Automatic Canonical Utterance Generation for Task-Oriented Bots from API Specifications

With the mind-blowing development of REST (REpresentational State Transfer) APIs (Application Programming Interfaces), many applications have been designed to harness their potential. As such, bots have recently become interesting interfaces to connect humans to APIs. Supervised approaches for building bots rely upon a large set of user utterances paired with API methods. Collecting such pairs is typically done by obtaining initial utterances for a given API method and paraphrasing them to obtain new variations. However, existing approaches for generating initial utterances (e.g., creating sentence templates) do not scale and are domain-speci!c, making bots expensive to maintain. The automatic generation of initial utterances can be considered as a supervised translation task in which an API method is translated into an utterance. However, the key challenge is the lack of training data for training domain-independent models. In this paper, we propose API2CAN, a dataset containing 14,370 pairs of API methods and utterances. The dataset is built by processing a large number of public APIs. However, deep-learning-based approaches such as sequence-to-sequence models require larger sets of training samples (ideally millions of samples). To mitigate the absence of such large datasets, we formalize and de!ne resources in REST APIs, and we propose a delexicalization technique (by converting an API method and initial utterances to tagged sequences of resources) to let deep-learning-based approaches learn from such datasets.