DXA-Net: Dual-Task Cross-Lingual Alignment Network for Zero-Shot Cross-Lingual Spoken Language Understanding

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Current zero-shot cross-lingual spoken language understanding (SLU) models use unsupervised contrastive learning for semantic alignment.
• Existing methods face limitations in cross-lingual knowledge transfer due to unmodeled dual-task correlations and ignored sample semantic differences.

Purpose of the Study:

• To propose a novel dual-task cross-lingual alignment network (DXA-Net) for zero-shot cross-lingual SLU.
• To address limitations in existing cross-lingual knowledge transfer methods.
• To enhance multilingual semantic alignment and improve zero-shot SLU performance.

Main Methods:

• Developed DXA-Net, a prompt-tuning paradigm for zero-shot cross-lingual SLU.
• Introduced a co-guiding prompt to model and transfer dual-task correlative knowledge.
• Proposed intent/slot contrastive prompts and multilingual semantics contrastive prompts to address semantic differences and enhance alignment.

Main Results:

• DXA-Net achieves new state-of-the-art performance on zero-shot cross-lingual SLU tasks.
• The proposed prompts effectively enable conditional label generation and discrimination of sample similarities.
• Significant improvements in multilingual semantic alignment were observed.

Conclusions:

• DXA-Net represents a significant advancement in zero-shot cross-lingual spoken language understanding.
• The novel prompt-based approach effectively tackles key challenges in cross-lingual knowledge transfer.
• The model demonstrates robust performance across multiple languages, setting a new benchmark.