Can Large Language Models Truly Understand Context: A Study on High-Context and Low-Context Speech Acts

This study focuses on the core question of whether large language models (LLMs) truly understand context, explores their performance differences in handling high-context and low-context speech acts, analyzes the correlation between surprisal metrics and human language comprehension, and discusses the significance of this research for model evaluation and cross-cultural, multi-scenario applications.

In computational linguistics, surprisal is used to measure the model's expectation of the next word or sentence (the lower the value, the more natural it is), and it is related to human cognitive load. In cross-cultural linguistics, high-context languages (e.g., Japanese, Chinese) rely on context, cultural background, and shared knowledge; low-context languages (e.g., English, German) emphasize direct and explicit expression. LLM training data is mostly English-dominated, which may affect their sensitivity to different contextual styles.

Core question: Do LLMs assign significantly lower surprisal to low-context speech acts? Theoretical significance: It relates to whether the model truly grasps contextual sensitivity or only imitates surface patterns; Practical significance: It affects model design and evaluation in multilingual and cross-cultural scenarios. Potential findings: If low-context expressions have lower surprisal, it may reflect training data bias or the model's limitations in understanding implicit meanings; conversely, it supports the model's true understanding of context.

Machine translation: Context sensitivity affects the naturalness and authenticity of the target text; Dialogue systems: Cross-cultural scenarios require understanding implicit meanings to enhance user experience; Content generation and analysis: Avoid cultural misunderstandings or inappropriate expressions to better control output.

Traditional evaluation metrics (perplexity, BLEU) are difficult to capture context understanding capabilities; it is necessary to design test sets covering different cultures and contextual styles; training data needs to be more balanced and diverse to improve model generalization and cultural sensitivity.

This study touches on the core of AI language understanding, emphasizing that language is an interweaving of culture, context, and shared knowledge. Developers need to attach importance to contextual factors, and users need to be aware of the model's limitations in implicit meanings and cultural nuances. We look forward to more research to unlock the potential of LLMs and avoid misunderstandings and biases.