Zero-shot cross-domain dialogue state tracking (DST) enables us to handle task-oriented dialogue in unseen domains without the expense of collecting in-domain data. In this paper, we propose a slot description enhanced generative approach for zero-shot cross-domain DST. Specifically, our model first encodes dialogue context and slots with a pre-trained self-attentive encoder, and generates slot values in an auto-regressive manner. In addition, we incorporate Slot Type Informed Descriptions that capture the shared information across slots to facilitate cross-domain knowledge transfer. Experimental results on the MultiWOZ dataset show that our proposed method significantly improves existing state-of-the-art results in the zero-shot cross-domain setting.

Check the packages needed or simply run the command

❱❱❱ pip install -r utils/requirements.txt

Dataset

❱❱❱ python create_data.py

use create_data_2_1.py if want to run with multiwoz2.1

Zero-shot cross-domain

❱❱❱ python T5.py --train_batch_size 16 --GPU 8 --except_domain ${domain} --slot_lang ${description type}

• --GPU: the number of gpu to use
• --except_domain: hold out domain, choose one from [hotel, train, attraction, restaurant, taxi]
• --slot_lang: slot description type, choose one from [none, human, naive, value, question, slottype]
• Note: real batch_size = train_batch_size * GPU_number * gradient_accumulation_steps

Few-shot cross-domain

❱❱❱ python T5.py --train_batch_size 16 --GPU 8 --slot_lang slottype --model_checkpoint ${checkpoint} --n_epochs 15 --fewshot 0.01 --mode finetune

• --model_checkpoint: saved checkpoint of zero-shot model
• --fewshot: ratio of in-domain data, choose one from [0.01, 0.05, 0.1]

Full-shot

❱❱❱ python T5.py --train_batch_size 16 --GPU 8 --slot_lang slottype --n_epochs 15