Personalized LoRA for Human-Centered Text Understanding

You Zhang, Jin Wang, Liang-Chih Yu, Dan Xu, Xuejie ZhangAAAI 2024 · DOI

Introduction: The Challenge of Human-Centered Text Understanding

Human-Centered Text Understanding (HCTU) recognizes that the meaning of a text often depends on its author. In personalized sentiment analysis, for example, a "3-star" review from a harsh critic conveys a different sentiment than the same rating from a lenient user. Adapting Large Language Models (LLMs) for such personalization faces three primary hurdles:

  1. Scalability: Storing unique weights or even standard LoRA adapters for millions of users is computationally prohibitive.
  2. Cold-Start: Models often struggle with new or anonymous users (zero-shot/few-shot scenarios) because they lack historical preference data.
  3. Efficiency: Traditional knowledge injection methods typically require full-model fine-tuning, which is too expensive for large-scale Pre-trained Language Models (PLMs).

Methodology: PLoRA and Plug-and-Play

The authors introduce PLoRA (Personalized Low-Rank Adaptation), a framework that bridges the gap between Personalized Knowledge Injection (PKI) and Parameter-Efficient Fine-Tuning (PEFT).

Overview of PLoRA compared to traditional PKI and PEFT methods, highlighting the Plug-and-Play (PnP) framework for cold-start scenarios.

Shared Low-Rank Structure

Instead of separate adapters, PLoRA utilizes a shared architecture where user attributes are converted into a unified embedding ($p$). This embedding is combined with the text representation ($h$) and passed through a shared output projection matrix. This coupling allows the model to learn a cohesive relationship between text and persona while drastically reducing parameter overhead.

Plug-and-Play (PnP) for Cold-Starts

For new users, PLoRA employs a "Plug-and-Play" approach: the model weights remain frozen, and only the user embedding ($p$) is updated via a quick backpropagation step. To ensure the model remains robust for anonymous users, the authors incorporate:

  • PDropout (Personalized Dropout): Randomly masking user tokens during training to maintain general text-understanding capabilities.
  • MIM (Mutual Information Maximization): A strategy to align task-specific and human-centered representations, acting as a regularizer to improve generalization.

Key Insights

  • Weight Sharing as Integration: By sharing the output matrix between the task-specific LoRA and the user-specific injection, personalization becomes an integral part of the task adaptation rather than an isolated feature.
  • The User as an "Access Token": In the PnP framework, the user embedding acts like a dynamic access token, allowing for seamless deployment where only the token is swapped per user without altering the underlying model.
  • Decoupling for Generalization: Strategies like MIM and PDropout effectively decouple personalization from general understanding, allowing the model to be highly personalized for known users while remaining accurate for strangers.

Experimental Results

PLoRA was evaluated across benchmark datasets including IMDB, YELP, GDRD, and PPR. It consistently outperformed state-of-the-art methods like MA-BERT and UserAdapter while requiring 30x to 50x fewer trainable parameters than full-model fine-tuning.

Performance of PLoRA in few-shot scenarios on IMDB-B and YELP-B datasets, demonstrating rapid adaptation as the number of shot samples increases.

As shown in the results, PLoRA demonstrates robust adaptation in few-shot scenarios, reaching competitive performance with as few as 15 samples. In zero-shot (anonymous) settings, it maintains high accuracy where other personalized models typically degrade.

Limitations and Future Work

While highly effective for sentiment analysis, the authors note that HCTU encompasses many other tasks, such as personalized dialogue or recommendation, which remain to be explored. Future work could also extend PLoRA beyond multi-head attention layers to other components like LayerNorm or hybrid structures, and investigate its utility in broader text-generative tasks.

Summary updated

BibTeX
@inproceedings{zhang2024personalized,
  author = {You Zhang and Jin Wang and Liang-Chih Yu and Dan Xu and Xuejie Zhang},
  title = {{Personalized LoRA for Human-Centered Text Understanding}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i17.29931},
}
Figure