This code is the official pytorch implementation of our paper: OSTOD: One-Step Task-Oriented Dialogue with Activated State and Retelling Response
Currently, with the advancement of large-scale pre-trained language models, the potential of conversational AI is gradually being realized, resulting in rapid development in related research. In particular, task-oriented dialogue systems have garnered significant attention due to their enormous potential for assisting people in achieving various goals, such as booking hotels, making restaurant reservations, and purchasing train tickets. Traditionally, task-oriented dialogue systems have been decomposed into a multi-step system consisting of four steps: spoken language understanding, dialogue state tracking, dialogue policy learning, and natural language generation. Recently, with the help of large-scale pre-trained language models, end-to-end task-oriented dialogue systems have integrated these steps into a single model, enabling joint optimization and avoiding error propagation. However, almost all previous end-to-end methods inevitably require predicting the intermediate result of the dialogue state for interaction with the database, and the dialogue state is often domain or task specific, which poses a significant challenge for generalization in previous task-oriented dialogue systems. To solve the problem, in this paper, we propose One-Step Task-Oriented Dialogue (OSTOD), which models task-oriented dialogue by synchronously generating activated states and retelling responses, where activated states refer to slot values that contribute to database access, and retelling responses are system responses that contain activated state information. Specifically, first, automatic methods are designed to build data containing activated states and retelling responses. Then, a joint generation model that synchronously predicts activated states and retelling responses in a single step is proposed for task-oriented dialogue modeling. Empirical results show that, on MultiWOZ 2.0 and MultiWOZ 2.1 datasets, our OSTOD model is comparable to the state-of-the-art baselines, but also has excellent generalization ability in few-shot learning and domain transfer.
- python >= 3.7
- pytorch >= 1.0
data: Data folder and contains data preprocessing
T5_generator: The generator to build data that contains retelling response
T5_filter: The filter to build data that contains retelling response
OSTOD: Our one step task-oriented dialogue model
domain_split_config.py, domain_split_data_loader.py and domain_split_generate_and_filtering.py: Retelling response generation and filtering using T5_generator and T5_filter
In the data folder, run preprocess_from_ubar.py and preprocess_for_our.py in turn to preprocess the dataset.
In the T5_generator folder, run T5_generate.py to train the generator model.
In the T5_filter folder, run T5_generate.py to train the filter model.
In the main folder, run domain_split_generate_and_filtering.py to build the data containing the retelling responses using T5_generator and T5_filter.
In the OSTOD folder, run train.py to train or test our OSTOD model, then run eval.py to calculate all metrics.