关键词: 意图识别, 槽位填充, 大语言模型, 对比学习, 注意力机制

Abstract: To address the semantic ambiguity and difficulty in candidate selection caused by the diverse results generated by large language models in intent recognition and slot filling tasks, a new approach for joint intent-slot recognition based on prompt-guided generation and tri-mutual attention was proposed，named PGT-Joint(Prompt-guided Generation and Tri-mutual Attention for Joint Intent–Slot Recognition). This approach first constructs prompts based on domain knowledge to guide the large language model in generating multiple candidate intent-slot combinations. Secondly, a tri-mutual attention mechanism is designed to explicitly model the high-order semantic relationships between user questions, candidate intents, and slots. Finally, a comparative learning-to-rank model is used to filter and optimize the candidate results. Experimental results on a Mini/Micro LED domain question-answering dataset show that, compared with baseline models such as BERT-Intent and BERT-Slot, PGT-Joint improves intent recognition accuracy by 0.4 to 3.8 percentage points and slot filling F1 score by 4.1 to 6.6 percentage points. Notably, the most significant gain is observed in joint matching accuracy, with an improvement of 5.2 to 10.2 percentage points. In addition, ablation experiments verified the key role of prompt generation and ternary interactive attention mechanism in performance improvement, indicating that PGT-Joint can effectively enhance semantic understanding capabilities in professional domain dialogue systems.

Key words: intent detection, slot filling, large language model, contrastive learning, attention mechanism