Learning to Ask: Conversational Product Search via Representation Learning

Jie Zou, Jimmy Xiangji Huang, Zhaochun Ren, Evangelos Kanoulas

Abstract

Traditional e-commerce search engines often struggle with the "vocabulary mismatch" problem and fail to capture the evolving nuances of user preferences. This paper introduces ConvPS (Conversational Product Search), a unified generative framework that treats search as a joint representation learning task. By embedding users, queries, items, and dialogue history into a shared semantic space and employing sophisticated "learning to ask" strategies, ConvPS effectively bridges the gap between user intent and product descriptions. Experimental results on Amazon datasets demonstrate that ConvPS significantly outperforms state-of-the-art baselines, achieving up to a 100% improvement in Mean Average Precision (MAP).

1. Introduction

Searching for products in massive online catalogs is frequently a frustrating experience for consumers. Users often lack the technical vocabulary to describe exactly what they need, leading to a disconnect between their search queries and the keywords used in product descriptions. While conversational search—where a system asks clarifying questions—offers a potential solution, existing approaches often suffer from three major limitations: they model conversation history independently of long-term user preferences, they rely on rigid keyword matching, and they often ask redundant or unhelpful questions.

To address these challenges, the authors propose ConvPS, which moves beyond simple keyword matching toward a deep semantic understanding of the search process. ConvPS integrates the entire conversational flow into a single latent space, allowing for "soft-matching" and more efficient information gathering.

2. The ConvPS Framework

The core innovation of ConvPS is its unified generative framework. Rather than treating the search engine and the chatbot as separate modules, the system learns a single semantic representation where all entities—users, queries, items, and conversation "slots" (e.g., brand, color, size)—coexist as vectors.

The overall architecture of the ConvPS framework, illustrating the four main stages: Question Pool Construction, Representation Learning, Learning to Ask, and Item Ranking.

As shown in the architecture above, the framework operates in four distinct stages:

  1. Question Pool Construction: Extracting slot-value pairs from a corpus of product reviews to create a library of potential clarifying questions.
  2. Representation Learning: Learning embeddings for users ($u$), words ($w$), queries ($Q$), slots/values ($C_{qa}$), and items ($v$).
  3. Learning to Ask: Selecting the most informative question to present to the user based on the current state of the search.
  4. Item Ranking: Calculating the probability $P(i)$ of an item being the target based on the combined vector of the user, query, and dialogue feedback.

Soft-Matching and Negative Feedback

By utilizing embeddings, ConvPS performs "soft-matching." This allows the system to understand that a user describing a product as "outstanding" likely refers to an item described as "excellent" in the metadata. Furthermore, the model explicitly learns from negative feedback (e.g., "I don't want a Sony"). This negative signal is treated as a powerful constraint that helps the system prune the search space more effectively than positive feedback alone.

3. Strategic Questioning

A critical component of conversational search is deciding what to ask. ConvPS evaluates four different strategies to optimize this process:

  • Generalized Binary Search (GBS): A greedy approach that attempts to select the question that most evenly splits the remaining pool of candidate items.
  • LinRel and Gaussian Processes (GP): These are advanced "explore-exploit" strategies. By using techniques like Expected Improvement (EI) and Upper Confidence Bound (UCB), the system balances asking questions to learn more about the user's general preferences (exploration) with asking questions to confirm a specific product match (exploitation).

4. Experimental Evaluation

The researchers evaluated ConvPS using three categories from the Amazon Product Datasets: Movies & TV, Health & Personal Care, and Cell Phones & Accessories. Since natural conversational search datasets are scarce, the team simulated user feedback using real product reviews and validated these results with an online user study involving 26 crowd workers.

Key Results

  • Accuracy: ConvPS significantly outperformed state-of-the-art baselines like HEM and AVLEM. In the "Cell Phones" category, it achieved a MAP of 0.235, nearly doubling the 0.125 MAP of the best baseline.
  • Efficiency: The model demonstrated rapid convergence. In several test cases, the target item's rank improved from over 2,000 to the top 3 within just ten rounds of questioning.
  • User Experience: In the live study, 71.4% of users found the system's questions easy to answer, and the system correctly interpreted user feedback 77% of the time.

5. Conclusion and Future Work

ConvPS represents a significant step toward more intuitive and effective product search. By unifying representation learning with strategic questioning, the system overcomes the limitations of traditional keyword-based search.

Future research will focus on moving beyond template-based questions toward Natural Language Generation (NLG) to create more human-like interactions. Additionally, the authors aim to improve the system's handling of user uncertainty (e.g., "I'm not sure") and address the "cold start" problem for new users who lack a historical purchase record.

Summary updated

BibTeX
@article{zou2023learning,
  author = {Jie Zou and Jimmy Xiangji Huang and Zhaochun Ren and Evangelos Kanoulas},
  title = {{Learning to Ask: Conversational Product Search via Representation Learning}},
  journal = {ACM Trans. Information Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=Dtmm7IYpqg},
}
Figure