Multiphase and Multitask Prompt Tuning for LLM-Based Context-Aware Machine Translation.

Area of Science:

• Natural Language Processing
• Machine Translation
• Artificial Intelligence

Background:

• Current context-aware machine translation (MT) methods using large language models (LLMs) often process intrasentence and intersentence contexts uniformly.
• This undifferentiated approach overlooks the distinct roles these contexts play in achieving accurate translations.

Purpose of the Study:

• To propose a novel strategy, multiphase prompt tuning (MPT), that enables LLMs to distinguish between intrasentence and intersentence contexts for improved MT.
• To enhance the model's ability to leverage intersentence dependencies for more coherent and contextually relevant translations.

Main Methods:

• MPT divides the MT task into three distinct phases: intersentence context encoding, source sentence encoding, and final decoding.
• Each phase utilizes unique continuous prompts to guide the LLM's focus.
• A multitask fine-tuning approach with auxiliary tasks (context-agnostic translation, cross-lingual next sentence generation) was employed to emphasize context differentiation and intersentence dependencies.

Main Results:

• The proposed MPT strategy allows LLMs to process intrasentence and intersentence contexts distinctly.
• Multitask fine-tuning enhances the model's capacity to capture and utilize intersentence dependencies.
• The methods improve the handling of discourse-related challenges in machine translation.

Conclusions:

• Multiphase prompt tuning (MPT) offers a more nuanced approach to context-aware machine translation with LLMs.
• Differentiating context types and emphasizing intersentence dependencies leads to more effective MT systems.
• This strategy holds promise for advancing the field of natural language processing and machine translation.