Subjective Attributes in Conversational Recommendation Systems: Challenges and Opportunities

Filip Radlinski, Craig Boutilier, Deepak Ramachandran, Ivan VendrovAAAI 2022 · OpenReview

Abstract

The ubiquity of recommender systems has increased the need for higher-bandwidth, natural and efficient communication with users. This need is increasingly filled by recommenders that support natural language interaction, often conversationally. Given the inherent semantic subjectivity present in natural language, we argue that modeling subjective attributes in recommenders is a critical, yet understudied, avenue of AI research. We propose a novel framework for understanding different forms of subjectivity, examine various recommender tasks that will benefit from a systematic treatment of subjective attributes, and outline a number of research challenges.

Introduction

Modern recommender systems (RSs) typically treat item attributes as objective, catalog-based truths. For instance, a search for a "cheap" camera often triggers a hard-coded price filter. However, in conversational settings, users frequently employ "soft" attributes that are deeply personal. A "cheap" camera for a professional photographer differs significantly from one for a student. This discrepancy highlights a "grounding problem": AI models often assume a unique, universal mapping from a word to its meaning. When this assumption fails, the resulting interaction feels rigid and fails to meet user expectations. This paper argues that subjective disagreement should not be treated as "noise" but as a fundamental feature of human language that must be systematically modeled.

A Taxonomy of Subjectivity

The core contribution of this work is a conceptual framework categorizing subjectivity into three distinct types:

  1. Degree Subjectivity: Users agree on the underlying scale but differ in their thresholds. For example, while everyone agrees a lower price is "cheaper," the specific dollar amount where "cheap" begins varies by individual.
  2. Semantic Subjectivity: Users have fundamentally different meanings or internal orderings for a term. One user might define a "funny" movie as dark political satire, while another associates it with slapstick comedy.
  3. Compositional Subjectivity: An attribute serves as a shorthand for a complex combination of other features. A "safe" car might imply high crash-test ratings to one user, but all-wheel drive and superior braking systems to another.
Three forms of attribute subjectivity: (a) Degree, (b) Semantic, and (c) Compositional.

Modeling and Data Challenges

To represent these subjective concepts, the authors discuss mathematical frameworks such as Ball Semantics (viewing attributes as regions in an embedding space) and Line Semantics (viewing attributes as directions or vectors).

A significant hurdle is the "small-data challenge." Learning a user's personal semantics must often occur within just a few interactions. This necessitates the use of few-shot learning and active learning techniques, where the system might explicitly ask for clarification (e.g., "By 'safe,' do you mean crash-test ratings or handling?"). Furthermore, subjectivity is dynamic and context-dependent; a "posh" restaurant in a small town has a different standard than one in a major metropolitan area.

Existing datasets, such as social tagging data (MovieLens, Goodreads) and natural language corpora (Amazon, Yelp reviews), provide a starting point. However, current benchmarks lack metrics specifically designed to measure a system's ability to handle subjective disagreement, often favoring consensus over personalization.

Conclusion and Future Directions

Addressing subjectivity is essential for the next generation of conversational recommenders. Future research should focus on developing personalized word embeddings that adapt to individual users and creating subjectivity-aware explanations. Instead of a generic "Recommended because it is 'funny'," a system should offer personalized justifications: "Based on your preference for dark humor, you might find this movie 'funny'." By embracing the personal nature of language, AI can move toward more natural and trustworthy user interactions.

Summary updated

Abstract

The ubiquity of recommender systems has increased the need for higher-bandwidth, natural and efficient communication with users. This need is increasingly filled by recommenders that support natural language interaction, often conversationally. Given the inherent semantic subjectivity present in natural language, we argue that modeling subjective attributes in recommenders is a critical, yet understudied, avenue of AI research. We propose a novel framework for understanding different forms of subjectivity, examine various recommender tasks that will benefit from a systematic treatment of subjective attributes, and outline a number of research challenges.
BibTeX
@inproceedings{radlinski2022subjective,
  author = {Filip Radlinski and Craig Boutilier and Deepak Ramachandran and Ivan Vendrov},
  title = {{Subjective Attributes in Conversational Recommendation Systems: Challenges and Opportunities}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  url = {https://openreview.net/forum?id=H8EtwFQPYT},
}
Figure