1 Introduction

Against the backdrop of the widespread application of large language models (LLMs) in society, the field of linguistics has begun exploring their application in linguistic and applied linguistic research. LLMs present new developmental opportunities for corpus-based translation studies, initiating a new phase in the research paradigm of the post-classical era. LLMs lower the technical barriers associated with tasks such as corpus construction; researchers can guide the models in text collection, processing, and data analysis using natural language prompts, thereby enhancing research efficiency. The research methodology has shifted from emphasizing search skills to competing in the era of "prompt engineering," with LLMs acting as advanced assistants to researchers.

2 Efficient Alternatives to Existing Corpus-Based Translation Research Methods

This section explores the application of LLMs in corpus-based translation studies and introduces new techniques that serve as alternatives to traditional methods. For corpus construction, LLMs can generate code to acquire and align corpora. For example, GPT-5 can be prompted to write Python code for scraping bilingual news from *China Daily*, and the Claude Sonnet 4 model can perform sentence alignment, outputting in TMX format to establish a parallel corpus. For corpus analysis, LLMs are capable of conducting multi-factor statistical analyses. For instance, R code generated by Claude Sonnet 4 can be used to analyze lexical features of translated text, including word count and type-token ratio. Furthermore, by learning from classic papers and following specific prompts, LLMs can analyze translation universals and translator style. Examples include analyzing the translation universals in Xu Yuanchong's English translation of the *Daodejing* and comparing the translator styles of the Hawkes and Yang translations of *A Dream of Red Mansions*. These alternative techniques demonstrate the potential and application prospects of LLMs in corpus-based translation research.

3 Qualitative Enhancement and Upgrading of Corpus-Based Translation Research Methods

This section delves into the innovative applications of LLMs in corpus-based translation research methodologies, highlighting their significant role in enhancing research techniques in the era of "intelligent post-editing." For sentiment analysis, LLMs, with their vast datasets, training parameters, and iteration cycles, exhibit stronger contextual understanding, enabling them to handle complex semantic structures and significantly improve the quality of sentiment analysis. Using the example of "被" (bei) constructions, a comparative sentiment analysis was conducted on the novel sections of two Chinese corpora based on the Claude Sonnet 4 model, revealing differences in emotional distribution and semantic fields between original Chinese and translated Chinese texts.

For translation strategy identification, LLMs can substantially improve the quality of translation strategy research. In previous studies, the identification of culture-loaded terms, metaphors, etc., suffered from low accuracy and high omission rates, while the judgment of translation strategies often relied on manual effort. Now, researchers can design prompts to guide LLMs in annotation and judgment, followed by modification and optimization based on the LLM outputs. For example, the GPT-5 model was used to extract Chinese culture-loaded terms and analyze their translation strategies in the Chinese-English parallel corpus of the *2023 Report on the Work of the Government*, and to annotate metaphors and parse their translation strategies in part of the parallel corpus from the *2020 Report on the Work of the Government*. Furthermore, the GPT-5 model was utilized to annotate and analyze interpreting strategies in the 2017 sub-corpus of the "Parallel Interpreting Corpus for Meetings and Events (PICME)."

For translation error identification, LLMs excel at error analysis and can be applied to error analysis in learner translation and interpreting as well as translation pedagogy research. Based on the Claude Sonnet 4 model, error analysis was performed on the Chinese-to-English written translation texts of five learners from the "Chinese College Students' English-Chinese & Chinese-English Translation and Interpreting Corpus," along with suggestions for revision.

For generating multilingual parallel corpora, LLMs can assist in constructing parallel corpora for low-resource languages. Taking the creation of a Chinese, English, and Tongan parallel corpus as an example, LLMs were used to generate parallel sentence pairs in these three languages for the category of "daily skills and recreational hobbies," providing an effective solution for corpus construction for low-resource languages.

4 Application Upgrading in Corpus-Based Translation Research Methods

This section discusses the application upgrades of LLMs in corpus-based translation research methodologies, covering four areas: bilingual term extraction, translation quality assessment, semantic similarity computation, and bilingual lexicography. For bilingual term extraction, LLMs improve the accuracy and efficiency of extraction by conducting deep learning analysis of parallel corpora with a focus on semantics. Using the *2020 Report on the Work of the Government* as an example, the process of extraction using DeepSeek-reasoner is demonstrated. For translation quality assessment, LLMs can conduct comprehensive evaluations of translated texts by considering multiple factors. Using the "TED English-Chinese Parallel Speech Corpus" as an example, GPT-5 was employed for assessment, drawing on House's model of translation quality assessment. For semantic similarity computation, LLMs can generate computational linguistics code, utilizing models like Sentence Transformer to compare the semantic similarity of parallel corpus sentences. Claude Sonnet 4 was used to generate the corresponding Python code. For bilingual lexicography, based on principles of "probabilistic contextual co-selection" and the "four-use principle" (frequency, range, availability, coverage), and taking hotel industry English as an example, the process is showcased from generating a common word list to creating illustrated entries, using GPT-5 and Tongyi Wanxiang models to complete related tasks.

5 Conclusion

This section summarizes the applications of LLMs in corpus-based translation research, noting their processing capabilities and potential to enhance research methodologies across multiple aspects, while also acknowledging limitations and potential issues that require attention.