Refocusing on Relevance: Personalization in NLG

Shiran Dudy, Steven Bedrick, Bonnie WebberEMNLP 2021 · DOI

Introduction

In current Natural Language Generation (NLG) research, a fundamental assumption persists: that for any given input, there exists a single, objective "gold standard" output. This position paper, originally presented at EMNLP 2021, challenges this "Universal Truth Assumption." The authors argue that by ignoring the specific context, intent, and identity of the user, modern NLG systems are reaching a ceiling in utility and risk perpetuating harmful data biases. To move forward, the field must transition from generic generation to a paradigm of situational relevance.

The Fallacy of the Universal Truth

Most NLG tasks—including summarization, machine translation, and question answering—are evaluated by comparing model outputs to a single reference text. This approach is fundamentally flawed for three primary reasons:

  1. Diminished Utility: A one-size-fits-all approach ignores the diverse needs of users. For instance, a medical research summary intended for a clinician requires a different level of technical detail and vocabulary than one intended for a patient.
  2. Inherent Ambiguity: Many user prompts are underspecified. A request for "instructions on how to put out a fire" requires vastly different responses depending on whether the user is facing a kitchen grease fire or explaining the concept to a child.
  3. Amplification of Bias: When models are trained to find a single "correct" answer from diverse data, they often default to the most frequent patterns. This frequently results in the reinforcement of societal biases, such as translation systems defaulting to gendered pronouns based on professional stereotypes found in training corpora.

A New Framework: Situational Relevance

To address these flaws, the authors propose borrowing the concept of situational relevance from Information Retrieval (IR). While "topicality" refers to whether an output matches the subject matter, "situational relevance" measures whether the output is useful for the user's specific task and context.

Transitioning to this framework requires a shift in both architecture and ethics. The authors advocate for Value-Sensitive Design (VSD), a proactive framework that identifies stakeholders and balances competing values—such as the trade-off between the benefits of high-quality personalization and the risks to user privacy.

Technical Paths to Personalization

While the paper is a conceptual roadmap, it identifies several promising technical avenues for implementing personalized NLG:

  • Prefix-tuning: This involves using small, learnable "prompts" to steer large, frozen language models toward a specific user’s style, expertise level, or context without retraining the entire model.
  • Distributional Constraints: Systems can be designed to adhere to specific attributes during the generation process, such as maintaining a certain level of formality or avoiding specific jargon.
  • Federated Learning: To mitigate privacy risks, personalized models can be trained on-device. This allows the system to learn from user behavior without sensitive personal data ever leaving the local environment.

Privacy and Data Governance

The "privacy-personalization trade-off" remains a significant hurdle. True relevance requires access to personal data, which increases the risk of data misuse. The authors point to the MIMIC-III (medical records) and N3C (COVID-19 data) datasets as successful examples of "data enclaves." These models demonstrate how the research community can share sensitive data through secure, controlled environments, potentially allowing for "Evaluation-as-a-Service" where models are tested on private data without researchers ever directly accessing the raw information.

The Future of Evaluation

The ultimate conclusion of the paper is a call to reform how we measure success in NLG. The field must move beyond static metrics like BLEU or ROUGE, which only measure similarity to a reference text. Instead, the authors advocate for extrinsic, task-based evaluation. Success should be defined by whether the generated text actually helped a human user achieve a specific goal in a real-world context.

By embracing personalization not just as a user-experience feature, but as a core safety and utility requirement, the NLG community can create systems that are more effective, more inclusive, and more resilient to the pitfalls of the Universal Truth Assumption.

Summary updated

BibTeX
@inproceedings{dudy2021refocusing,
  author = {Shiran Dudy and Steven Bedrick and Bonnie Webber},
  title = {{Refocusing on Relevance: Personalization in NLG}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.emnlp-main.421},
}
Figure