基于LSTM模型与情感词典的翻译情感一致性分析

Sentiment Consistency Analysis of Translations Based on LSTM and Sentiment Lexicon

摘要: 随着大语言模型广泛应用于翻译领域，机器译文质量评估模式受到更多关注。目前的译文质量评估框架多聚焦于词汇准确性和语法正确性，却往往忽略了原文与译文之间的情感一致性。情感表达的偏差不仅影响信息传递的完整性，还可能引发跨文化沟通中的误解。针对这一问题，本文提出在译后编辑过程中引入LSTM (长短期记忆网络)模型与情感词典的双路径融合情感分析方式进行情感分析，为翻译质量评估增加新的评估维度从而优化机器译文的情感表达效果。实验结果表明，引入该框架后，提高了对译文中的情感弱化、极性偏差等问题的有效识别，突破了原有质量评估框架的局限性。

Abstract: With the widespread adoption of large language models (LLMs) in the field of translation, increasing attention has been drawn to the evaluation paradigms of machine translation (MT) quality. Existing MT quality assessment frameworks primarily focus on lexical accuracy and grammatical correctness, while the consistency of sentiment between the source text and its translation is often overlooked. Deviations in sentiment expression not only compromise the integrity of information transfer but may also lead to misunderstandings in cross-cultural communication. To address this issue, this study proposes incorporating a dual-path sentiment analysis approach—combining a Long Short-Term Memory (LSTM) network with a sentiment lexicon—into the post-editing process. This method introduces a new evaluation dimension for translation quality assessment, thereby enhancing the emotional fidelity of machine-generated translations. Experimental results demonstrate that the proposed framework significantly improves the detection of sentiment weakening and polarity shifts in translations, overcoming limitations inherent in existing evaluation models.