A critical review of state‐of‐the‐art chatbot designs and applications

Abstract

Chatbots are intelligent conversational agents that can interact with users through natural languages. As chatbots can perform a variety of tasks, many companies have committed numerous resources to develop and deploy chatbots to enhance various business processes. However, we lack an up‐to‐date critical review that thoroughly examines both state‐of‐the‐art technologies and innovative applications of chatbots. In this review, we not only critically analyze the various computational approaches used to develop state‐of‐the‐art chatbots, but also thoroughly review the usability and applications of chatbots for various business sectors. We also identify gaps in chatbot‐related studies and propose new research directions to address the shortcomings of existing studies and applications. Our review advances both academic research and practical business applications of state‐of‐the‐art chatbots. We provide guidance for practitioners to fully realize the business value of chatbots and assist in making sensible decisions related to the development and deployment of chatbots in various business contexts. Researchers interested in the design and development of chatbots can also gain useful insights from our critical review and identify fruitful research topics and future research directions based on the research gaps discussed herein. This article is categorized under: Technologies > Machine Learning Application Areas > Business and Industry

A Performance Comparison of Different Cloud-Based Natural Language Understanding Services for an Italian e-Learning Platform

“…A more complete and specific definition is provided in [2,3], which defines a chatbot as "an intelligent conversational system that is able to process the human language and formulate a response which has to be relevant to the user input sentence". Although the original idea of a chatbot comes from the Turing test, in which a computer program talks with humans trying to convince them that their interlocutor is not artificial [4], in current applications, chatbots are simply designed to provide information and satisfy humans' requests [5,6], providing them with a more user-friendly interface in comparison, for instance, to a static content search of a website [7]. In our context, the chatbot has to answer the students' most frequently asked questions, so as to significantly reduce the workload of tutors, who are the ones that usually assist the learners.…”

“…Focusing on the structure and implementation of the Application Core, the work in [6] presents a division of chatbots into several categories. The template-based chatbots select responses, computing the similarity between the user request and a predefined template using pattern matching techniques.…”

“…The corpus-based chatbots integrate knowledge engineering techniques to store information which can be used to satisfy the users' requests [10] or exploit word embedding techniques [11] for better computing the similarity between questions and answers [12]. Sequence-to-sequence (Seq2Seq) models, in which a deep learning model (such as a Recurrent Neural Network) elaborates the request and then automatically generates the answer, are defined by [6] as RNN-based chatbots. However, in our opinion, other Seq2Seq models based on the Transformer architecture [13] (such as Meena by Google [14]) can also be included in this category for their ability to generate an answer without a set of predefined questions.…”

“…Given the urgency to implement this kind of system, the company decided to use one of the cloud-based services that allows the fast development of a virtual assistant. Most of these services can be defined (according to the schema presented in [6]) as intent-based chatbots. The main activity that an intent-based chatbot has to perform is to recognize the intentions of the user, which can be expressed in many different ways.…”

“…Several works perform this kind of comparison, mainly from an architectural and implementation perspectives and simply showing the results of the intent recognition task [15][16][17], often considering an already existing application [18][19][20]; in this paper, we tackle several fundamental aspects that have to be considered in developing a chatbot using a cloud-based platform, such as the number of the training instances for each intention and its relation with the performance. As pointed out in [6], service-oriented chatbots (such as the one we implemented for Mega Italia Media) have their specific lexicon and require a customization process which can be based on a very low number of training instances. Moreover, since in terms of usability the response time has to be considered, we compared the response time of the different platforms in order to identify if there are relevant differences between the cloud-based platforms; in addition, we evaluate the performance also in terms of misspellings and errors.…”