2. Related Work
2.1. Limitations of LLMs (e.g., GPT, BERT) in Interpreting Complex Meaning
Large Language Models (LLMs), such as GPT (Brown et al., 2020; OpenAI, 2023) and BERT (Devlin et al., 2019), have demonstrated remarkable capabilities in a wide range of natural language processing (NLP) tasks, from machine translation and summarization to question answering and dialogue generation. These systems achieve impressive fluency and coherence by leveraging large-scale transformer architectures and training on massive corpora to predict word sequences based on contextual embeddings.
Despite these achievements, there are increasing concerns regarding the depth and authenticity of the understanding that these models exhibit. Fundamentally, both autoregressive models (like GPT) and masked language models (like BERT) operate on statistical inference principles---namely, predicting the most probable next token or masked token given a context. This probabilistic mechanism, while effective in producing plausible surface-level outputs, often fails to capture latent, symbolic, emotional, or culturally nuanced meaning that is essential to human communication.
Several key limitations are widely documented:
1. Surface Semantics Bias
LLMs excel at syntactic pattern recognition but frequently conflate statistical proximity with semantic depth. For example, phrases such as "crocodile tears" or "burning silence" carry metaphorical and cultural significance that cannot be derived solely from frequency-based token associations.
2. Inadequate Contextual Memory
While transformer-based models are designed to attend to prior context, they often exhibit limited ability to model discourse-level coherence or long-range semantic dependencies, especially when meaning emerges across multiple utterances or relies on shared background knowledge.
3. Lack of Interpretability and Conceptual Grounding
LLMs operate as black-box statistical machines, lacking explicit symbolic reasoning or semantic representations. Consequently, they fail to reason about underlying concepts, analogies, or symbolic relations that go beyond surface tokens.
4. Cultural and Aesthetic Blind Spots
 LLMs trained on broad web corpora often internalize dominant language patterns, but lack sensitivity to non-dominant cultural metaphors, regional idioms, or linguistic aesthetics (e.g., poetic rhythm, emotional undertones). This limits their utility in creative writing, intercultural dialogue, or artistic co-creation tasks.
5. Figurative Language Challenges
Empirical studies show that LLMs perform poorly on tasks involving metaphor, irony, sarcasm, or allegory (Chakrabarty et al., 2022; Bisk et al., 2020). These forms of expression often rely on multi-layered interpretation, which probabilistic token-level prediction struggles to replicate without structured conceptual modeling.
Several attempts have been made to mitigate these issues, such as incorporating external knowledge graphs (Zhang et al., 2019), grounding in multimodal contexts (e.g., CLIP by Radford et al., 2021), or aligning outputs using reinforcement learning from human feedback (Ouyang et al., 2022). However, these augmentations still operate within the overarching statistical paradigm and are limited in their capacity to model meaning as a dynamic interaction system.
The CAS-6V framework proposed in this paper builds upon these insights but shifts the perspective entirely---from treating meaning as emergent from prediction over static tokens to modeling it as the result of structured, multi-variable, and adaptive interactions between linguistic elements. This distinction is critical for progressing toward AI systems that do not merely simulate language, but genuinely interpret and generate it in a manner aligned with human-level abstraction, affect, and creativity.
2.2. Insights from Cognitive Semantics, Emergent Semantics, Semiotics, and Complex Adaptive Systems
The limitations of probabilistic language modeling in current LLMs have motivated scholars to explore alternative paradigms that draw from cognitive linguistics, philosophy of language, semiotics, and complex systems theory. These interdisciplinary perspectives offer critical insights into how meaning emerges, evolves, and is interpreted beyond surface-level token prediction.
2.2.1 Cognitive Semantics and Conceptual Metaphor Theory
Cognitive semantics, particularly as developed by George Lakoff and Mark Johnson (1980), and further expanded by Gilles Fauconnier and Mark Turner (2002), posits that meaning arises from embodied experience and conceptual mapping, rather than from purely syntactic relations. Core ideas such as image schemas, conceptual blends, and metaphorical projection emphasize that linguistic meaning is deeply tied to human cognition, perception, and culture.
