API2CAN: a dataset & service for canonical utterance generation for REST APIs.

Area of Science:

• Natural Language Processing
• Human-Computer Interaction
• Software Engineering

Background:

• Natural language interfaces (NLIs), such as chatbots, are increasingly used to interact with applications by executing underlying APIs.
• Supervised methods for building NLIs require extensive datasets of user utterances paired with corresponding APIs.
• A significant challenge is the scarcity of training data for developing domain-independent NLI models.

Purpose of the Study:

• To address the lack of training samples for domain-independent NLI models.
• To propose a novel dataset specifically designed for training supervised models to generate initial utterances for APIs.
• To facilitate the creation of more robust and versatile natural language interfaces.

Main Methods:

• A dataset of 14,370 API method-utterance pairs was automatically generated.
• API method descriptions were converted into user utterances.
• The dataset underwent manual cleaning to ensure high quality and usability.
• Accompanying microservices were developed to aid in sample collection.

Main Results:

• The creation of a large-scale, manually curated dataset for API-to-utterance generation.
• Development of supporting microservices to streamline the data collection process.
• The dataset enables training of models for generating diverse and accurate initial utterances for APIs.

Conclusions:

• The proposed dataset and accompanying tools significantly lower the barrier to entry for developing supervised natural language interfaces.
• This resource facilitates research and development in domain-independent NLI systems.
• Enables more efficient creation of high-quality training data for translating API methods into natural language utterances.