While some may confuse chatbots with Conversational AI, one must realize that both possess a slight difference in their abilities. Where the latter is based on Cognitive Architecture, AGI orientation, and is generic, the former is based on Machine Learning, Deep Learning techniques. Srini Pagidyala, Co-Founder of Aigo.ai in one of his LinkedIn blogs defines Chatbot as a narrow or weak AI-oriented, purpose-driven technology.

Srini believes that Conversational AI can serve as the Universal User Interface by Humanizing Interactions with machines and systems using natural language abilities while 'narrow or weak AI' provides the inputs to the Conversational AI. The two technologies are complementary and when they are combined effectively they can enhance customer experience and add significant value to both the customer & the company.

Chatbots are believed to be the early development stage of Conversational AI. How? Let's dive into it.

Evolution Timeline

In the pursuit to adopt more humane features, not only menial tasks but modern chatbot tends to be highly specialized. Chatbots have evolved magnificently to become more diverse and creative with a combination of AI. But how did AI capabilities like NLP aided the evolution?

According to a Medium report, in the early stages of chatbot development, core NLP methods were used to design them as machine learning wasn't exactly viable then. Slowly machine learning methods came into effect to channel more data and code. Then came google's 2018 groundbreaking paper BERT, helping researchers transcend all research. One should understand the fact that unless the initial brute force and core methods were utilized there would have never been enough data for the best of ML algorithms today.

Across the decades chatbots have evolved with great abilities. Here is the timeline.

ELIZA

(1964–1966)

Eliza is one of the first natural language processing computer program created in 1964 by Joseph Weizenbaum. It was developed at the MIT Artificial Intelligence Laboratory. Its main intention was to demonstrate the superficiality of communication between humans and machines. It rose to fame while psychiatric patients perceived it to be human. Eliza simulated the chatbot experience by using a "pattern matching" and substitution methodology. This gave users the illusion of understanding on the part of the program but had no built-in framework for contextualizing events. Eliza interacted by provided "scripts". They were written originally in MAD-Slip (a programming language). It allowed ELIZA to process user inputs and engage in discourse following the rules and directions of the script. It was one of the first programs capable of attempting the Turing test.

Cleverbot

(1997–1998)

Cleverbot, created by British AI scientist Rollo Carpenter. It is a chatterbot web app that uses artificial intelligence (AI) to strike conversations with humans. It approaches the technique with core natural language processing and fuzzy logic. Fuzzy logic is utilized to tackle a million records stored and their utilization in a heuristic approach. With over 279 million interactions, about 3–4% of the data it has already accumulated Cleverbot is now taking the next big step and aiming to improve its efficiency by implementing ML techniques. It rose to fame after being featured in the popular creepypasta ARG web serial Ben Drowned by Alexander D. Hall.

Mitsuku 

(2002)

Mitsuku is a web-based chatbot developed by Pandorabots was awarded the annual Loebner Prize twice — in 2013 and 2016 — for being the most human-like chatbot around. Stamped as a "virtual friend", Mitsuku can answer questions, play games and do tricks at the user's request, and is capable of basic reasoning. It is also available on Kik Messenger. The underlying development was done by Steve Worswick for a good 13 years as he was frustrated with his I.T support job. The codebase is around 350,000. It is based on the AI/ML technology which consists of pattern and template elements.

Rose

(2011)

Rose is an award-winning chatbot created by Brillig Understanding, Inc. Brillig has based Rose on the average teenage girl. They have a claim of giving the chatterbot its own personality but one understands on usage that it is just set on a code basis. The main objective here is a starting base for chatbots to develop in a way where the chatbot displays its own emotion and thought process. The program is based majorly on powerful pattern matching aimed at detecting meaning and a simple rule layout combined with C-style general scripting.

Xiaoice

(2014)

Xiaoice is the AI system developed by Microsoft STCA in 2014 based on an emotional computing framework. XiaoIce is uniquely designed as an AI companion by forming an emotional connection to satisfy the human need for communication, affection, and social belonging. Intelligent quotient (IQ) and emotional quotient (EQ) are both considered in the system design. The chat experience is based on Markov Decision Processes (MDPs) which optimize XiaoIce for long-term user engagement. The established metric is Conversation-turns Per Session (CPS). With a huge user base, it is backed by none other Bill Gates himself. He has gone so far to claim that "she's gotten good enough at sensing a user's emotional state that she can even help with relationship breakups". Xiaoice is "much more" than just a chatbot as described by Microsoft.

Melody

(2015)

Melody's objective is to help both doctors and patients. By focusing on the medical assistant space, they've built a conversational bot that can give highly-customized and situation-appropriate responses to a patient's query. Melody is designed to save time but also serves as a sort of stop-gap solution for the worldwide issue of doctor shortages. The project is headed by none other than the deep learning pioneer Andrew Ng. The parent company is Baidu. Melody is based on advanced deep learning and NLP technologies. It is said that it continues to learn through more usage which means its an online learning system where the weights are constantly changed on usage. Sensely, HealthTap, and Koko are examples of other chatbots that focus on the same goal. Melody is currently available on the Baidu's Doctor App, accessible only in China.

DialoGPT: Toward Human-Quality Conversational Response Generation via Large-Scale Pretraining

(2019, November)

DialoGPT is a joint project between MSR AI and Microsoft Dynamics 365 AI Research team to develop state of the art chatbot systems. The project provides a foundation for building versatile open-domain chatbots that can deliver engaging and natural conversational responses across a variety of conversational topics, tasks, and information requests, without resorting to heavy hand-crafting. It is trained on data gathered from 147 million Reddit comment chains. The main code is utilized on the base of hugging face PyTorch transformers.

Meena

(January 28, 2020)

Meena is a chatbot that learns to respond sensibly to a given conversational context. Google has a fancy term for it which is the "neural conversational model". The training objective is to minimize perplexity, the uncertainty of predicting the next token (in this case, the next word in a conversation). At the core lies the Evolved Transformer seq2seq architecture, a Transformer architecture discovered by evolutionary neural architecture search to improve its main objective, the perplexity. It has around 2.6 billion parameters that have been trained on 2.5 TB of information. Until the recent announcement of Facebook's Blender, it was the state of the art system and was ahead of the entire race with its high SSA scores. The main aim of Meena is to address the critical flaw in chatbots of them not making sense. Google plans to improve the model by reducing its unnecessary biases which are an aim in making the model explainable. For the above reason, there is no current model that can be tested but the paper can be found here.

Blender

(April 29, 2020)

Blender, Facebook's latest chatbot in collaboration with ParlAI, is named for its ability to merge multiple conversational skills at once. The chatbot is built from upon 9.4 billion parameters and trained using 1.5 billion examples of conversation, making it so large that it had to be broken up into pieces in order to handle larger sets of data. The AI uses what Facebook calls Blended Skill Talk (BST) to merge various chatbot abilities. Though there is a large area for improvement, around 49% of the judges chose blender over humans and 3 quarters of the panel chose it over Meena! The main objective of this chatbot was to target open-domain conversations while maintaining empathy, knowledge, and personality. The paper and code can be found in the below link. Facebook has kept all the data as open source in hopes of someone improving the efficiency of the model.