TexShape: Information Theoretic Sentence Embedding for Language Models

TL;DR

TexShape tackles the problem of encoding sentences to maximize task-relevant information while minimizing sensitive information using an information-theoretic objective. It introduces a trainable encoder built on top of pretrained sentence embeddings and an MI evaluator that uses the DV representation to compute $\mathcal{I}(T_\Theta(X);X)$, $\mathcal{I}(T_\Theta(X);L)$, and $\mathcal{I}(T_\Theta(X);S)$. The approach demonstrates improved privacy-utility trade-offs and fairness through experiments on SST-2, Corona-NLP, and MultiNLI, achieving lower leakage with comparable downstream accuracy. The framework offers a principled, tunable mechanism for bandwidth-constrained data sharing and bias mitigation, with future work aimed at broader modalities and theory-guided hyperparameter optimization.

Abstract

With the exponential growth in data volume and the emergence of data-intensive applications, particularly in the field of machine learning, concerns related to resource utilization, privacy, and fairness have become paramount. This paper focuses on the textual domain of data and addresses challenges regarding encoding sentences to their optimized representations through the lens of information-theory. In particular, we use empirical estimates of mutual information, using the Donsker-Varadhan definition of Kullback-Leibler divergence. Our approach leverages this estimation to train an information-theoretic sentence embedding, called TexShape, for (task-based) data compression or for filtering out sensitive information, enhancing privacy and fairness. In this study, we employ a benchmark language model for initial text representation, complemented by neural networks for information-theoretic compression and mutual information estimations. Our experiments demonstrate significant advancements in preserving maximal targeted information and minimal sensitive information over adverse compression ratios, in terms of predictive accuracy of downstream models that are trained using the compressed data.

Figures (3)

• Figure 1: Architecture for designing and utilizing TexShape.
• Figure 2: 2D visualizations of sentences in Dataset A, colored based on their public label (left) and private label (right). The top panel is dedicated to the original embedding, and the bottom panel is dedicated to TexShape embedding.
• Figure 3: Validation ROC and AUROC for classifiers trained on Dataset B for public task (left) and private task (right).