eduzhai > Applied Sciences > Engineering >

Automatic Discovery of Novel Intents & Domains from Text Utterances

  • Save

... pages left unread,continue reading

Document pages: 11 pages

Abstract: One of the primary tasks in Natural Language Understanding (NLU) is torecognize the intents as well as domains of users spoken and written languageutterances. Most existing research formulates this as a supervisedclassification problem with a closed-world assumption, i.e. the domains orintents to be identified are pre-defined or known beforehand. Real-worldapplications however increasingly encounter dynamic, rapidly evolvingenvironments with newly emerging intents and domains, about which noinformation is known during model training. We propose a novel framework,ADVIN, to automatically discover novel domains and intents from large volumesof unlabeled data. We first employ an open classification model to identify allutterances potentially consisting of a novel intent. Next, we build a knowledgetransfer component with a pairwise margin loss function. It learnsdiscriminative deep features to group together utterances and discover multiplelatent intent categories within them in an unsupervised manner. We finallyhierarchically link mutually related intents into domains, forming anintent-domain taxonomy. ADVIN significantly outperforms baselines on threebenchmark datasets, and real user utterances from a commercial voice-poweredagent.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×