Virtual assistants such as the Google Assistant, Alexa, Siri, Cortana etc. help users accomplish tasks by providing a natural language interface to service providers (backends/APIs). For the democratization of such assistants, it is important to seamlessly support an ever-increasing number of services and APIs. In order to do so, a virtual assistant needs to have the following qualities:
- Scalable - Easy to support new APIs without requirement of skilled developers. Massively multidomain.
- Low Maintenance - Seamless addition of new APIs without retraining. Robust to addition of new entities in the API or changes in API's interface.
- Data Efficient - Massive parameter sharing. Reduced requirement of complex human-generated annotations.
This repository is a collection of research projects aimed at building a data-driven virtual assistant with the above mentioned qualities.
This is not an officially supported Google product. All resources in this repository are provided "AS IS" without any warranty, express or implied. Google disclaims all liability for any damages, direct or indirect, resulting from their use.
We created the Schema-Guided Dialogue Dataset to benchmark the above qualities of a virtual assistant. The dataset includes annotated multi-domain conversations involving 45 APIs spanning across 20 domains. The evaluation sets contain some domains and APIs which are not present in the training set in order to test the zero-shot generalization capability of the dialogue systems.
The schema-guided LU and DST model uses natural langauge description of slots and intents to obtain their embedded semantic representation. Then a single unified model is employed to make predictions conditioned on these embeddings. With the use of BERT, the model is capable of zero-shot generalization to APIs which are not present in the training set.
The baseline model for LU and DST was released along with the Schema-Guided Dialogue Dataset. The model is described in our paper and code is available here. The code will be moved to this repository soon.
The Schema-Guided State Tracking track in the 8th DSTC focussed on improving the baseline model for LU and DST. The participants were able to significantly improve the performance of the baseline model while reducing the gap in performance between seen and unseen APIs. More details about the competition and the submissions may be found in the overview paper.
Our paper
investigates three different apporaches to building a multi-domain NLG system
without any domain or API dependent parameters, by using a pre-trained
T5
model. The code may be found in the directory generation/.
@inproceedings{rastogi2019scalable,
title={{Towards Scalable Multi-domain Conversational Agents: The Schema-Guided Dialogue Dataset}},
author={Abhinav Rastogi and Xiaoxue Zang and Srinivas Sunkara and Raghav Gupta and Pranav Khaitan},
year={2019},
booktitle={{Proceedings of the AAAI Conference on Artificial Intelligence}},
url="https://doi.org/10.1609/aaai.v34i05.6394",
}
@inproceedings{kale-rastogi-2020-template,
title = "Template Guided Text Generation for Task Oriented Dialogue",
author = "Kale, Mihir and Rastogi, Abhinav",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.emnlp-main.527",
pages = "6505--6520",
}