From Eliza to Siri and Beyond

Abstract

Since Eliza, the first chatbot ever, developed in the 60s, researchers try to make machines understand (or mimic the understanding) of Natural Language input. Some conversational agents target small talk, while others are more task-oriented. However, from the earliest rule-based systems to the recent data-driven approaches, although many paths were explored with more or less success, we are not there yet. Rule-based systems require much manual work; data-driven systems require a lot of data. Domain adaptation is (again) a current hot-topic. The possibility to add emotions to the conversational agents’ responses, or to make their answers capture their “persona”, are some popular research topics. This paper explains why the task of Natural Language Understanding is so complicated, detailing the linguistic phenomena that lead to the main challenges. Then, the long walk in this field is surveyed, from the earlier systems to the current trends.