A GAN and LLM-Driven Data Augmentation Framework for Dynamic Linguistic Pattern Modeling in Chinese Sarcasm Detection
A Chinese-language sarcasm detector built on synthetic data outperforms benchmarks — but narrow dataset scope and platform-specific training limit real-world reach.
The Thesis
Detecting sarcasm in Chinese social media is genuinely hard: the language relies heavily on context, implied meaning, and individual speaking style rather than obvious verbal cues. This paper builds a new tool that combines two data-generation techniques — a GAN (Generative Adversarial Network, a system where one model generates fake text and another tries to catch it, pushing both to improve) and GPT-3.5 prompting — to create a larger training dataset than previously existed for Chinese sarcasm. The model then uses that richer data alongside each user's posting history to pick up on personal linguistic habits that signal sarcasm. The catch is that everything is trained and tested on Sina Weibo, a single Chinese platform, and the evaluation relies heavily on the authors' own synthetic dataset rather than independent benchmarks. If the claims hold up, the approach matters most for content moderation, brand-sentiment tools, and public-opinion monitoring in Chinese-language markets.
Catalyst
Two things converged: GPT-3.5 became cheap and capable enough to act as a data factory for low-resource NLP (natural language processing) tasks, and GAN-based text generation matured enough to produce plausible social-media prose in Chinese. Before roughly 2023, neither tool was accessible at the cost and quality needed to bootstrap a labeled sarcasm corpus from scratch.
What's New
Earlier Chinese sarcasm detection work — including BERT-based classifiers and multi-modal models — was bottlenecked by tiny labeled datasets, often fewer than a few thousand examples, and treated all users as interchangeable. Those models focused purely on the text of a single post, ignoring who wrote it. This paper adds a user-history dimension and inflates the training corpus synthetically, which the authors argue lets the model learn that the same phrase can be sarcastic from one user and sincere from another.
The Counter
The headline F1 scores of ~0.91 look impressive, but they are measured on a dataset the authors built themselves using the same GPT-3.5 system that generated part of the training data — a circularity that should make any reader cautious. There is no evaluation on a fully independent, human-labeled Chinese sarcasm corpus, so it is unclear whether the gains reflect genuine linguistic understanding or the model simply recognizing its own synthetic writing style. The user-history feature sounds powerful in theory, but Weibo's API restrictions mean real-world deployments would struggle to collect the same depth of user history at scale without violating platform terms. The paper also does not test cross-platform generalization: sarcasm on Weibo may look quite different from sarcasm on Douyin, WeChat, or Xiaohongshu. Finally, GAN-based text augmentation has a known failure mode — mode collapse, where the generator produces repetitive outputs — and the paper does not provide enough detail about training stability to rule this out.
Longs
- BIDU (Baidu) — dominant Chinese NLP infrastructure and search-ad trust signals
- BZUN (Baozun) — Chinese e-commerce operator reliant on brand-sentiment monitoring
- CSGP (CoStar Group) — adjacent: data-enrichment firms that sell social-signal analytics
- KWEB (KraneShares China Internet ETF) — broad exposure to Chinese platform companies that would deploy content moderation tools
Shorts
- Generic multilingual sentiment vendors (e.g., Brandwatch, Sprinklr) — if user-history-aware, language-specific models outperform their one-size-fits-all classifiers, enterprise buyers may prefer specialized tools
- Static lexicon-based Chinese NLP services — this approach makes hand-crafted sarcasm dictionaries look crude by comparison
Enablers (Picks & Shovels)
- OpenAI API (GPT-3.5/4) — used directly for data augmentation in this paper
- Hugging Face Transformers — BERT variants and fine-tuning infrastructure the model builds on
- Sina Weibo open data — the raw corpus; access policy changes would break reproducibility
- PyTorch GAN libraries — underlying generation framework
Private Watchlist
- Meltwater — social listening and sentiment analytics, active in APAC
- Synthesio (part of Ipsos) — multilingual social intelligence platform
- Moonshot AI — Chinese LLM lab whose models could power or compete with this approach
Resources
The Paper
Sarcasm is a rhetorical device that expresses criticism or emphasizes characteristics of certain individuals or situations through exaggeration, irony, or comparison. Existing methods for Chinese sarcasm detection are constrained by limited datasets and high construction costs, and they mainly focus on textual features, overlooking user-specific linguistic patterns that shape how opinions and emotions are expressed. This paper proposes a Generative Adversarial Network (GAN) and Large Language Model (LLM)-driven data augmentation framework to dynamically model users' linguistic patterns for enhanced Chinese sarcasm detection. First, we collect raw data from various topics on Sina Weibo. Then, we train a GAN on these data and apply a GPT-3.5 based data augmentation technique to synthesize an extended sarcastic comment dataset, named SinaSarc. This dataset contains target comments, contextual information, and user historical behavior. Finally, we extend the BERT architecture to incorporate multi-dimensional information, particularly user historical behavior, enabling the model to capture dynamic linguistic patterns and uncover implicit sarcastic cues in comments. Experimental results demonstrate the effectiveness of our proposed method. Specifically, our model achieves the highest F1-scores on both the non-sarcastic and sarcastic categories, with values of 0.9138 and 0.9151 respectively, which outperforms all existing state-of-the-art (SOTA) approaches. This study presents a novel framework for dynamically modeling users' long-term linguistic patterns in Chinese sarcasm detection, contributing to both dataset construction and methodological advancement in this field.