Bootstrapping Chatbots for Novel Domains

Abstract

We tackle the problem of automatically generating chatbots from Web API specifications using embedded natural language metadata, focusing on the intent classification subtask. One of the main challenges for such a use case comes from the lack of a sufficiently representative training sample for utterance classification, which hinders the traditional supervised model’s ability to generalize to unseen inputs. We apply several unsupervised and distantly-supervised techniques to refine the model’s representation and to augment its coverage. These include sentence similarity estimation in the embedding space and decision tree learner-based feature selection leveraging structural regularities in API specifications. In addition, we harness linguistic resources such as web corpora, PPDB, and WordNet. We use the resulting representation for prediction and perform a small-scale evaluation, where our approach compares favorably to the supervised baseline.