1　Introduction

This section points out that although large language models provide support for translation, current applications are limited to conversational interaction and output production, with their potential for multidimensional language quality assessment yet to be fully recognized. This paper explores new ways of embedding intelligent assessment into the translation process, utilizing multilingual large language models through metrics such as semantic similarity and perplexity to achieve dynamic interaction between translators and models, thereby providing solutions for digital intelligence-empowered translation education.

2　Applications of Large Language Models in Translation and Translation Teaching

This section reviews the current status and limitations of large language model applications in translation and translation teaching. In the translation field, "machine translation plus post-editing" has become the mainstream model, significantly improving translation efficiency and consistency; however, issues persist regarding accuracy, semantic deviation, and inadequate capture of cultural connotations when processing complex texts such as legal, traditional Chinese medicine, and literary texts, as well as video subtitle translation, with significant performance variations within the same text. In translation teaching, large language models have promoted the transformation of teaching models toward intelligent, interactive human-machine collaboration, demonstrating feasibility and effectiveness in digital resource construction and translation quality assessment feedback, with related research confirming that intelligent assessment can improve feedback quality and timeliness and enhance learners' cognitive engagement. However, current applications suffer from single scenarios and students' tendency to over-rely on model outputs; existing intelligent assessment models mainly depend on prompts for qualitative evaluation, lacking numerical assessment and making intuitive comparison difficult. Moreover, they mostly serve as teacher assistance tools, limiting students' real-time assessment access during autonomous learning and teachers' timely diagnosis of learning situations, with their educational value in cultivating critical thinking and self-efficacy yet to be fully developed. Based on this, the study proposes constructing a multidimensional intelligent assessment method based on semantic computation and creating an AI-empowered operational platform to achieve deep integration of intelligent assessment and autonomous learning.

3　Language Capabilities and Assessment Dimensions of Large Language Models

This section discusses the core technologies of large language models and their application dimensions in language assessment. First, it elaborates on cross-lingual word embedding technology, explaining how it achieves semantic correspondence of multilingual vocabulary in a shared space, providing a new path for solving semantic computation problems. Second, it analyzes the core components of large language models, including transformer architecture, self-attention mechanism, positional encoding, and chain-of-thought technology, noting that these technologies enable models to simulate human reasoning processes and produce more coherent and logical language. It also introduces key technologies such as shared token mechanisms and embedding alignment, explaining their roles in maintaining contextual consistency and achieving knowledge transfer. Regarding assessment dimensions, it focuses on two measurements: semantic similarity measurement, which calculates semantic relationships of words, sentences, and segments through distance computation in high-dimensional vector space, transforming the traditional linguistic notion of context-determined semantics into computable digital expression; and naturalness measurement, which uses perplexity indicators to assess the degree of fit between sentences and the language patterns learned by models, combining log-likelihood and masked prediction methods to judge language naturalness and verifying the model's characteristics of integrating grammatical capabilities with world knowledge. Finally, it points out that generative artificial intelligence, through high-dimensional vector computation, provides holistic semantic judgment and can collaborate with humans to complete translation quality analysis tasks based on semantics and pragmatics.

4　Embedding Large Language Models into Translation and Translation Teaching Processes

This section discusses the application of large language models in the post-editing stage, proposing a new model of collaboration between human intelligence and artificial intelligence. This model achieves rapid response through agents, supports providing modification suggestions for translation segments, and allows human adjustment of agents to accelerate document editing, with finalization after human review. The research focuses on an intelligent collaboration model that can be locally deployed and facilitates management of translation process data; its core workflow includes: large language models providing real-time intelligent assessment results for translations, covering source-target semantic similarity and target naturalness/perplexity assessment. Assessment can be based on perplexity from the Chinese-bert-wwm-ext model, or calculated through weighted combination of perplexity and similarity to user-created Chinese reference sentence embedding libraries.

To embed intelligent assessment into the translation process, the study constructed a user-friendly platform whose working interface includes source language, target language, sentence-level alignment (initial translation and post-editing), and instantly updated numerical values for semantic similarity, naturalness, and composite language quality assessment. Assessment dimensions are determined as semantic similarity (reflecting fidelity) and target naturalness (reflecting target language normativity), with display methods based on vectorized language data for instant updating. Data is stored in semi-structured form in MySql database, supporting timestamp-based recording of multiple post-editing assessment data and retrieval through search pages.

The design addresses two key issues: external knowledge bases and composite language quality indices. Regarding the unsatisfactory perplexity assessment of short translations by BERT-class models, the study established a self-built vectorized Chinese sentence embedding library as an external target language reference library, obtaining comprehensive naturalness assessment values through sentence embedding comparison combined with perplexity (weights of 70% and 30%). The Chinese-English translation platform also adopts traditional language quality assessment indices, including TTR, average sentence length, clause count, Flesch_Kincaid readability, and lexical repetition rate, forming interpretable composite assessment indices through normalization processing to provide reference when artificial intelligence assessment shows deviation.

Translation process data is stored in structured form, covering student identity information, source texts, each modified translation, submission times, and various assessment data. This mechanism supports students in querying their own translation processes to promote reflection and provides direct data for teachers to observe learning behaviors. Teachers can design content based on teaching objectives, use large language models to generate source texts and interfering translations, have students complete post-editing through human-machine collaboration, and revise translations under intelligent assessment feedback guidance. Through system backend monitoring, teachers can observe translation quality, conduct horizontal comparisons or longitudinal tracking, achieving targeted teaching. Teaching trials found that students demonstrate high efficiency in handling spelling, grammar, and obvious mistranslations and omissions, while requiring considerable cognitive effort in addressing language idiomaticity and subtle semantic collocation issues. Students can identify errors based on feedback, reflect on agent preferences, and adjust expressions; teachers need to guide students to correctly understand the relative reference nature of assessment scores and maintain cognitive subject status. The longitudinal data stored by the platform can effectively track students' multi-round modification processes, providing an empirical foundation for teaching.

5　Conclusion

This section discusses the value and practical paths of embedding intelligent assessment into translation teaching. Intelligent assessment promotes learners' cognitive engagement and deep learning through real-time feedback and numerical assessment, driving classroom transformation toward student-centeredness, with teachers' roles shifting to "instructional facilitators" and "content innovators" and large language models becoming "collaborators" for teachers and students. It also emphasizes that teachers need to treat assessment results prudently and guide students to pay attention to discourse context. Finally, it notes that the multidimensional language capabilities of large language models still await further development and application to engender new translation teaching paradigms.