Identifying Questions & Requests in Conversation

Abstract

In building an automated conversation agent, that attempts to converse with a human with as human-like as possible manner, we require the agent to identify which dialogue act, or class, the sentence belongs to. Determining the sentence class of a spoken phrase is helpful in building an intelligent companion, because without it the response may seem out of place. In written language, sentences can be classified into three classes: Declarative, Interrogative, and Imperative. These classes indicate which dialogue act the sentence belongs to. What our system does is take spoken text, which contains no punctuation, and classify the text into the three aforementioned classes. In conversation, the type of spoken text can indicate the type of required response. Our system is able to classify spoken text with 82% accuracy on our semi-automatically constructed dataset.