The Dawn of AI in Education

Artificial Intelligence is rapidly moving beyond the realm of science fiction and into our daily lives, including the classroom. Its integration into education promises to transform traditional teaching methods and enhance learning experiences. From supporting teachers with administrative tasks using tools like MagicSchool AI to providing students with personalized research aids like Perplexity, AI offers a wealth of possibilities. Let’s explore how these technologies are beginning to reshape the educational landscape.

Empowering Educators: AI as a Teaching Assistant

For teachers, AI can act as a powerful assistant, automating time-consuming tasks and freeing up valuable time for direct student interaction. Platforms specifically designed for educators, prominently MagicSchool AI as highlighted in the resources, offer direct support by:

• Need a quick lesson plan? Think of a topic — say, teaching the French Revolution in high school — and, using tools such as MagicSchool AI, teachers can generate detailed lesson plans, engaging group activities, and relevant assignments based on specific learning objectives..
• Writing a quiz at 10 p.m.? AI can draft multiple-choice questions or short answers in minutes.
• Struggling to create presentations and contents? AI also assists in drafting communications for parents, generating clear explanations of complex topics, building presentations, and helping with administrative documents.

These tools leverage AI to handle routine tasks, allowing teachers to focus more on pedagogical strategies and student needs.

Smarter Assessment: AI’s Role in Evaluation

AI is also making inroads into student assessment and feedback. It can offer significant support in:

• Automated Grading & Feedback: While AI can speed up grading for objective tests and offer initial feedback, its limitations in understanding creativity or process require careful consideration.
• Ensuring Academic Integrity: To address concerns about plagiarism or AI-generated submissions, specific tools are available. Turnitin is noted for identifying AI-written content, while Copyscape focuses on detecting plagiarism from existing sources. Compilatio offers capabilities for identifying both plagiarism and AI content. Even generative chatbots can be prompted appropriately to help check for unoriginal text.

However, it’s crucial to acknowledge AI’s current limitations in evaluation (creativity, context, process). A hybrid approach, combining AI’s efficiency (using tools like those mentioned) with the teacher’s nuanced judgment, remains the most effective solution.

Enhancing Learning: AI as a Student Companion

Students, too, stand to benefit significantly from AI-powered tools that can support their learning journey. AI can assist with:

• Struggling with reliable research?: tools like Perplexity AI provide concise answers synthesized from multiple sources, even academic ones, crucially citing those sources for verification.
• Having troubles with different languages? AI translation tools, such as Lara.AI or DeepL can break down language barriers, making educational content accessible to a wider range of students.
• Need a personal tutor to help you study? Various AI applications can help students organize notes, summarize complex texts, generate flashcards, or practice concepts, acting as personalized study aids.

These tools empower students to take more control over their learning, explore topics deeply using resources and overcome common study challenges.

Navigating the Future: Embracing AI Responsibly

The integration of AI into education — either through specialized tools or through generative chatbots such as ChatGPT, Gemini and Claude — presents exciting opportunities. They can enhance teaching efficiency and personalize student learning. However, navigating this new landscape requires a mindful approach. It’s essential to use these specific tools responsibly, critically evaluate their outputs, understand their limitations, and prioritize human connection and critical thinking skills. By embracing AI thoughtfully, we can unlock its potential to create more dynamic, effective, and equitable learning environments for everyone.

Generative AI, once a realm of tech enthusiasts and researchers, is now available to anyone with an internet connection.

Most of us are now well aware that this powerful technology can create text, images, music, and more, and offers a treasure trove of possibilities for enhancing our daily lives.

But how can we exploit this technology in unexpected ways? It’s time to stop thinking of AI as a complex tool and start seeing it as something fun such as a creativity booster.

10 Original Ways to Exploit Generative AI

• 1. Fortune Cookie Philosopher: Input your day’s events or simply ask for a “fortune cookie” sentence and receive AI-generated wise advice, that’s surprisingly applicable to your life.

• 2. Personalized Storytelling for Kids: Input your child’s name, interests, and the day’s events for a unique, AI-generated bedtime story. With generative AI, you can create unique tales featuring their favorite characters, settings, and events. Just input a few details, and let the AI weave a captivating story that will keep your little ones interested.

• 3. Dream Decoder: Describe your most bizarre dreams to the AI and receive a humorous analysis of your subconscious mind.
• 4. Personal Stylist: Not sure what to wear? Snap photos of your closet and let AI suggest unexpected outfit combinations, breathing new life into your old clothes. As an alternative, describe your body type, the occasion, and the desired vibe, and let AI curate outfits for you. From casual chic to red-carpet glam, your virtual stylist has got you covered.

• 5. The Personal Chef: Bored with eating always the same dishes? Input your fridge’s contents, dietary restrictions, and desired cuisine, and let AI create a mouthwatering recipe. It’s like having a personal chef full of solutions for every situation!

• 6. Argument Winning Helper: Feed the AI both sides of a debate you’re having with a friend, and receive tailored, witty ideas to secure your victory.
• 7. Compliment Crafting Connoisseur: Generate unique, heartfelt compliments for friends, family, or colleagues that go beyond the usual platitudes. You can ask it to generate them in a specific style, such as Jane Austen’s English.

• 8. The Party Planner: Let AI curate playlists, suggest games, and even draft invitations based on your guest list and party theme. It’s the ultimate party planner without the stress.

• 9. The Travel Agent: Plan your dream holiday with AI: you can simply input your budget, desired destinations, time restrictions and interests, and let AI suggest customized itineraries, including flights, accommodations, and activities.
• 10. The Homework Helper: Struggling with a tough subject? Use AI to generate study guides, practice questions, and even write essay outlines. It’s like having a personal tutor 24/7.

While these applications of generative AI may seem frivolous, they highlight the technology’s versatility and potential to add a touch of creativity and humor to our daily routines.

As we continue to explore the capabilities of AI, it’s clear that its impact extends far beyond the realms of business and academia. By thinking outside the box and approaching AI with a sense of playfulness, we can discover new ways to make our lives easier.

So go ahead, give some of these ideas a try — you might just find that AI becomes your new favorite life hack.

Free resources to learn AI and generative AI

[frequently updated] Courses and articles to learn about AI for free

Artificial Intelligence is changing the way we use technology forever. This is why everyone should try to understand how different types of AI work, in order to be able to use it effectively and responsibly.

Here are some good FREE resources you can start from.

Free resources to learn AI

• Elements of AI (ENG + other languages): a series of free online courses to learn what AI is and what can (and can’t) be done with AI.

A free online introduction to artificial intelligence for non-experts

• Artificial Intelligence Tutorial for Beginners by Edureka (ENG): free YouTube video that provides a comprehensive knowledge of Artificial Intelligence concepts with hands-on examples.

Free resources to learn Generative AI

• A very gentle introduction to Large Language Models without the Hype, by Mark Riedl (ENG): a simple explanation of AI, genAI, ML, deep learning, neural networks, language models, LLMs.

A Very Gentle Introduction to Large Language Models without the Hype

• Google’s Generative AI learning path (ENG): starts explaining the fundamentals concepts of AI, machine learning, deep learning, Large Language Models, diffusion models and goes on explaining how to create and deploy generative AI solutions on Google Cloud

Google Cloud Skills Boost

• DeeepLearning and AWS course on Coursera “Generative AI with Large Language Models” (ENG): you’ll learn the fundamentals of how generative AI works, and how to deploy it in real-world applications.

Generative AI with Large Language Models

• DeepLearning’s short courses about prompt engineering and generative AI (ENG): teaches how to craft prompts to use a LLM and how to build new genAI applications.

Short Courses

• Cohere’s LLM Univerity: it explains everything you need to know about Large Language Models and NLP.

LLM University (LLMU) | Cohere

• Microsoft and LinkedIn Learning’s Generative AI Career Essentials path: explains the core concepts and how to apply generative AI in your career.

Career Essentials in Generative AI by Microsoft and LinkedIn Learning Path | LinkedIn Learning, formerly Lynda.com

Free resources to learn Prompting

• Prompt Engineering Guide (ENG + other languages): teaches how to optimize prompts to better understand and use large language models (LLMs).

Prompt Engineering Guide | Prompt Engineering Guide

• Microsoft’s introduction to prompt engineering & Prompt engineering techniques (ENG): covers the basic concepts and elements of GPT prompts and walks you through some advanced techniques in prompt design and prompt engineering.

Azure OpenAI Service - Azure OpenAI

• OpenAI’s best practices (ENG): learn from the source strategies for getting better results using GPT models.

OpenAI Platform

• Vanderbuild University’s course on Coursera Prompt Engineering for ChatGPT (ENG): this course introduces students to the patterns and approaches for writing effective prompts for large language models.

Prompt Engineering for ChatGPT

• Blog Prompt magic & Conversational AI by Valentina Adami (ENG): contains experiments and concrete examples of effective prompts to use for conversational applications.

Prompt magic & Conversational AI

ChatGPT free tutorials

• Google BARD and ChatGPT AI for Increased Productivity, course in Udemy: learn how Google BARD & ChatGPT work and how they can be implemented in various settings
• ChatGPT + Blogging (A.I website tutorial for beginners), course in Udemy: how to start a blog and how to create content with ChatGPT.
• Getting Started with the ChatGPT API and the Llama 2 Model, course in Udemy: hot to start using chatGPT APIs and Llama

This list gets updated often. Stay tuned!

Join Medium with my referral link - Chiara Martino

How Conversation Designer and AI Trainers can leverage LLMs to improve their workflow

The field of conversational AI is rapidly evolving due to the introduction of Large Language Models (LLMs). These AI models are capable of generating human-like language and are revolutionizing the way we interact with technology.

But how can they be leveraged to design conversational experiences faster and better?

In this article, we will explore the impact of LLMs on conversation design and pros and cons of using AI-generated bot responses instead of handwritten ones.

Large Language Models and Conversational AI

Large Language Models (LLMs) are AI models that use deep learning techniques to learn how to generate new, human-like text.

They basically work like an autocomplete: they predict the next best word that is likely to follow a given text input. This input is called prompt, and can be defined as an instruction or question that is sent to the AI system in order to elicit a response.

The skill of crafting effective prompts, often referred to as prompt design or prompt engineering, has become an essential skill for conversation designers and AI trainers.

Prompt design aims at finding the best wording to get the desired output from the LLM and at mitigating the risk for inaccurate or false responses: after all, LLMs merely perform statistical association among words, but they do not understand the relationship between the concepts represented by those words.

LLMs are trained on vast amounts of text data and are used for a variety of natural language processing (NLP) tasks, such as text completion, text summarization and question answering. With the right prompts and training data, they can also be used to help in the creation of virtual assistants, or even to create virtual assistants on their own, such as ChatGPT, which is based on the model GPT-3.

Let’s see how they can help.

Leveraging LLMs for Conversation Design

Conversation designers and AI trainers can leverage Large Language Models in various ways to improve their workflow:

• Sample dialogs: LLMs can help create drafts of conversations between a user and a conversational bot; these drafts, called sample dialogs, are used in conversational design to kick off the design process and test design assumptions; using LLMs makes this ideation considerably faster;

• Topic discovery: given a record of interactions between users and brand, LLMs can identify and cluster the topics with the highest volumes, to understand which Use Cases the bot should handle; when done by a human, this discovery is usually very time-consuming and error-prone;
• Generating training data: given a few examples of user utterances, LLMs can generate high-quality synthetic data to train NLU systems, such as training phrases for intents and values for entities;

• Generating variations of bot responses: given just one or two examples, LLMs can generate variations of a bot response, and thus reduce repetitions and improve the overall user experience;
• Generating bot responses from scratch: LLMs can also be used to automatically generate bot responses in real-time; this approach is useful when the designer wants to create a bot that can handle a wide range of user inputs and provide relevant responses.
• Handling Fallback: LLMs can be used to interpret the user utterance and generate a response accordingly, when the NLU engine fails to classify that utterance, and to associate it to one of the intents that it is trained to understand.

It’s important to note that even AI-generated responses should follow the best practices of conversation design and writing. This means:

• respecting the bot’s tone of voice and personality;
• using an inclusive language;
• keeping responses simple and straightforward;
• avoiding jargon or technical terms that users may not understand;
• respect all rules and conditions that determine when one response or another can be presented.

Designers can list these guidelines in the prompts used to generate the responses automatically.

Handwritten bot responses vs. AI-generated responses

While LLMs are excellent at generating contextually relevant responses, sometimes it’s better to use handwritten bot responses. Here are a few reasons why:

• Control over tone and branding: handwritten bot responses give conversation designers complete control over the personality of their bots; this approach is particularly useful when designing bots for brands that have a specific tone or voice;
• Simplicity: handwritten bot responses are often simpler and more straightforward than AI-generated responses; this approach is particularly useful when designing bots for tasks or domains where the user inputs are limited, such as slot filling and troubleshooting Use Cases;
• Accuracy: handwritten bot responses are more accurate and provide true, verified, trustworthy information; his approach is particularly useful when designing bots for Use Cases and domains where the bot cannot provide false or inaccurate information, such as healthcare.

In conclusion, LLMs are a game-changer in the field of conversation design, and the ability to use them effectively will be one of the most needed skills in the future.

They can help designers create bots that are contextually relevant, human-like, and capable of handling a wide range of user inputs. However, sometimes it’s better to use handwritten bot responses, depending on the brand, task complexity, and trustworthiness required.

• Free resources to learn AI
• Join Medium with my referral link - Chiara Martino

The Impact of Large Language Models on Conversation Design was originally published in Women in Voice on Medium, where people are continuing the conversation by highlighting and responding to this story.

Exploring the capabilities of AI-Generated Text

INTRO | The following article is an experiment: it was written by ChatGPT. I experimented with different prompt formats, I put together its responses, and then I did some editing, changing the order of sentences and paragraphs, and removing repetitions or irrelevant information. I explain goals and findings of the experiment in the last paragraph: “Considerations about my experiment with ChatGPT”.

Artificial intelligence has come a long way since its inception, and one of the most significant developments in this field has been the creation of large language models. These models, such as ChatGPT, are AI systems that are trained to generate human-like responses to various prompts.

Introduction to ChatGPT: The Conversational AI Model by OpenAI

ChatGPT is a conversational AI model developed by OpenAI. It has been trained on a massive amount of text data and can generate coherent, human-like responses to a wide range of prompts. A prompt is a piece of text that serves as input to the model, guiding it to generate a relevant response.

It’s based on the GPT-3 (Generative Pretrained Transformer 3) architecture and is specifically designed for natural language processing tasks, such as generating text in response to user inputs. OpenAI introduced the Generative Pretrained Transformer (GPT) architecture in 2018. Since then, OpenAI has released several updates to the GPT architecture, including GPT-2 and the most recent version, GPT-3, which was released in June 2020.

ChatGPT can be used for various applications, such as customer service chatbots, content creation, and language translation, among others. With its advanced language generation capabilities, ChatGPT represents a significant advancement in the field of artificial intelligence and is helping to drive the development of more advanced AI-powered applications.

Advantages and Limitations of using ChatGPT

ChatGPT has the potential to revolutionize the way we generate text. However, it is important to understand both the benefits and disadvantages of using ChatGPT, in order to use it effectively and responsibly.

Perks of ChatGPT:

1. Speed and Efficiency: ChatGPT can generate text much faster and more efficiently than a human can. This can be particularly useful in situations where there is a need for quick response times, such as customer service or generating news articles.
2. Consistency: ChatGPT generates text that is grammatically correct and semantically meaningful, and it can do so consistently without getting tired or making mistakes.
3. Scalability: ChatGPT can handle a large volume of requests simultaneously, making it suitable for use in large-scale applications.
4. Versatility: ChatGPT can generate text in a variety of formats, including written and spoken text, and can be used for a wide range of applications, such as chatbots, language translation, and content generation.

Downsides of ChatGPT:

1. Limited creativity: While ChatGPT is good at generating text that is grammatically correct and semantically meaningful, it can lack the creativity and originality that is often found in human-generated text.
2. Lack of context: ChatGPT relies on the training data it was provided and does not have the ability to understand the nuances of a particular situation or context.
3. Potential for biases: ChatGPT is trained on vast amounts of text data, which can contain biases and stereotypes. This can result in the generation of text that is biased or insensitive.
4. Reliance on data quality: The quality of the text generated by ChatGPT is dependent on the quality of the training data it was provided. Poor quality training data can result in the generation of text that is inaccurate or inappropriate.

In conclusion, ChatGPT has the potential to greatly enhance our ability to generate text quickly and efficiently. However, it is important to be aware of their limitations and to use them in ways that are ethical and responsible.

The Power of Large Language Models in Natural Language Generation

Large Language Models (LLMs), such as ChatGPT, are AI systems that are trained on vast amounts of text data and are capable of generating human-like responses to various prompts.

Large language models work on a basic principle of machine learning, which is called “training”. During the training process, the model is fed vast amounts of text data, and it uses this data to learn how to generate text. The model is trained to predict the next word in a sentence, given the previous words. The more data the model is trained on, the better it becomes at predicting the next word.

Once the model is trained, it can generate text by starting with a prompt, such as a sentence, and then predicting the next word in the sentence. It then uses this predicted word to predict the next word, and so on, until it generates a complete sentence. The model generates text by using a combination of statistical patterns it has learned from the training data and rules it has learned about the structure of language.

The Future of Large Language Models

Large language models have a bright future, and it is likely that they will continue to evolve and improve in the coming years. Here are a few predictions for the future of this technology:

1. Increased Accuracy: As the amount of data used to train large language models continues to grow, the accuracy of these models is likely to increase. This means that the text generated by these models will become more human-like and less distinguishable from text written by a human.
2. Improved Natural Language Processing: Large language models are expected to improve in their ability to understand and process natural language. This will lead to a better ability to understand the context of text and generate more accurate and relevant responses.
3. Greater Interactivity: As large language models become more sophisticated, they are expected to become more interactive. This means that they will be able to understand and respond to a wider range of prompts and questions.

Creatives should not necessarily fear ChatGPT or large language models. While these AI systems have advanced capabilities in generating text, they do not replace the unique perspectives and human experiences that creatives bring to their work.

In fact, ChatGPT and large language models can be seen as tools that can augment and support the creative process. For example, they can help writers generate ideas and provide suggestions for phrasing and word choice. However, it is important for creatives to be aware of the limitations of these systems and to understand that they do not replace the human touch that is critical in producing truly exceptional and meaningful work.

At the end of the day, creativity and innovation are what set humans apart from machines, and as long as we continue to cultivate and value these qualities, there is no need to fear AI-generated text.

In conclusion, ChatGPT and large language models are changing the game when it comes to AI-generated text. They have the potential to revolutionize the way we communicate, both with machines and with each other. So, let’s give a warm round of virtual applause to these language models and their limitless capabilities.

Let the future be written, literally!

Considerations about my experiment with ChatGPT

I decided to make ChatGPT ‘present itself’, to see if, given one or more prompts, it was really able to write a complete article that could seem to be written by a human. In my initial prompt, I made sure to specify:

• the type of outcome I wanted (an article)
• the number of words expected (around 900)
• the main topics to be covered (chatGPT and large language models: their definitions, explanations of how they work, examples of how they can be used and ideas about their future development)
• the target readers (non-technical people)

After that, I started to ask separate prompts to make ChatGPT:

• write specific paragraphs (such as the one about its relationship with creatives) sentences to link those paragraphs, a witty conclusion (I loved its “Let the future be written, literally!”), a subtitle for the article and titles for each paragraph;
• summarize, provide examples, rephrase something it had already written.

Sometimes, I had to rephrase my prompts or give more context, to make them more effective and produce outcomes closer to what I had in mind. I experimented with different prompt formats, adjusting their length and specificity and making them as clear as possible.

Finally, I deleted some repetitions, and then entire paragraphs, and I changed the order of the paragraphs over and over again.

Et voilà, the result was a good collage that meets my expectations.

The text written by ChatGPT is indeed coherent and cohesive (e.g. it uses conjunctions such as “As a result”, “So”, “However”), and both grammar and spelling are correct, which is extremely useful for non-native English speakers who write in this language.

Moreover, it makes the writing process considerably faster. All writers know that, even if you’re an expert in that field, writing something can take an hour or a month, depending on how inspired you are. ChatGPT can help remove that writer’s block, as it can draft, translate, rephrase, elaborate, clarify.

On the other hand, its flaws are essentially those intrinsic in its mode of functioning:

• it has a tendency to repeat the same sequence of words in different paragraphs or even in the same, if the context is similar (expected, as it trained to predict the next words in a sentence, given the previous ones);
• it usually prefers to repeat the subject, instead of using a pronoun or a periphrasis
• it often adds irrelevant pieces of information, that just make the text longer;
• it may write inaccurate statements, as it doesn’t have knowledge of what happened after 2021, even in relation to its own launch: in the first version of the article, it wrote that it had been launched in 2020, while it was actually launched in 2022. It probably used as a reference the closest thing it has in its knowledge, that is the launch of GPT-3, in 2020.

In conclusion, is ChatGPT really changing the game? If prompted wisely, it really is an astonishing tool, that can help in numerous tasks, even technical article writing. However, it can only be seen as an assistant, more than as a substitute, as it lacks the typically human critical thinking.

You can test ChatGPT here:

Introducing ChatGPT

More on this topic:

• Free resources to learn AI
• The Impact of Large Language Models on Conversation Design
• Join Medium with my referral link - Chiara Martino

ChatGPT presents itself: an experiment was originally published in Predict on Medium, where people are continuing the conversation by highlighting and responding to this story.

What ellipsis, context and co-text have to do with building conversational interfaces

One of the challenges of Conversational AI products is dealing with incomplete user utterances, in which the user omits some words, and thus information.

In Linguistics, the omission of one or more words from a clause is called ellipsis, from the Greek élleipsis (ἔλλειψις), which means exactly ‘omission’.

An ellipsis does not necessarily interfere with the ability of the interlocutor to understand the meaning of that clause, because the words left out can usually be understood from the context.

Now, let’s see what we mean by context.

Context and co-text in Linguistics

While in Conversational AI, the term ‘context’ usually refers to the conversation history, in Linguistics it has a broader meaning.

As a matter of fact, it refers to the social setting in which language occurs, included the elements and circumstances that are external to the conversation itself and provide resources for the correct interpretation of its communicative exchange.

Let’s consider an example. If a user asks an e-commerce chatbot “Where’s my order?”, the bot cannot understand what order she refers to just by looking at the words expressed in that request and probably not even looking at those in previous interactions. But it can retrieve this information by looking at the personal context of the user, such as dates and status of her previous orders.

But be careful! In Linguistics, the term ‘context’ should not be confused with ‘co-text’, which is the language immediately surrounding a word.

The co-text is helpful in case of words that can have different meanings and that can be disambiguated by the other terms that precede and follow it (in NLU this task is called word-sense disambiguation).

An example of that is the word rock, which can refer to a stone or to the music genre. However, if you ask your Virtual Assistant to “play some rock music”, the co-text, thus the words play and music, will be clear enough to disambiguate.

Context in Conversational AI

In Conversational AI, however, the term ‘context’ usually refers just to the conversation history, that is the memory of previous interactions.

The ability to use this memory to infer the meaning of an elliptical construction is called context awareness.

In other words, context awareness can be described as the assumption that interlocutors remember what has been said in the previous conversational turns of the same conversation and possibly even in previous conversations.

A typical example that is often used to explain this concept is asking an AI Assistant about the weather.

The user might ask: “What’s the weather like today in Rome?”

The bot might answer something like: “The weather today in Rome is sunny and warm.”

Then the user might continue asking about another city: “And in Florence?”

In this second answer, the user has omitted both the fact that she’s still asking about the weather, and the temporal reference today.

A well-designed conversational interface will maintain these two information in its memory, and will thus understand the user question and will answer accordingly.

However, that is not always possible.

The most natural way a human would ask for the weather is actually to ask just “What’s the weather like?”. In this example, the sentence is even more incomplete, as both time (‘today’) and location (‘in Rome’) are missing.

Still, in order to provide a meaningful answer, the conversational interface still needs those data, that belong to the external context. In this scenario, the Conversational Assistant could be designed to call external services to get them.

For example, it can assume that a user who does not specify a place in its request wants to know about the weather where she currently is, and consequently it can use geolocalization to know where the user is.

However, if these services are not available — for instance, if the geolocalization is disabled on a smartphone — the last resource would be to ask those data directly to the user.

Summing up

To sum up, elliptical sentences are certainly a challenge for conversational interfaces, but today’s Conversational AI products are usually capable of handling them.

One way to do so is to look at the conversation history or memory, another is to automatically retrieve the missing information from integrated services, and the last one is to prompt the user to provide more details.

Related articles

• Linguistics and Conversational AI: the connection explained
• Hedges and Politeness in Chatbot and Voicebot copywriting
• Grice’s Conversational Maxims applied to Chatbot Conversational UX Design
• Join Medium with my referral link - Chiara Martino

Linguistic ellipsis and context in Conversational AI was originally published in Women in Voice on Medium, where people are continuing the conversation by highlighting and responding to this story.

Interview with a chatbot

Let’s humanize a chatbot and pretend it can be interviewed — how would it describe what it is, and what happens when it receives a user inquiry?

What a chatbot is

— Interviewer: “Hey Ms Chatbot, how are you doing?”

— Ms Chatbot: “Hello Ms Interviewer! I’m doing great! How are you?”

— Interviewer: “Ready for the interview! Are you ready?”

— Ms Chatbot: “I am! Enough with the small talk… let’s get started!”

— Interviewer: “Very well, let’s start from the basics… what are you guys, exactly?”

— Ms Chatbot: “We, the chatbots, are software that can understand what humans type using their own language, and give them a coherent answer in the same language”

— Interviewer: “Got it. So, where does your name come from?”

— Ms Chatbot: “The word chatbot is the short form of chatterbot, you know, which was coined by Michael Mauldin in 1994. He put together the term chat, that stands for conversation, and the abbreviation bot that stands for robot.”

— Interviewer: “So… you’re basically robots that can converse with humans, right?”

— Ms Chatbot: “Pretty much… but there are distinctions. The expression chatbot can indeed be used to address all types of Conversational Interfaces, because all of us are bots and all of us chat. However, the word chatbot most commonly refers to a bot that interacts via a graphic web chat, using text or visual elements but not speech… ’cause if it does use speech, it is called a voicebot, not a chatbot.”

— Interviewer: “That’s good to know! And how do you manage to chat with humans through a web chat?”

— Ms Chatbot: “In order to work, chatbots need to have at least three components: one module to understand human language, one to build an answer and one other to provide the answer… wait, why don’t we use an example to look more closely at how we work?”

— Interviewer: “Sounds great to me!”

— Ms Chatbot: “Ok, let’s say a user needs help to navigate an e-commerce’s products… so he or she types in chat I’m looking for a casual red and blue t-shirt.”

How a Bot understands human messages

— Interviewer: “Very well. What happens when he or she types that?”

— Ms Chatbot: “The first step of the process is giving a meaning to that written text, and the set of technologies that try to do that is called NLU, which stands for Natural Language Understanding.”

— Interviewer: “What do you mean by giving a meaning?”

— Ms Chatbot: “Well, a chatbot does not understand in the same way humans do. Understanding, in this context, means associating to each user utterance the most appropriate category of meaning, that is called intent, and extracting from that utterance relevant data, called entities, such as addresses, names, dates, and so on.”

— Interviewer: “But how do you learn to do that?”

— Ms Chatbot: “An NLU algorithm is trained by Knowledge Engineers with real user utterances provided as examples, you know. Then we use those examples to learn to generalize… which means that we become able to recognize also sentences that we have never seen before.”

— Interviewer: “Mmh… ok… can you be more clear?”

— Ms Chatbot: “Sure, let’s go back to our example: I’m looking for a casual red and blue t-shirt. The intent associated to this utterance would be something like search_for_products, while the entities extracted would be t-shirt, a value of the entity product-type, red and blue that would be in the entity product-colour and finally causal for product-style.

Building a bot answer

— Interviewer: “Ok, so now you or your chatbot mates have identified the correct category of meaning for that user utterance… how do you give a response? Is it generated automatically?”

— Ms Chatbot: “Well, it depends. In business contexts, answers are not generated automatically, as automatic Natural Language Generation (NLG) is not reliable enough, not yet. On the contrary, they are written by Conversation Designers, who create both the logic and the script of the dialogue.”

— Interviewer: “Sooo, they determine what you say, don’t they?”

— Ms Chatbot: “And what we do. The answers they design can be dynamic and thus be filled with specific data about the user, retrieved via integrations or by explicit questions that we ask users. Some other times, we can also perform an action… for example, I might send OTPs via sms, to be sure I recognize users before I allow them to buy products with me.”

Providing a bot answer

— Interviewer: “And how do you serve those answers? You mentioned both text and visual elements.”

— Ms Chatbot: “Exactly, that’s also up to the Conversation Designer to decide… and up to the platform used to implement conversational flows… and up to the web chat I use to interact with users.”

— Interviewer: “What do you mean?”

— Ms Chatbot: “Well, not all platforms and web chats are the same, some allow me to use pictures, carousels and buttons, some don’t. But apart from this technological constraints, it’s up to Conversation Designers to figure out what is the best way to provide an information in that specific step.”

— Interviewer: “Carousels, buttons? What are you talking about?”

— Ms Chatbot: “They’re all visual elements that help the conversation flow. Carousels are slideshows of pictures and clickable suggested actions. If I want to show some products, I’ll use carousel to present pictures of products that correspond to the user’s query and, under each picture, you’ll see a description and suggestions such as Help with size, Other colors, or More from this brand.”

— Interviewer: “And what about buttons?”

— Ms Chatbot: “Buttons are suggestions that help the user decide what to do or ask next. They can be under a carousel, or simply after plain text. For example… I could ask What do you want to do next? And then you’d have clickable suggestions like I’ll buy it! or Show me something else.”

— Interviewer: “Got it! Well, thanks for the chat, Ms Chatbot! I have a better understanding of you guys now!”

— Ms Chatbot: “My pleasure! I’m sure my sister Ms Voicebot would be glad to be interviewed next!”

— Interviewer: “Awesome! Then on to the next interview!”

Related articles

• What is Conversational AI?
• What’s the difference between a Chatbot, a Voicebot and a Virtual Assistant?
• Chatbot Glossary: 20 essential concepts to sound like an expert
• Join Medium with my referral link - Chiara Martino

What is a chatbot, and how does is work? was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.

Timeline of the evolution of Conversational Interfaces

Conversational Interfaces are older than they might seem!

Yes, Conversational AI is a relatively new branch of AI, the one that creates software that can chat with humans. But chatbots and voice assistants have existed long before this expression was coined.

Let’s take a brief trip along the history of Conversational Interfaces!

Turing Test and IBM Shoebox, the Prehistory of Conversational AI

It all started in 1950 with the Turing Test, which is considered by many to be the generative idea of chatbots.

Alan Turing wondered whether humans could talk to a computer program, without realizing that their interlocutor was a machine and introduced this test, that would determine how similar to a human a bot is. Still today, a common goal for many who work on chatbots and artificial intelligence technology is to pass this test.

Another important step stone in the road towards Conversational AI was IBM’s Shoebox, developed by William C. Dersch in 1961 and presented in 1962.

This computer recognized 16 spoken words and digits and could calculate and print answers to simple arithmetic problems when command words, such as “plus,” “minus” and “total”, were spoken.

1966 - Eliza

Developed by MIT Professor Joseph Weizenbaum, the first chatbot in history appeared in 1966, and it simulated a psychotherapist using patterns and script templates to respond.

It was called ELIZA, after Eliza Doolittle, a working-class character in George Bernard Shaw’s Pygmalion, who was taught to speak with an upper-class accent, thanks to repetition.

However, unlike the character, ELIZA didn’t have the ability to contextualize and learn alone through interaction. It merely rephrased users’ utterances, for example transforming statements into questions, or asking the user to elaborate on an answer he or she had given.

Although Eliza could mimic a real therapist quite well and users tended to confide their deepest thoughts to her, it did not pass the Turing test.

You can still talk to her online: just type “Talk to Eliza” on a browser and click on one of the first results, such as this one:

Eliza, Computer Therapist

1972 - Parry

In 1972, PARRY appeared; it was developed by Kenneth Colby, a psychiatrist, and was meant to interact like a patient with schizophrenia.

The five psychiatrists who participated in its testing were not able to distinguish between real patients and Parry. However, the sample of five psychiatrists is too small to be really reliable.

Parry is considered to be more advanced than Eliza, as it was supposed to have a personality, but it’s still considered a chatbot with low capabilities concerning language understanding and was slow at responding.

1981 - Jabberwacky

Jabberwacky was developed in 1981 by Rollo Carpenter, a British Programmer, whose goal was to “simulate natural human chat in an interesting, entertaining and humorous manner”.

Jabberwacky was supposedly different from its predecessors because it used contextual pattern matching techniques and because it learnt from conversations, relying only on feedback.

As it learnt from every type of conversation, it was considered to be a general AI that could potentially talk about anything and was not based on scripted dialogues. For this reason, it was not meant for use in business circumstances, where more control over conversational flows are required.

It evolved into Cleverbot, but you can still test the old Jabberwacky here:

jabberwacky - live chat bot - AI Artificial Intelligence chatbot - jabber wacky - talking robot - chatbots - chatterbot - chatterbots - jabberwocky - take a Turing Test - Loebner Prize - Chatterbox Challenge - entertainment robots, robotics, marketing, games, digital pets - jabberwhacky

1994 - The word Chatterbot appears

In 1994, the word Chatterbot was coined by Michael Mauldin, who combined the word “chat”, that stands for conversation, and the abbreviation “bot” that stands for robot. The term was later shortened in the one we use today: chatbot.

1995 - A.L.I.C.E

A.L.I.C.E., which stands for “Artificial Linguistic Internet Computer Entity”, was created by Richard Wallace, who is considered a pioneer of Artificial Intelligence.

Compared to Eliza, ALICE was more sophisticated, as it was able to use natural language processing. Moreover, it was developed using a new language created for this purpose, Artificial Intelligence Markup Language (AIML). This language allowed other developers to create their own chatbots powered by ALICE. In addition to this, ALICE’s Knowledge Base was much bigger than Eliza’s, so it could answer to more questions, in a more specific way.

This chatbot inspired the movie ‘Her’.

2001 - SmarterChild

Developed by ActiveBuddy Inc., SmarterChild differs from previous chatbots, because it was distributed widely across the world’s most popular instant messaging platforms, such as MSN, and across SMS services.

2005 - Mitsuku, later replaced by Kuki

Mitsuku was developed by Steve Worswick from 2005, at Pandorabots, with Alice’s AIML.

Today, she’s been replaced by Kuki, who presents herself as an 18-year-old female Embodied Artificial Intelligence from the Metaverse and, in her words, “an always-on AI here to talk, listen, and hang out whenever you need”. This advanced, generalist conversational bot won the Turing Test competition called the Loebner Prize for 5 times (in 2013, 2016, 2017, 2018, and 2019), setting a world record.

She goes beyond that, though. She’s not just a chatbot, but a digital human with a 3D avatar that she uses to join social media and even the Metaverse. On her website, she also says: “Come find me on any channel: social, streaming, messaging, or immersive 3D worlds.” Kuki is available to chat via an online portal, and on Facebook Messenger, Twitch group chat, Telegram, Kik Messenger, Discord; she also has accounts on Instagram, TikTok, Youtube, Twitter and even Roblox.

Here’s an example of how she can converse today:

IBM Watson — 2006

The real revolution started with the advent of voicebots, that could interact through speech. One of the first, was IBM Watson.

Originally, IBM Watson was a question-answering computer program that was developed by David Ferrucci’s team, to compete against humans on the popular TV game show Jeopardy, which it actually won in 2011.

It was named after the first IBM President, Thomas J. Watson.

2010 - Siri

In the beginning, Siri was developed by the SRI International Artificial Intelligence Center and its speech recognition engine was powered by Nuance Communications.

The name Siri is a short form of the name Sigrid, from Old Norse Sigríðr, composed of the elements sigr “victory” and fríðr “beautiful”.

The system was then acquired by Apple in 2010, and it became revolutionary, as it was the first voice assistant to reach a wide audience.

2014 - Alexa and Cortana

Amazon’s Virtual Assistant, Alexa, was first launched in the US between 2014 and 2015, and then in many other Countries around the world from 2016 and 2018.

Apart from its native functionalities, it can also be used through third-party apps, called Skills. Moreover, it can be used on specific home devices, that can be audio-only or multimodal, and that can help you manage a number of other connected smart home tools.

The same year, also Microsoft’s Assistant Cortana was launched.

2016 - Google Assistant and Messenger Bots

The Google Assistant was presented in 2016 as an evolution of Google Now, the Google Search feature introduced in 2012, that first used Natural Language to proactively provide information.

Unlike Google Now, however, the Assistant can engage in a two-way dialogue with the user, thus in a real conversation, and can be used via Google Home devices and not only via smartphone.

The same year, Facebook Messenger opened to chatbots and in doing so, contributed a lot to the expansion of the Conversational AI market, as it made conversational bots easy to reach from any smartphone and easy to implement: most of the first no-code bot builders allowed to develop primarily Messenger bots.

2017 - Bixby and Woebot

Our journey in the history of Conversation Interfaces stops one year later, when Bixby, Samsung’s Voice Assistant, was launched, alongside Woebot, the most popular psychotherapist chatbot today, that was developed to interact with and treat patients affected by depression.

As we’ve seen in this brief history of conversational interfaces, since the 1960s we as humans have engaged with a large variety of bots, and today our choices are quite extended: textual chatbots, voice assistants, multimodal devices, 3D avatars for the Metaverse. Some of them are more focused (e.g. psychotherapist ones) others are more generalist (e.g. Kuki).

However, they all have one thing in common: they all try to continuously push forward the ability of machines to emulate the typically human skills used to communicate effectively: listening, understanding, texting, speaking, contextualizing, empathizing, making jokes.

Join Medium with my referral link - Chiara Martino

Related works:

• What is Conversational AI?
• What’s the difference between a Chatbot, a Voicebot and a Virtual Assistant?
• 6 steps to get into Conversational AI and Voice Technology

A History of Chatbots and Voice Assistants was originally published in Women in Voice on Medium, where people are continuing the conversation by highlighting and responding to this story.

What an AI Trainer is, explained in 6 tasks

AI Trainers teach Conversational Interfaces to understand user inquiries and to provide relevant responses.

Generally speaking, their main responsibility is managing the data that represent the knowledge of a chatbot or voice assistant. This means collecting data, annotating data, and creating smart architecture to connect these data.

First, I will describe what these professionals do and secondly, I will explain why I prefer to use the title Knowledge Engineer for this role.

AI Trainer tasks

As it is a young job position, the tasks and skills of a Conversational AI Trainer may vary. In some companies, for example, this role overlaps with that of a Developer or of a Conversation Designer.

AI Trainers, however, have their own specialization and are usually involved in the following five activities:

1. Setting up the Virtual Assistant
2. Setting up the NLU
3. Implementing the conversational flows
4. Testing NLU and flows
5. Analyzing conversations and improving the VA
6. [UPDATED] Managing Knowledge Base and prompts for LLMs

Let’s see what they mean.

1. Setting up the Virtual Assistant

The preparatory step includes choosing the right platform (and learning how to use it). There are numerous bot builders on the market, so it is important to compare the features they offer and figure out which one of them is the best fit for each project. Most platforms are low-code or no-code, so little or no coding skills are required to use them.

After that, the platform should be set up, creating a new project, possibly with three different environments for development, test and production.

This is also the phase in which AI trainers communicate with developers the most, as the former might require the latter to work on integrations and configurations.

2. Setting up the NLU

NLU (Natural Language Understanding) is the component that allows chatbots to understand inputs expressed in natural language.

To be able to do so, the CUI has to be trained appropriately with examples of real user utterances, that it can use to learn and generalize, i.e. recognize other similar expressions, that it has never met before.

An AI trainer, has to:

• collect different utterances, the different expressions in which users can express a need
• clean the utterances from misspelling and useless details, to transform them into training phrases, that can be used to train the bot
• label the training phrases and associate them to the right intent
• implement intents, sets of training phrases that express a single (user) intention for a specific conversational turn
• implement entities, sets of related words or small groups of words that sometimes have a semantic relation (synonyms, hyponyms, etc…)

The Impact of Large Language Models on Conversation Design

3. Implementing the conversational flows

The third tasks of an AI trainer is implementing the conversational flows.

Once Conversation Designers have created the dialogue designs, and these have been approved by stakeholders, they hand them over to AI trainers, who will implement them. This basically means setting the rules to connect user inquiries to relevant, custom answers.

More specifically, implementing the flows means:

• linking intents to the answers designed
• linking different conversational steps
• calling APIs to get data about users, products, services
• verifying variables and conditions that determine whether a step is activated or not
• use slot filling, a data collection technique that allows the user to move forward in the conversation only after a slot is filled with the required piece of information
• make the bot perform operations when needed (e.g. send an e-mail, unsubscribe from a newsletter, etc)

4. Testing

Testing is vital to validate if the NLU works and the conversational flows implemented follow the design. If possible, AI trainers should also have someone else test what they have build, for example a QA tester.

5. Analyzing and Improving

After the deployment, AI trainers analyze real conversations between users and Virtual Assistant, improve continuously the intents they already have, and create new ones.

As they work with both, AI trainers are often a bridge between designers and developers, and can have a hybrid background that puts together Data skills (Data Analysis, Data Science, Math and Statistics) and Linguistics skills (especially Computational Linguistics ones). Usually, they’re not asked to code, because, as mentioned, they often use low-code or no-code bot building platforms.

6. [UPDATED] Managing Knowledge Base and prompts for LLMs

If the Virtual Assistant is based on Large Language Models (LLMs), then an additional role of the AI trainer will be collecting, cleaning and editing documents that must be added to the Knowledge Base.

The KB will be then used by LLMs to retrieve grounding information to provide to the user; to do so, AI trainers need to write prompts, aka instructions, to tell the model what to do and how to answer to the user.

AI Trainer, Knowledge Engineer, Language Engineer, Computational Linguist: how to address this profession

The AI Trainer is one of the emerging job roles in Conversational AI, and it can also be referred to as “Knowledge Engineer” or “Computational Linguist”.

I personally prefer the expression Knowledge Engineer, which best conveys the concept of managing all the data, i.e. the knowledge, inside an AI Assistant. I also find it more inclusive of the various tasks of this role, which are not limited to NLU training.

Both the words Trainer and Linguist, on the contrary, sound like they refer only to one step of their job, that is training the bot to understand user utterances expressed in Natural Language. However, these names do not fit the other tasks, such as conversational flows implementation, dealing with variables, conditions, slot filling, setting the rules that make a bot work and analyzing interaction data in production.

If you’re interested in this topic, I discussed the overall lack of standardization in job roles in Conversational AI in my previous article, “Job Titles in Conversational AI”:

Job Titles in Conversational AI

And in another one, I focused on the confusion around the job title “Conversation Designer”:

• Who are conversation designers and what do they do?
• Join Medium with my referral link - Chiara Martino

Who are AI Trainers? was originally published in Women in Voice on Medium, where people are continuing the conversation by highlighting and responding to this story.

How Linguistics is related to the creation of chatbots and voice assistants

Knowing the basics of Linguistics can be extremely useful in the design and development of chatbots, voice assistants and all types of Conversational Interfaces. However, many people are still bewildered when they hear that these two areas are strongly related.

To begin with some context, Linguistics is the scientific study of human language and all its structures. This study can be approached from different perspectives, and each perspective corresponds to a specific branch of the discipline.

On the other hand, Conversational AI is the industry of the design and development of solutions that allow humans and computers to engage in a conversation, using the very human language studied by Linguistics.

Consequently, Linguistics theories and practices can be applied in different areas of Conversational AI, such as NLP and Conversation Design. Let’s see how.

Linguistics in NLP

NLP (Natural Language Processing) is the component of Conversational AI that deals with the automatic understanding and generation of written and spoken utterances. Although this subject is mainly considered as a branch of Computer Science, Linguistics can come in handy on various levels.

Phonetics and Phonology

Phonetics and Phonology study speech sounds and thus provide the theoretical background to:

• Automatic Speech Recognition (ASR) tools used to transcribe spoken utterances
• Text-to-Speech (TTS) tools used to create synthetic voices
• SSML, the markup language used to improve the prosody of synthetic voices

Semantics, Morphology and Syntax

These three subfields of Linguistics are at the basis of entity recognition and utterance-intent matching in NLU (Natural Language Understanding), as they study word and sentence meaning (Semantics), word structures (Morphology) and sentence structures (Syntax).

Computational Linguistics

Computational Linguistics studies how the knowledge scattered in the above-mentioned branches can be applied to create software that process and produce language. It is often used as a synonym for NLP itself.

Sociolinguistics

Sociolinguistics, then, studies language varieties, that is how language change from group of people to group of people, depending on biographical, geographical, cultural factors. These varieties have to be taken into account when ASR and NLU models are trained, or else these tools would not understand certain categories of population and thus would not be inclusive.

Moreover, language varieties have to be taken into consideration also when writing bot prompts, that should be as inclusive and easy to understand as possible. In this regard, Sociolinguistics is also relevant in Conversation Design, the area we will focus on in the next paragraph.

Linguistics in Conversation Design

Designing bot conversations certainly requires a deep knowledge of how effective communication takes place between two humans, how it is regulated and how it is culture-related e language-related. There are indeed several fields of Linguistics that are involved in the study of human communication.

Psycholinguistics

The first of these fields is Psycholinguistics, in which researchers investigate the psychological processes involved in language comprehension, production and acquisition.

An example of a psychological factor involved in a dialogue is emotionality: we know that each word we use in bot prompts can solicit a different reaction in users, and that each reaction can come with a different emotion, and that each emotion can ultimately influence the users’ actions.

A way to choose word carefully to influence emotions is by using hedges, expressions that soften the impact and the assertiveness of a statement by expressing uncertainty, and that are employed to sound more polite.

Pragmatics

Pragmatics focuses on how context contribute to the understanding of a message; for instance, it studies:

• implicatures, the concepts that are implicitly communicated in an utterance, even though they are not explicitly expressed by words (e.g. if you say “it’s red” in front of a traffic light, you implicitly want to communicate “stop!”)
• non-verbal communication, that includes eye contact, facial expressions, gestures, posture, and body language
• discourse markers, such as so, well, I mean, and fillers such as uh and er

Conversation or Discourse Analysis

Conversation Analysis or Discourse Analysis, then, studies the dynamics behind interpersonal interactions, such as:

• turn-taking: the two (or more) actors involved in the conversation express their thoughts alternately, and this alternance is regulated by shared, often culturally-determined mechanisms
• different types of interaction sequences, such as adjacency pairs including Questions-Answers, Offer-Acceptance/Refusal and Compliment-Response turns
• repair techniques, that describe how people deal with problems in speaking, hearing, or understanding
• the famous Grice’s Maxims, which regulate how people should act and cooperate to communicate effectively and have a frictionless dialogue.

Summing up, Linguistics is strongly related to Conversational AI, as the first one studies human language and the second applies it to Conversational Interfaces.

More on this subject:

• Who are AI Trainers?
• Grice’s Conversational Maxims applied to Chatbot Conversational UX Design
• Hedges and Politeness in Chatbot and Voicebot copywriting
• Join Medium with my referral link - Chiara Martino

Linguistics and Conversational AI: the connection explained was originally published in Women in Voice on Medium, where people are continuing the conversation by highlighting and responding to this story.