Accounting for Named Entities in Intent Recognition from Short Chats

The operational cost of call centres accounts for a large part of the total spending of any modern organization. As a consequence, the automated conversation agent powered by Artificial Intelligence (AI) through Natural Language Processing (NLP) alternative has gained major attraction over the past years. Efforts to achieve such level of automation generally rely on predefined business intents or intents, which are in turn tightly related to business entities or processes that they represent. As the success of the automated conversation agent fully relies on its ability to accurately recognise user intents, a good automated agent will be the one that recognises intents it is meant to recognise. In light of the strong relationship that exists between business entities and business intents or entities and intents in general, we propose two approaches for accounting for named entities in the task of intents recognition from short chats. The first approach relies on Bi-Normal Separation (BNS) to weight term features that are named entities more than other features, whereas, the second approach takes advantage of word embedding to encode the relationship between entities and chats. Evaluation of proposed methodologies, on a data set composed of one to one conversations between human actors, suggests that accounting for named entities improves the performance of the intents recognition task.