DelTA: An Online Document-Level Translation Agent Based on Multi-Level Memory

TL;DR

DelTA introduces an online document-level translation agent built on a multi-level memory system to address translation inconsistency and inaccuracy in long texts. By organizing information into Proper Noun Records, Bilingual Summaries, and Long/Short-Term Memory, and by performing sentence-by-sentence translation with memory-informed prompts, DelTA achieves stronger consistency (LTCR improvements up to 4.58 percentage points) and quality (COMET gains up to 3.16) across multiple datasets and backbones. The approach also enhances pronoun and context-dependent translation and enables a complementary summarization capability via its bilingual and segment summaries. In addition to improved translation reliability, DelTA demonstrates memory efficiency compared to Doc2Doc methods, making it suitable for practical deployment on local hardware and adaptable to varied domains.

Abstract

Large language models (LLMs) have achieved reasonable quality improvements in machine translation (MT). However, most current research on MT-LLMs still faces significant challenges in maintaining translation consistency and accuracy when processing entire documents. In this paper, we introduce DelTA, a Document-levEL Translation Agent designed to overcome these limitations. DelTA features a multi-level memory structure that stores information across various granularities and spans, including Proper Noun Records, Bilingual Summary, Long-Term Memory, and Short-Term Memory, which are continuously retrieved and updated by auxiliary LLM-based components. Experimental results indicate that DelTA significantly outperforms strong baselines in terms of translation consistency and quality across four open/closed-source LLMs and two representative document translation datasets, achieving an increase in consistency scores by up to 4.58 percentage points and in COMET scores by up to 3.16 points on average. DelTA employs a sentence-by-sentence translation strategy, ensuring no sentence omissions and offering a memory-efficient solution compared to the mainstream method. Furthermore, DelTA improves pronoun and context-dependent translation accuracy, and the summary component of the agent also shows promise as a tool for query-based summarization tasks. The code and data of our approach are released at https://github.com/YutongWang1216/DocMTAgent.

Figures (9)

• Figure 1: Framework of DelTA. The modules outlined with dashed lines represent the multi-level memory components, while those outlined with solid lines denote the LLM-based components. Memories closer to the top are more global, abstract, and densely packed with information. During translation, memory information is retrieved and incorporated into the translator LLM's prompt. After the translation of each sentence, the LLM-based components extract key information from both the source and target documents and update the multi-level memory components.
• Figure 2: Proportions of consistent translations in different sentence-wise distances.
• Figure 3: Memory cost of the Doc2Doc approach and our approach.
• Figure 4: Prompt template for the Proper Noun Extractor. We provide several few-shot exemplars preceding the current input. This template is designed for the En $\Rightarrow$ Zh translation direction. For other translation directions, adjust the corresponding content to match the specific languages.
• Figure 5: Prompt template for Source Summary Writer.