Introduction
The field of Natural Language Processing (NLP) has historically excelled at objective information extraction—tasks with clear, factual answers. However, human communication is fundamentally subjective, rich with nuance, emotion, and figurative intent. This survey paper provides a comprehensive roadmap of the transition from objective fact-extraction toward Subjective Language Understanding (SLU). It defines SLU as a unified field and explores how Large Language Models (LLMs) are transforming our ability to interpret what humans actually mean, rather than just what they literally say.
A Unified Framework for Subjectivity
The core challenge of SLU lies in its ambiguity and context-dependence. Unlike objective tasks, subjective language often lacks a single "ground truth." The paper categorizes SLU into eight primary tasks: sentiment analysis, emotion recognition, sarcasm detection, humor detection, metaphor recognition, stance detection, intent recognition, and aesthetic evaluation.
As illustrated in the figure above, modern foundation models like GPT-4o, Claude, and DeepSeek serve as the roots for a wide array of specialized techniques and benchmarks across these domains. This represents a paradigm shift from narrow, task-specific models toward general-purpose architectures that leverage vast world knowledge and emergent reasoning capabilities.
Methodological Evolution
The survey identifies four primary approaches to tackling SLU with LLMs:
- Prompt-based Learning: Techniques such as Chain-of-Thought (CoT) prompting allow models to "reason" through subjective nuances. For instance, a model might explain why a specific context makes a statement sarcastic before providing a final label.
- Supervised Fine-Tuning (SFT): Using Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA, researchers adapt general models to specialized subjective domains, such as clinical empathy or financial sentiment.
- Reasoning-based Architectures: Novel frameworks like "Reasoning through Perspective Transition" (RPT) enable models to switch between direct and third-person perspectives, better capturing social nuances.
- Multi-model and Multimodal Integration: Recognizing that subjectivity is often conveyed through delivery, frameworks like Commander-GPT coordinate specialized models to analyze text alongside acoustic or visual cues.
Key Insights: Knowledge and Theory of Mind
A significant finding of the survey is the "Understanding Gap." While LLMs have reached human-level performance in sentiment analysis, they still hit a stumbling block with the high-level pragmatic inference required for humor and sarcasm. However, their success in metaphor recognition highlights the importance of world knowledge; to understand a metaphor, a model must understand the underlying concepts, not just the words (e.g., knowing who Usain Bolt is to understand a metaphor about speed).
Furthermore, the authors argue that SLU is essentially an exercise in Theory of Mind (ToM)—the ability to attribute mental states to others. LLMs are the first models to demonstrate sparks of ToM, allowing them to infer a speaker's true attitude even when it contradicts their literal phrasing.
Benchmarking the "Gray Areas"
The paper synthesizes results across diverse benchmarks:
- Affective Tasks: SST-5 and GoEmotions for sentiment and fine-grained emotion.
- Figurative Tasks: SARC and MUStARD for sarcasm, and MUNCH for metaphors.
- Opinion/Intent Tasks: SemEval-2016 and VAST for stance detection.
- Emerging Tasks: AesBench and Textual Aesthetics, which challenge models to judge the "beauty" or "engagement" of content—perhaps the most subjective task of all.
Results and Performance
In many low-resource tasks, zero-shot LLMs often outperform older supervised models due to their massive pre-training. However, efficiency remains a factor; while large models like GPT-4 set the state-of-the-art, smaller fine-tuned models (like RoBERTa) remain competitive for high-volume, basic classification. A major advantage of LLMs is explainability—their ability to provide rationales for their classifications is vital for human-AI trust.
Challenges and Future Directions
Despite progress, several hurdles remain. The "black box" nature of LLM reasoning makes it unclear if models truly "understand" emotion or simply predict patterns. Cultural bias is another major concern, as most datasets are Western-centric, risking the amplification of specific cultural norms regarding what is considered "offensive" or "aesthetic."
The survey concludes by calling for Unified Subjective Language Modeling—a single, cohesive approach to handling the entire spectrum of human nuance simultaneously. Moving forward, the community must develop more human-aligned evaluation frameworks that move beyond simple accuracy to model the natural distribution of human opinions.