TL;DR
This paper tackles the challenge of privatizing text data while preserving its meaning, especially for ambiguous words. They achieve this by replacing standard word embeddings with "sense embeddings" and incorporating a word-sense disambiguation step before applying privacy mechanisms, leading to improved utility and comparable privacy guarantees.
What Problem Does This Paper Address?
The increasing use of language models trained on vast amounts of text raises significant privacy concerns, as personal information can be inadvertently leaked. Text-to-text privatization, often using differential privacy, aims to mitigate this by substituting words with others that are semantically similar. However, a key limitation of traditional word embeddings is their inability to capture the multiple meanings of ambiguous words (like "bank"). This leads to inappropriate substitutions that can degrade the utility of the privatized text and potentially hinder downstream tasks. Contextualized embeddings, while better at handling ambiguity, are difficult to use for privatization due to their dynamic nature. This paper seeks to bridge this gap by developing a privatization method that accounts for word senses.
Methods
The core of this work is the adaptation of the metric differential privacy mechanism for text-to-text privatization. The authors replace the standard word embeddings used in prior work with sense embeddings. These sense embeddings represent words with multiple vectors, each corresponding to a distinct meaning. To effectively utilize these sense embeddings, they introduce a crucial word-sense disambiguation (WSD) step before the privacy perturbation. This WSD step uses a context window to determine the most likely sense of an ambiguous word in a given text.
The sense embeddings themselves are constructed using a graph-based approach, clustering nearest neighbors to induce word senses. The privatization process then involves:
- Sense Disambiguation: Mapping an ambiguous word to its most probable sense vector based on context.
- Noise Injection: Adding calibrated noise to the selected sense vector, controlled by a privacy budget ($\epsilon$).
- Substitution: Projecting the noisy vector back to the discrete vocabulary space of sense embeddings and finding the nearest neighbor as the substitute word.
Datasets
The paper utilizes several datasets for their experiments:
- Wikipedia: Used for constructing the word embeddings and sense embeddings.
- Words in Context (WiC): Used for evaluating the impact of privatization on downstream classification tasks.
- WordSim-353, SimLex-999, and SCWS: Standard datasets for evaluating word similarity to measure the utility of the privatized text.
Key Insights and Ideas
- Ambiguity is a major hurdle for text privatization: Standard word embeddings conflate meanings, leading to poor substitutions for ambiguous words.
- Sense embeddings offer a middle ground: They provide multiple representations for a word, allowing for more nuanced privatization.
- Pre-privatization disambiguation is key: Applying a WSD step before adding noise is essential for leveraging sense embeddings effectively. This ensures the privacy mechanism operates on the correct word meaning.
- Contextual information improves utility: Privatizing with sense disambiguation leads to substitutions that better reflect the original word's meaning in its context, as evidenced by improved performance on downstream tasks.
- Privacy budget recalibration is necessary: Using sense embeddings can shift the required privacy budget, meaning a direct comparison of $\epsilon$ values might not be straightforward, but the shape of the privacy-utility trade-off remains similar.
Key Results
- Improved Classification Accuracy: On the Words in Context dataset, their proposed method (sense embeddings + WSD) achieved a 6.05% increase in classification accuracy compared to baseline privatization using only word embeddings.
- Enhanced Word Similarity: For the SCWS dataset (which provides context), their method showed a higher correlation (0.1911 vs. 0.0696 for word embeddings) between privatized word similarities and human annotations, indicating better utility. For WordSim-353 and SimLex-999 (without explicit context), the performance was comparable to word embeddings.
- Privacy Analysis: The study found that the privacy proxies (Nw and Sw) for sense embeddings followed a similar shape to word embeddings but allowed for a larger privacy budget to achieve comparable plausible deniability. This suggests that sense embeddings can be more efficient in terms of the privacy-utility trade-off.
- Qualitative Analysis: Visualizations showed that lexical privatization resulted in a blend of substitutions, while sense-disambiguated privatization created distinct boundaries for substitutions based on context (e.g., financial vs. geographical "bank").
Limitations and Future Work
- Interpretability of Senses: The paper acknowledges that the sense embeddings used are not mapped to an external, interpretable lexicon, making the individual senses difficult to understand.
- Superseded by Contextual Embeddings: The authors note that contextual embeddings (like BERT) are more advanced and offer better sense awareness. However, their dynamic nature still poses challenges for direct application in privatization.
- Future Work: The authors suggest that sense embeddings and contextual embeddings are not mutually exclusive and that future work could explore alternating between them for privatization and optimization. They also highlight the need to address the interpretability of senses.
Summary updated