Abstract
Traditional approaches to building natural language (NL) interfaces typicallyuse a semantic parser to parse the user command and convert it to a logicalform, which is then translated to an executable action in an application.However, it is still challenging for a semantic parser to correctly parsenatural language. For a different domain, the parser may need to be retrainedor tuned, and a new translator also needs to be written to convert the logicalforms to executable actions. In this work, we propose a novel and applicationindependent approach to building NL interfaces that does not need a semanticparser or a translator. It is based on natural language to natural languagematching and learning, where the representation of each action and each usercommand are both in natural language. To perform a user intended action, thesystem only needs to match the user command with the correct actionrepresentation, and then execute the corresponding action. The system alsointeractively learns new (paraphrased) commands for actions to expand theaction representations over time. Our experimental results show theeffectiveness of the proposed approach.