UNSUPERVISED INDUCTION OF USER INTENTS FROM CONVERSATIONAL CUSTOMER SERVICE CORPORA

A methodology and system are presented for inducing user intent in a corpus and storing this intent in an intent library. To accurately detect intent, the corpus is first cleaned of nonsensical words and symbols and then syntactically analyzed to extract words and dependencies between them, which are then semantically analyzed to select keywords that are indicative of intent, and map the keywords to ordered broad semantic categories of the types of action, modifier and object. Keywords are then converted into embedding vectors whose dimensions are reduced and clustered according to category and order. Relations are calculated for the clusters across the semantic categories and intent is then calculated with the help of intent templates and word dictionaries.