When a user interacts with an automated system, their commands are rarely perfectly specified. A user might say "I cannot open" to a customer service bot, ask a database for "hiring trends" without specifying a date range, or tell a robot to fetch "something to cut with."
Traditional natural language interfaces treat intent understanding as a single-turn classification problem: the system takes an input, maps it to a single internal representation, and executes it. If the user's input is ambiguous or low-confidence, the system typically defaults to a hard failure or makes an unverified guess. As Yaning Zhang et al. observe, this inability to handle low-confidence scenarios prevents systems from guiding users toward a resolution. When a system guesses incorrectly, it forces users into frustrating, multi-turn repair dialogues to undo the error, as noted by Liliana Ardissono et al..
Interactive clarification reframes this problem. Instead of treating ambiguity as a fatal processing error, modern dialogue systems treat uncertainty as a signal to initiate a structured sub-dialogue. By proactively asking targeted questions, the system can narrow down the user's true intent before executing an action.
Quantifying Ambiguity: Knowing When to Ask
Before a system can clarify an ambiguous request, it must first recognize that it is confused. This requires moving beyond simple confidence scores to explicitly model denotational uncertainty—the state where an utterance could validly map to multiple distinct system actions.
[User Utterance] ---> [Conformal Prediction / Sampling] ---> Multiple Valid Intents?
|
+----------------------------+----------------------------+
| Yes | No
v v
[Interactive Clarification Loop] [Direct Execution]
To achieve this, systems leverage techniques that expose the boundaries of their own knowledge:
- Interpretation-First Prompting: Jeremy R. Cole et al. demonstrate that forcing a model to explicitly state its interpretation of an ambiguous query before answering surfaces hidden assumptions. By using Sampling Repetition—counting how often an answer repeats across multiple random model generations—the system can detect when it is guessing between multiple valid interpretations.
- Conformal Prediction: Rather than relying on uncalibrated softmax probabilities, Floris den Hengst et al. introduce the Conformal Intent Classification and Clarification (CICC) framework. CICC processes a user's query and outputs a set of likely intents that is statistically guaranteed to contain the correct answer at a user-defined confidence level. If the set contains more than one intent, the system knows it must clarify.
- Uncertainty Decomposition: In high-stakes domains like healthcare, Daeyoung Kim et al. apply uncertainty decomposition to electronic health record question-answering. This allows the system to flag ambiguous clinical questions and transition from passive detection to active, interactive intent resolution.
Designing the Clarification Loop: What to Ask
Once a system detects ambiguity, its goal is to resolve it with the least amount of conversational friction. Asking generic questions like "What do you mean?" places the cognitive burden back on the user. Instead, effective systems algorithmically construct questions that maximize information gain.
Collection Partitioning
Xiang Hu et al. frame clarification as a collection partitioning problem. If a user's query matches a cluster of potential intents, the system uses reinforcement learning to generate a diverse set of "refinement labels." These labels act as a mathematical partition over the remaining candidate space, allowing the user to eliminate large swaths of incorrect interpretations in a single turn.
Bayesian Experimental Design
To handle open-ended tasks, Kasia Kobalczyk et al. propose an Active Task Disambiguation framework rooted in Bayesian Experimental Design. The model explicitly reasons about the entire space of possible solutions and computes which clarifying question will most aggressively reduce the entropy of the user's intent distribution.
Downstream Relevance
Crucially, a system should not clarify details that do not affect the outcome. Peter van Beek et al. demonstrate that plan ambiguity only requires resolution if the underlying uncertainty leads to different system actions. If all possible interpretations of a user's vague request map to the exact same execution path, the system can safely proceed without interrupting the user.
Grounding Clarification Across Domains
The transition from static, single-turn execution to interactive disambiguation is fundamentally changing how AI agents operate across diverse environments.
Conversational Databases (Text-to-SQL)
Real-world database users rarely know the exact schema of the data they are querying. Benchmarks like CoSQL, introduced by Tao Yu et al., force models to handle situations where users ask for information missing from the database. Rather than generating a broken SQL query, the system initiates a clarification dialogue to align the user's natural language with the database's structural reality. As Zijin Hong et al. note in their survey, this capability is a prerequisite for moving text-to-SQL systems out of idealized benchmarks and into messy, real-world deployments.
Interactive Data Analysis
Data science is inherently iterative. Hosein Hasanbeig et al. introduce ConDABench to evaluate language models as interactive analysts. They emphasize that real-world data queries are fundamentally under-specified, requiring models to proactively identify missing parameters or data cleaning requirements through dialogue before writing code.
Robotics and Physical Navigation
In robotics, interactive clarification serves as a "virtual sensor" to resolve ambiguities in physical space. When a human gives an unconstrained command to a mobile robot, the robot can use dialogue to fill in missing parameters.
For instance, the ScoutBot system developed by Stephanie M. Lukin et al. prompts users for missing spatial parameters (such as distance or target landmarks) rather than guessing a trajectory. Similarly, the DRAGON robot developed by Shuijing Liu et al. uses interactive clarification to help visually impaired users navigate indoor environments, using free-form dialogue to resolve visual ambiguities and confirm exact destinations safely.
| System Type | Input Ambiguity | Clarification Mechanism | Downstream Goal |
|---|---|---|---|
| Text-to-SQL (Tao Yu et al.) | Vague phrasing; missing schema alignment | Proactive schema-matching questions | Valid SQL execution |
| Robotics (Stephanie M. Lukin et al.) | Underspecified spatial commands | Parameter prompting (distance, targets) | Safe physical navigation |
| Intent Classification (Floris den Hengst et al.) | Overlapping semantic intents | Conformal prediction set selection | Rapid intent resolution |
By treating human-computer interaction as a dynamic negotiation of meaning rather than a static translation task, interactive clarification allows AI systems to operate reliably in complex, real-world environments where ambiguity is guaranteed.
Key Foundational Work
TORUS - A Natural Language Understanding System For Data Management (1975) — contributed 1 concept elements to this direction.
On an Approach for Designing Linguistic Processors (1982) — contributed 1 concept elements to this direction.
Flexible Parsing of Discretely Uttered Sentences (1982) — contributed 1 concept elements to this direction.