A Multi-Task Multi-Stage Transitional Training Framework for Neural Chat Translation.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Existing Neural Machine Translation (NMT) models struggle with Neural Chat Translation (NCT) due to limited bilingual dialogue data, poor modeling of conversational properties, and training discrepancies.
• These limitations hinder the performance of context-aware NMT models in real-world cross-lingual chat scenarios.

Purpose of the Study:

• To propose a novel Multi-task Multi-stage Transitional (MMT) training framework to enhance Neural Chat Translation (NCT) performance.
• To address the inherent problems of limited annotated bilingual dialogues, neglect of conversational properties, and training discrepancies in existing NCT models.

Main Methods:

• Developed a Multi-task Multi-stage Transitional (MMT) training framework for NCT.
• Incorporated two auxiliary tasks: utterance discrimination and speaker discrimination, to model dialogue coherence and speaker characteristics.
• Implemented a three-stage training process: sentence-level pre-training, intermediate training with auxiliary tasks on monolingual dialogues, and context-aware fine-tuning with gradual transition.

Main Results:

• The proposed MMT framework significantly improved NCT performance compared to existing methods.
• The auxiliary tasks effectively introduced dialogue coherence and speaker characteristics into the NCT model.
• The three-stage training process with gradual transition alleviated training discrepancies and smoothed stage transitions.

Conclusions:

• The MMT training framework is effective and superior for Neural Chat Translation (NCT).
• The proposed methods successfully address key challenges in cross-lingual chat translation, offering improved performance and robustness.