Disentangled and Robust Representation Learning for Bragging Classification in Social Media

Xiang Li, Yucheng Zhou

Introduction

Bragging, or self-promotion, is a pervasive social phenomenon on platforms like Twitter and LinkedIn. While identifying these behaviors is crucial for understanding social dynamics and personality traits, computational linguistics faces a significant challenge: extreme data imbalance. In most social media datasets, non-bragging posts vastly outnumber instances of actual bragging, often by a ratio of 50 to 1.

Traditional methods to address this imbalance typically rely on external knowledge sources, such as emotion lexicons (NRC) or linguistic inquiry tools (LIWC). However, these external dictionaries often introduce noise that can degrade model performance. This paper introduces a novel framework that bypasses external dependencies by mining latent information directly from the text through feature disentanglement and adversarial learning.

Methodology: Disentanglement and Augmentation

The authors propose a two-pronged approach: Disentangle-based Feature Augmentation (DFA) and a Domain-aware Adversarial Strategy (DAS). The core idea is to treat bragging as a linguistic "wrapper" that can be separated from the underlying topic or content of a post.

1. Disentangle-based Feature Augmentation (DFA)

The model utilizes a pre-trained BERTweet encoder to generate initial text representations. These representations are then processed through specialized Multi-Layer Perceptrons (MLPs) to isolate two distinct vectors:

  • Content Feature ($h_c$): Captures the semantic topic of the post (e.g., a family memory or a specific possession).
  • Bragging-type Feature ($h_t$): Captures the linguistic markers and stylistic elements associated with self-promotion.

By "mixing" the content feature of one post with the bragging-type feature of another, the model generates synthetic training samples ($h^{ji}$). This allows the model to create new, valid examples of rare bragging classes without needing new raw data.

The proposed framework featuring Disentangle-based Feature Augmentation (DFA) and Domain-aware Adversarial Strategy (DAS).

2. Domain-aware Adversarial Strategy (DAS)

To ensure that synthetic features remain realistic and follow the distribution of real social media posts, the authors implement an adversarial discriminator. While the encoder attempts to create augmented features that are indistinguishable from real ones, the discriminator tries to identify the synthetic samples. This adversarial "cat-and-mouse" game forces the model to learn robust, high-quality representations that generalize well to real-world data.

Experimental Setup and Datasets

The study utilizes the first large-scale bragging dataset (Jin et al., 2022), consisting of 3,382 training and 3,314 test samples sourced from Twitter. The dataset includes six specific bragging categories—Achievement, Action, Feeling, Trait, Possession, and Affiliation—alongside a dominant Not Bragging class. The imbalance is stark: the "Not Bragging" class contains over 2,800 examples, while some specific categories like "Affiliation" have fewer than 60.

Key Results

The proposed model achieved State-of-the-Art (SOTA) performance in multi-class classification, significantly outperforming previous benchmarks that relied on external clusters or lexicons.

Performance comparison of the proposed method against various baselines on the bragging classification task.

As shown in the results table, the model reached a Macro F1 score of 39.86, a substantial improvement over the previous best of 35.59 (BERTweet-Clusters). In binary classification (Brag vs. Not Brag), the model achieved an F1 of 73.07. Ablation studies further confirmed the importance of the framework: removing the adversarial strategy (DAS) resulted in a 1-point drop in F1, while removing both DFA and DAS led to a massive 5-point decline, highlighting that the "mix-and-match" augmentation is the primary driver of success.

Discussion and Conclusion

The success of this method suggests that the "signal" for bragging is already present within latent text representations; it simply needs to be isolated from the noise of the general content. By treating bragging as a stylistic layer, the model learns the essence of self-promotion regardless of the topic being discussed.

While a Macro F1 of ~40% underscores the inherent difficulty of interpreting nuanced human social behavior, this work provides a robust path forward for low-resource classification tasks. Future research may explore how these "bragging-type" features translate across different languages or professional platforms like LinkedIn, where self-promotion styles may differ significantly from Twitter.

Summary updated

BibTeX
@inproceedings{li2023disentangled,
  author = {Xiang Li and Yucheng Zhou},
  title = {{Disentangled and Robust Representation Learning for Bragging Classification in Social Media}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10096519},
}
Figure