1 Introduction

Human language has a rich vocabulary, complex grammatical structures, and the ability to adapt to different literary genres, demonstrating a high degree of variability. Noam Chomsky's syntactic structure theory provides a new perspective for understanding language, but there is still no comprehensive and accurate grammar. Modern computer technology, especially Large Language Models (LLMs), has demonstrated significant capabilities in text generation, reasoning, question answering, and other areas, becoming a new hotspot in language research. This article will review the development history and key technologies of LLM, especially how deep learning technology has driven breakthroughs in this field. Then, we will delve into the similarities and differences between LLM and humans in language acquisition, understanding, and underlying processing mechanisms, and explore whether LLM truly understands language. This article aims to provide a new perspective for understanding LLM and human language comprehension, and to offer insights for future research and applications.

The Development History and Key Technologies of the Two Major Language Models

The development of big language models began with rule-based systems and statistical methods in the mid-20th century, gradually evolving into the widespread application of deep learning techniques in the 21st century. Deep learning technology uses multi-layer neural networks for feature extraction and learning, especially after AlexNet's success in the ImageNet competition in 2012, neural networks have been widely used in fields such as syntactic analysis and image recognition. Neural machine translation (NMT) adopts an end-to-end learning approach, using models such as recurrent neural networks (RNN) and long short-term memory networks (LSTM) to capture long-range dependencies, reducing the need for manual feature engineering. The emergence of Transformer models has further improved the performance of syntactic analysis and become one of the main methods. In recent years, joint training and multimodal translation have received attention, and emerging statistical learning methods such as transfer learning and reinforcement learning have also begun to be applied in the field of pattern recognition. Deep learning models have achieved significant results in multiple fields, and hybrid intelligence has become a research hotspot. LSTM、 Attention mechanism, RNN, and Transformer are important models for processing sequential data with memory capabilities, which can be simulated and applied in computer systems through mathematical modeling and training. The development of neural computing has led to the rise of LLM, especially with deep learning, machine learning attention mechanisms, LSTM, and Transformer based generative models accelerating the widespread application of LLM. ChatGPT, born in 2022, is an LLM based chatbot that can generate realistic human like text responses, solving the challenges of natural language machine understanding and generation. ChatGPT uses deep learning techniques to automatically learn language rules and contextual information from massive corpora, generating natural and fluent text based on input and contextual information. Understanding, knowledge recall, and reasoning improve as the model size and instruction prompts are adjusted. GPT and BERT are both pre trained language models based on Transformer, using pre training and fine-tuning methods, and using word embedding techniques to capture semantic relationships between words. GPT is more suitable for generative tasks, while BERT is more suitable for discriminative tasks. When the training corpus size is medium or below, BERT performs better than GPT, but the opposite is true when the training corpus size is extremely large.

Comparative analysis of three major language models and human language

In depth exploration was conducted on the similarities and differences between the Large Language Model (LLM) and human language processing, analyzing from three dimensions: language acquisition, language comprehension, and underlying cognitive processing mechanisms. In terms of language acquisition, LLM relies on large-scale data training, which is different from the way children learn language through multimodal sensory experiences. LLM enhances interactive training capabilities through reinforcement learning and simulates human communication patterns. Research has found that the language learning content of LLM is not closely related to syntactic structure and focuses more on semantic knowledge. At the level of language comprehension, LLM exhibits reasoning abilities similar to those of humans, particularly exceeding expectations in abstract reasoning tasks. LLM also exhibits humanoid responses in Theory of Mind (ToM) tests, but lacks adaptability to changes in task conditions, indicating that it may only be imitating patterns in training data. In terms of underlying processing mechanisms, LLM is based on statistical learning and probability prediction, which contrasts with the discrete rules of traditional generative linguistics. The grading and probability calculation methods of LLM do not conflict with human language processing mechanisms, but instead provide a new perspective for understanding human language processing. Overall, LLM has shown good performance in language comprehension, but it is still inferior to humans in deep reasoning and context recognition.

Can the four major language models really 'understand' language?

Explored the question of whether the Large Language Model (LLM) truly understands language. There is extensive discussion in the academic community about this, and some experts believe that as the model parameters and training data increase, the performance of LLM is significantly improving, potentially achieving levels of intelligence and understanding close to those of humans. They believe that MLL can not only truly understand language, but also perform conventional reasoning, although it may not yet reach human level. However, some argue that LLM, no matter how human like it may appear, lacks true understanding ability because it lacks world experience and can only achieve a superficial level of 'understanding'. In addition, some scholars have pointed out that humans are prone to falling into the "Eliza effect" and overestimating the understanding and self-awareness of machines. Overall, LLM's "understanding" is based on statistical pattern recognition of large amounts of data, which goes beyond the general understanding of human concepts. As the scale and capabilities of LLM grow, we need to expand the existing boundaries of intelligent science and provide a broader definition of 'understanding' that applies to both human cognition and machine intelligence performance. LLM can be considered as a new type of "understanding" that can bring "understanding" beyond human predictive ability.

5 Conclusion

Reviewed the development history and key technologies of LLM, and compared the similarities and differences between LLM and human language processing. LLM has potential in specific language tasks, but it has not replaced human language abilities. The future development of LLM should be combined with human cognition and socio-cultural practices, playing an auxiliary role while ensuring ethics and safety. The academic community should explore the interpretability and transparency of LLM, promote human-machine collaboration, and advance the development of language science and artificial intelligence. In depth research on LLM helps to understand machine language processing capabilities and deepen our understanding of the complexity of human language.