is needed is closely related to such works on dialogue act prediction. Furthermore, Sorensen et al. (2024) describes the challenge and goal of developing pluralistically aligned language models that can accommodate a variety of user perspectives. Clarifying questions represent one such strategy, and our evaluation methods using sampled user interpretations are connected to the concept of distributional pluralism discussed in this work. Here, systems are evaluated on their ability to accommodate perspectives based on their distribution across the user population. We are able to develop an evaluation setting for this by relying on sampled interpretations to inputs where users may have diverse views and expect opposing outputs.

5548 Uncertainty Estimation Several existing works have studied methods for disentangling different sources of uncertainty. Kamath et al. (2020) studies identifying out-of-distribution test examples, a source of epistemic uncertainty. Other works have studied uncertainty estimation techniques for LLMs (Kadavath et al., 2022; Lin et al., 2022), but they do not explicitly model or evaluate their ability to disentangle different sources of uncertainty. These works also explore supervised methods for uncertainty estimation in LLMs, but find that these methods generalize poorly to new domains.

Method. Setting Here, we provide some basic definitions for our setting and the three-step pipeline used for our evaluation framework depicted in Figure 1. We say that users provide an initial input request, x, to the LLM assistant. Some inputs may be ambiguous, resulting in many feasible output responses for the system to choose from, which we denote as Y = {yi}k 1. One of these outputs, y∗∈Y , represents the gold output corresponding to the user’s intent behind their ambiguous request. To determine the users intent, systems may ask the user a clarifying question, q. The user then responds with the clarifying answer corresponding to their intent, a∗∈A = {ai}k 1. For simplicity, we assume a bipartite matching between the sets of clarifying answers, A, and feasible final responses, Y . Each input, x, has its own distribution over intended interpretations given by P(y = y∗|x). Gathering annotations for the true distribution over intents is intractable and temperamental (i.e., subject to changes over time, location, and individual preferences); however, in Section 3 we describe several methods we use for generating our datasets of (x, y∗) tuples where y∗are realistic samples from the true distribution over intents.

2.1 Task Definition Determining when to ask for clarification is a complex task that builds upon prior works on modeling ambiguity and estimating uncertainty in NLP. Existing approaches for modeling ambiguity in NLP applications has primarily focused on treating ambiguity as a binary label: either an input does or does not posses multiple valid interpretations (Min et al., 2020; Pilault et al., 2023). While it may seem natural to always ask for clarification for ambiguous inputs, the distribution over intended interpretations, given by P(y = y∗|x), is often dominated by a single most-likely interpretation. In such cases, it may be preferable to forego clarification and respond to the user’s request directly. How frequently systems should forego clarification for ambiguous requests depends on the demands of the domain and preferences of the user.

4 Ambiguity Resolution Pipeline To evaluate systems for determining when to clarify, we first construct systems for performing the latter two steps of our three-step pipeline for resolving ambiguity through interaction.

Generating Oracle Clarifications To minimize the dependence on the LLM’s ability to generate high quality clarifying questions, we use an oracle system for generating clarifying questions and answers. Our oracle makes use of few-shot prompting with GPT-3.5 (OpenAI, 2022), providing systems with instructions and two hand-written exemplars to accomplish the following task: Given the ambiguous input, x, and its different interpretations, each corresponding to a different output y ∈{yi}k 1, systems must generate (1) clarifying question differentiating each interpretation, q, and (2) then clarifying responses, {ai}k 1, corresponding to each interpretation. Interpretations are provided in different formats depending on the available annotations in each dataset: we use disambiguated revisions of x for QA and NLI and the different target translations, {yi}k 1, for MT (details in Appendix B). Table 1 includes examples of generated clarification questions generated using this oracle.

Generating LLM Predictions To generate LLM predictions with and without providing clarification, we use standard four-shot prompting. We provide LLMs with demonstrations from the target task where exemplars include or do not include clarifications to match whether the test case does. We sample exemplars per-example and perform greedy decoding (exact prompts in Appendix E).

4.1 Do LLMs Utilize Clarifying Interactions?

5.1 INTENT-SIM Unsupervised methods for uncertainty quantification in LLMs generally rely on estimating entropy over the output distribution, using high entropy to identify erroneous outputs (Kadavath et al., 2022; Kuhn et al., 2023). While these methods perform well at identifying incorrect predictions, they fail to identify why predictions are incorrect.

Discussion. As expected, we also observe that clarification is often not necessary infer the correct interpretation and that models still frequently produce errors, even after clarification. These observations reinforce the challenges we highlighted above when determining when to clarify: systems must be able to identify when ambiguous inputs have a dominant, inferable interpretation, and they must be able to disentangle different forms of uncertainty to distinguish when incorrect predictions can and cannot be resolved through clarification.

Another notable trend is that systems tend to perform better with clarifying questions and answers than with disambiguated inputs, particularly for QA and NLI. We attribute this the way our QA and NLI datasets construct disambiguated interpretations. These datasets create disambiguated revisions of each ambiguous input by applying a minimal set of token-level edits to the initial input. While this makes disambiguations easier to annotate and compare, it comes at the cost of the naturalness of the resulting disambiguations. In contrast, our clarifying interactions do not have the same minimal-edit constraints and more closely resemble these systems’ pretraining distributions.

Conclusion. We present a unified framework for resolving ambiguity with clarifying questions, applying it to QA, MT, and NLI. Our framework exposes the challenges in modeling clarifying interactions, and motivates the further study of disentangling uncertainty estimation and identifying when uncertainty can be attributed to ambiguity. We present a novel uncertainty estimation approach for this objective, INTENT-SIM, which we demonstrate improves detection of when to clarify.

Limitations. There is a computational overhead associated with our proposed INTENT-SIM uncertainty estimation method that comes from running an NLI model over samples from the LLM. In practice, we find that this additional cost is dominated by the need to sample multiple continuations from the LLM, and the overhead from running our NLI model or clustering algorithm are negligible in comparison. While this work makes strides toward modeling ambiguity as a distribution over intents, rather than a binary classification as done in previous works on ambiguity in NLI and MT, we note that these distributions are highly dependent on a variety of extralinguistic factors including time, location, and individual preferences. Further exploration is needed into exploring how this framework may be adapted to also account for such factors in both modeling and evaluation. Finally, our work does not consider several other important challenges, including clarifying question generation, and handling arbitrary length interactions. We hope that our work will aid future efforts exploring these unaddressed challenges.