Automated PII Extraction from Social Media for Raising Privacy Awareness: A Deep Transfer Learning Approach

Yizhi Liu, Fang Lin, Mohammadreza Ebrahimi, Weifeng Li, Hsinchun Chen

Abstract

Internet users have been exposing an increasing amount of Personally Identifiable Information (PII) on social media. Such exposed PII can cause severe losses to the users, and informing users of their PII exposure is crucial to raise their privacy awareness and encourage them to take protective measures. To this end, advanced automatic techniques are needed. While Information Extraction (IE) techniques can be used to extract the PII automatically, Deep Learning (DL)-based IE models alleviate the need for feature engineering and further improve the efficiency. However, DL-based IE models often require large-scale labeled data for training, but PII-labeled social media posts are difficult to obtain due to privacy concerns. Also, these models rely heavily on pre-trained word embeddings, while PII in social media often varies in forms and thus has no fixed representations in pre-trained word embeddings. In this study, we propose the Deep Transfer Learning for PII Extraction (DTL-PIIE) framework to address these two limitations. DTL-PIIE transfers knowledge learned from publicly available PII data to social media to address the problem of rare PII-labeled data. Moreover, our framework leverages Graph Convolutional Networks (GCNs) to incorporate syntactic patterns to guide PIIE without relying on pre-trained word embeddings. Evaluation against benchmark IE models indicates that our approach outperforms state-of-the-art DL-based IE models. Our framework can facilitate various applications, such as PII misuse prediction and privacy risk assessment, protecting the privacy of internet users.

Introduction

Internet users constantly leak PII, such as names, phone numbers, and birthdates, on social media, leaving them vulnerable to cybercriminals. Building Deep Learning (DL) models for Information Extraction (IE) in this domain faces two primary hurdles. First, the Data Scarcity Problem: DL models require large labeled datasets, yet publicly available social media posts labeled with PII are rare due to privacy concerns. Second, the Embedding Problem: Standard models rely on pre-trained word embeddings (e.g., GloVe) that often fail to recognize irregular social media PII formats, such as abbreviations or non-standard phone number layouts.

Methodology

To address these challenges, we introduce the Deep Transfer Learning for PII Extraction (DTL-PIIE) framework. The architecture combines Character Bi-LSTMs for morphological clues, Transformers for global context, and a novel PII-GCN layer for syntactic analysis.

Overview of the DTL-PIIE framework, illustrating the end-to-end workflow from data collection to PII extraction and user warning.

Transfer Learning Strategy

To bypass the lack of social media data, the model is first trained on a "source domain" consisting of 24,762 documents from six public PII datasets (e.g., i2b2, CoNLL 2003). The learned network layers are then transferred and fine-tuned on a smaller "target domain" of 8,000 manually annotated tweets.

PII-GCN: Syntax Over Semantics

The core novelty of DTL-PIIE is the use of Graph Convolutional Networks (GCNs) based on dependency graphs. Instead of processing text purely as a sequence, the model maps the grammatical syntax of a sentence as a graph where words are nodes and grammatical relationships are edges. This allows the model to identify PII based on its grammatical role—such as recognizing the object of "lives in"—rather than relying solely on semantic embeddings.

Illustration of the GCN layer using dependency graphs to capture syntactic relationships between words.

Key Results and Insights

Evaluation against state-of-the-art baselines demonstrates the efficacy of the DTL-PIIE approach:

  • Superior Performance: On the social media target domain, DTL-PIIE achieved an F1 score of 71.1%, significantly outperforming the Dependency-Guided LSTM-CRF baseline (57.2%) and a standard Bi-LSTM+CRF transfer model (67.4%).
  • Source Domain Excellence: The model set a new benchmark on public source datasets with a 94.6% F1 score.
  • Ablation Analysis: Removing the transfer learning component resulted in an 11.4% drop in F1, while removing the GCN/Dependency Graph caused a 3.9% drop, confirming that both components are critical for high-accuracy extraction.

Discussion and Future Work

The study highlights that grammatical structures surrounding PII are highly transferable across domains, even when the text style differs significantly (e.g., news vs. tweets). By prioritizing syntax over semantics, DTL-PIIE effectively handles the irregular nature of social media communication.

Future research will focus on expanding this framework to non-textual media, such as images and videos, where PII leakage is also prevalent. Additionally, the authors suggest integrating this system into real-world behavioral studies to evaluate how automated privacy warnings influence user protective measures.

Summary updated

Abstract

Internet users have been exposing an increasing amount of Personally Identifiable Information (PII) on social media. Such exposed PII can cause severe losses to the users, and informing users of their PII exposure is crucial to raise their privacy awareness and encourage them to take protective measures. To this end, advanced automatic techniques are needed. While Information Extraction (IE) techniques can be used to extract the PII automatically, Deep Learning (DL)-based IE models alleviate the need for feature engineering and further improve the efficiency. However, DL-based IE models often require large-scale labeled data for training, but PII-labeled social media posts are difficult to obtain due to privacy concerns. Also, these models rely heavily on pre-trained word embeddings, while PII in social media often varies in forms and thus has no fixed representations in pre-trained word embeddings. In this study, we propose the Deep Transfer Learning for PII Extraction (DTL-PIIE) framework to address these two limitations. DTL-PIIE transfers knowledge learned from publicly available PII data to social media to address the problem of rare PII-labeled data. Moreover, our framework leverages Graph Convolutional Networks (GCNs) to incorporate syntactic patterns to guide PIIE without relying on pre-trained word embeddings. Evaluation against benchmark IE models indicates that our approach outperforms state-of-the-art DL-based IE models. Our framework can facilitate various applications, such as PII misuse prediction and privacy risk assessment, protecting the privacy of internet users.
BibTeX
@misc{liu2021automated,
  author = {Yizhi Liu and Fang Lin and Mohammadreza Ebrahimi and Weifeng Li and Hsinchun Chen},
  title = {{Automated PII Extraction from Social Media for Raising Privacy Awareness: A Deep Transfer Learning Approach}},
  year = {2021},
  eprint = {2111.09415},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2111.09415},
}
Figure