1　Introduction

This section introduces how deep learning and Transformer architecture have propelled the rise of large language models in the NLP field, points out the relative lack of current research on AI-assisted quantitative and qualitative analysis in linguistics, and clarifies that this study will take Sinclair's extended unit of meaning model as its object and use GPT-5 as an example to explore the auxiliary utility of artificial intelligence for researchers.

2　Research Background and Literature Review

This section discusses the current application status of generative AI in linguistic research and the theory of extended unit of meaning. The application of generative AI is mainly manifested in two aspects: research topic expansion and methodological innovation—the former promotes rethinking of traditional linguistics and generates new issues, while the latter provides tool support for various research levels; however, domestic research mostly focuses on macro-level areas, lacking fine-grained discussion of specific linguistic research tasks. Extended unit of meaning, proposed by Sinclair, comprises five elements: core, collocation, colligation, semantic preference, and semantic prosody, forming a unified form-function whole. Despite its important theoretical value, it suffers from insufficient operability due to conceptual ambiguity and the complexity of co-selection relations, while the intervention of generative AI is expected to help break through this bottleneck.

3　Research Design and Methods

Research design and methods, aiming to examine GPT-5's performance in processing extended unit of meaning tasks. The study selected 471 "Cover Story" articles from China Daily website as corpus, with "development" as the node word, and through comparative analysis between traditional corpus tools and GPT-5, evaluated its performance in tasks such as obtaining collocates and colligations, and analyzing semantic preference and semantic prosody. Research steps include: determining the node word, using AntConc and Python scripts to obtain collocations and colligations, utilizing GPT-5 to assist semantic analysis, and exploring its capability in describing semantic prosody at the discourse level. To ensure analytical effectiveness, the researcher designed English prompts for each环节.

4　Results and Analysis

With "development" as the node word, this section comparatively analyzes the performance of script retrieval and GPT-5 processing in extended unit of meaning analysis. The study found that GPT-5 controlled errors within 5 for collocation frequency statistics and within 10 for word class statistics, effectively extracting lexical information at specified grammatical positions and compensating for oversights in manual observation. In colligation pattern recognition, GPT-5 accurately extracted structures meeting requirements with only acceptable minor frequency deviations, and its mechanism of actively prompting internal models enhanced the interpretability of analysis. The study further tested GPT-5's semantic preference analysis capability, finding that although it cannot completely replace researchers, it can provide ideas and improve efficiency for analysis. Through debugging, the researcher categorized GPT-5's output adjectives into three semantic preferences: "manner," "aspect," and "significance," and divided verbs into categories such as "promote," "hinder," "observe," and "pursue," ultimately constructing composite lexical item structures and determining corresponding semantic prosodies. Additionally, GPT-5 performed well in semantic prosody discourse management, accurately identifying linear relationships between composite lexical items, but its effectiveness declined during batch processing, requiring optimization of prompts and分批输入 to improve analysis quality.

5　Discussion

This section discusses GPT-5's application performance in four levels of extended unit of meaning analysis: in quantitative statistics of collocation and colligation, GPT-5 has problems with data accuracy and shows no obvious advantage compared to traditional corpus tools; in qualitative analysis levels such as semantic preference and semantic prosody, GPT-5 demonstrates significant advantages through its broad applicability, rich embedded knowledge, and theoretical application capabilities. The section also introduces three modes of human-AI collaboration—Embedding mode, Copilot mode, and Agents mode—pointing out that current linguistic research mainly adopts the first two modes, and emphasizing that researchers should optimize usage strategies to enhance AI's practical value, while maintaining a prudent attitude toward generative AI, reasonably guiding its tool attributes to advance research development.

6　Conclusion

This section evaluates GPT-5's performance in extended unit of meaning analysis, pointing out that it can significantly improve research efficiency in quantitative extraction and qualitative analysis, covering tasks such as word frequency statistics, word class identification, semantic feature summarization, attitudinal semantic analysis, and semantic prosody interaction parsing; it also reveals problems such as declining generation quality when processing large-scale corpora and lack of response consistency, proposes solutions including batch processing and API temperature parameter adjustment, emphasizes the key role of prompt engineering, and prospects the strategic use and continuous development of generative AI in linguistic research.