User Willingness-Aware Sales Talk Dataset

• The paper introduces a meticulously annotated dataset distinguishing three dimensions of user willingness (CD, PI, GA) in sales interactions.
• It employs ecological validity with staged, realistic sales dialogues and utterance-level ratings to capture dynamic user behaviors.
• Strategic conditioning of dialogue agents using these willingness signals demonstrably improves engagement and final purchase intent.

User willingness-aware sales talk datasets are specialized corpora designed to capture, annotate, and model the multidimensional notion of user willingness as it manifests during sales-focused human–computer dialogue. These datasets serve as the foundation for developing automated systems and analytics tools capable of tailoring interaction strategies to a user's fluctuating likelihood to engage, disclose information, or accept sales initiatives. The construction and utilization of such datasets require methodological rigor, fine-grained annotation protocols, and an ecologically valid experimental setting, with recent advances encompassing utterance-level willingness labeling, contextual profiling, and strategic conditioning for modern dialogue agents (Hentona et al., 27 Dec 2024).

1. Definition of User Willingness and Its Dimensions

User willingness, as formalized in the applied sales dialogue literature (Hentona et al., 27 Dec 2024), denotes a set of user behavioral and attitudinal indicators reflecting their propensity to engage further in conversation, share relevant information, and ultimately accept the persuasive objective of the interaction. The foundational dataset distinguishes three willingness dimensions:

Willingness Aspect	Behavioral Manifestation	Typical Metric
Continuing Dialogue (CD)	Propensity to remain in the conversation	Engagement ratio
Providing Information (PI)	Readiness to disclose purchase-relevant personal information	Disclosure rate
Goal Acceptance (GA)	Openness to move toward purchase intent	Acceptance score

During data collection, participants evaluate each sales-system utterance along these dimensions using a discrete scale (positive, neutral, negative), enabling high temporal resolution and contextual sensitivity. This annotation schema facilitates longitudinal modeling of willingness fluctuations and supports dynamic strategy adaptation by conversational agents.

2. Ecological Validity and Data Collection Protocols

To approximate realistic sales scenarios, dataset creators employ principles from ecological validity in HCI (Hentona et al., 27 Dec 2024). Sales conversations are staged in environments resembling authentic e-commerce interfaces or advertisement landing pages, preserving user agency (e.g., enabling disengagement at any turn and avoiding forced purchases).

The Wizard-of-Oz paradigm is commonly adopted: expert salespersons interact with real users recruited via crowdsourcing platforms. After each sales utterance, users provide per-utterance willingness ratings, thus generating annotated multi-turn dialogues with fine-grained willingness trajectories.

The protocol also includes measurement of pre-conversation purchase intent (via Likert scale), tracking user attrition, and recording underlying meta-data (number of turns, tokens, scenario context). This approach ensures data utility for both behavior analysis and supervised machine learning.

3. Annotation Methodology and Scoring Strategies

User annotations are mapped to numerical scores (positive: +1, neutral: 0, negative: -1), yielding temporally structured series for each dialogue (Hentona et al., 27 Dec 2024). Aggregated and turn-level statistics (means, counts of negative evaluations) are used to analyze the progression of willingness within dialogues:

• Early dialogue stages often emphasize sustaining CD willingness.
• Middle turns prioritize eliciting PI willingness, focusing on information gathering.
• Final dialogue phases target GA willingness, driving toward user acceptance of the sales goal.

Statistical visualizations (e.g., violin plots, turn-wise trend graphs) provide diagnostics on sales strategies and their differential impact across willingness dimensions. Notably, any negative annotation is shown to be highly detrimental to overall purchase intent.

4. Integration into Conversational AI Training and System Design

Constructed datasets are utilized for attribute-conditioned supervised fine-tuning of large pre-trained models. For example, fine-tuning GPT-3.5 on utterance-level willingness labels, combined with strategic conditioning (e.g., prioritizing CD in initial turns, PI mid-dialogue, GA near closing), yields dialogue agents that dynamically adjust sales tactics as user willingness evolves (Hentona et al., 27 Dec 2024).

Empirical evaluation demonstrates that leveraging willingness-aware conditioning improves system-generated outcomes, specifically increasing final purchase intent compared to baseline models ignorant of willingness signals.

Furthermore, strategic systems may incorporate scoring models to avoid negative utterances and maximize engagement through the distinct dialogue stages, effectively operationalizing findings from dataset trend analysis.

5. Contextual Profiling, Personalization, and Multidimensional Data Sources

Recent expansions of willingness-aware datasets incorporate contextual user profiling derived from large-scale behavioral data such as purchase history, review metadata, and personality traits (MBTI dimensions) (Kim et al., 28 Mar 2025, Cheng et al., 25 Apr 2025). User simulators (e.g., CSUSER) grounded in empirical data can model fine-grained user profiles (preferences, budget constraints, decision-making styles, openness) to generate realistic simulated sales interactions.

Metrics such as Success Rate (SR) and Sales-Win Rate (SWR) are introduced to quantify both general recommendation success and the agent's ability to persuade users beyond their stated price boundaries:

$$\text{SWR} = \frac{\sum_n I(a_n \in I_n^{(\text{out})})}{\sum_n I(a_n \in (I_n^{(\text{in})} \cup I_n^{(\text{out})}))}$$

where $a_n$ is the user's accepted item, and $I(\cdot)$ is an indicator function (Kim et al., 28 Mar 2025).

Integration of persona-defined simulators allows controlled benchmarking and provides actionable interfaces for adaptive agent training across conversational domains.

6. Analytical and Practical Insights for Sales Dialogue Systems

Analysis of willingness-aware datasets yields several strategic insights:

• Avoidance of utterances tagged with negative willingness is crucial for increasing conversion.
• Effective sales strategies require dynamic adaptation: emphasizing engagement first, information gathering mid-dialogue, and persuasion or closure in final turns.
• Willingness dynamics vary by product category and user profile, necessitating personalized conversational trajectories and diversified persuasive tactics.
• Incorporation of real-time willingness feedback into agent policy selection (e.g., via chain-of-thought reasoning or prompt-based conditioning) improves system consistency, naturalness, and overall effectiveness (Hentona et al., 27 Dec 2024, Kim et al., 28 Mar 2025, Cheng et al., 25 Apr 2025).

Such datasets not only advance the design of conversational recommender systems but also facilitate research in adaptive strategy selection, intent detection, and user-centered dialogue evaluation.

7. Significance, Limitations, and Future Directions

User willingness-aware sales talk datasets represent a rigorous methodological advancement over previous corpora that lacked explicit user intent tracking or ecological validity. Their multi-dimensional annotation schema, contextual profiling, and strategic conditioning frameworks provide critical infrastructure for developing high-performance, interactive sales agents and for studying user engagement, persuasion, and decision-making.

Limitations include the complexity of eliciting robust user willingness annotations at scale and challenges associated with generalizing behavioral findings across domains and cultures. Future research directions, as outlined in the literature, include the development of larger, cross-category datasets, advanced fusion models (e.g., CRS-LLM hybrids), and dynamic adaptation algorithms combining willingness signals with real-time conversational feedback (Hentona et al., 27 Dec 2024, Kim et al., 28 Mar 2025, Liu et al., 2023).

These datasets form the empirical and methodological backbone for a new generation of user-willingness-aware, context-sensitive conversational agents in sales and related domains.