Democratizing AI: Creating Cognitive AI Assistants with No Coding

Title: Democratizing AI: Creating Cognitive AI Assistants with No Coding
Date: Thursday, April 13, 2021, 1:00 PM ET/10:00 AM PT
Duration: 1 hr

SPEAKER:
Michelle Zhou, Juji, Inc.; ACM Distinguished Member

MODERATOR:
Wenxi Chen, AI Software Engineer, Juji, Inc.

TechTalk Registration

More Resources
Build Better Chatbots: A Complete Guide to Getting Started with Chatbots (Skillsoft book, free for ACM Members)
Developing Enterprise Chatbots: Learning Linguistic Structures (Skillsoft book, free for ACM Members)
Exploring Chatbots (Skillsoft course, free for ACM Members)
Understanding Bots: Chatbot Architecture (Skillsoft course, free for ACM Members)
UI/UX for Chatbot and Voice Interface AIs (Skillsoft course, free for ACM Members)

Question 1: What is books of materials to create cognitive AI bots?

Answer: You can check out the documentation on our website:Juji Documentation A set of scientific publications (https://juji.io/publications/) by us or others may also serve as a good resource on the creation of cognitive AI.

Question 2: Mostly we see in chatbots the questions are pre-programmed there is no actual intelligence going on. How to create something that can be intelligent enough so that it can tackle new questions?

Answer: Depending on the training models used behind the scene. If a chatbot uses only keywords to process user questions or requires a user input matches with training samples exactly, the chatbot’s capability is very limited. It would be the best to use a combination of data-driven (e.g., deep learning models) and symbolic AI approaches (e.g., Finite State Machines) to match a user’s input with trained models. This way even a user input is not an exact match, the chatbot can still answer the questions. Even if it cannot answer, it will make an “educated guess” to help users. For example, the training sample is “how long will I receive my refund”, the chatbot could recognize that a new user input “when will I get my refund” asks the same thing and then answers the user’s question accordingly. However, there are always questions that a chatbot cannot answer, e.g., if a chatbot does not know “how much will I receive my refund” – in such a case, a cognitive AI should notify a human admin who can then teach the chatbot new knowledge. Check out these two article to read more about what AI can do (https://juji.io/blog/where-is-artificial-intelligence-ai-how-to-make-a-chatbot-smart/) and how to create real AI chatbots to answer questions (https://juji.io/blog/a-step-to-step-guide-to-customer-service-chatbots-with-nlp-no-coding-required/).

Question 3: Is it possible to deploy a chatbot built in Juji as Alexa skill or Google Action?

Answer: I suppose you should be able to do so. Just give it a tr by signing up on Juji and it’s free to try: juji.io/signup

Question 4: Can AI read people’s mind by reading their facial and body gestures?

Answer: Yes, if you are very good at reading people :slight_smile: People’s facial expressions and body gestures are in fact good cues to read about a person. Such cues can be integrated with one’s communication text to gain a deeper understanding of a person. Currently, it is just that the technologies are not mature enough to parse such cues reliably.

Question 5: How are the sentences generated, random words, or just retrieved from a corpus? Or other ways?

Answer: Just like processing natural language expressions, our system uses various ways to generate chatbot responses, e.g., pattern-based, template-based, and algorithm-based.

Question 6: Can the bot keep track of the context of a conversation? For example, if we are talking about physics and I ask about entropy, will it figure out that I am talking about thermodynamic entropy, or will it think I am talking about Shannon entropy from information theory?

Answer: Yes! Cognitive AI assistants must understand context. As mentioned in my talk, Juji’s chatbots converse with a user by topics. The beauty of using topic-based conversation is to use the topic context to drive and manage a conversation. Using your example, you can easily define two topics, one on physics and the other on information theory. When a person asks about entropy in a particular context, the chatbot should know “which entropy” is referred to and answers accordingly. If you watched our talk, we showed that the exactly same user input “what about you” invoked very different chatbot responses in different contexts (e.g., in the context of favorite movie vs. hobby).

Question 7: Conversation w/Juji sounds more like a scenario-based one; it may not even require intelligence beyond DB access. It sounds that Juji would succeed in a limited domain, such as a company’s products & processes. I wonder what Juji would respond to “What’s up, buddy?”

Answer: Juji actually has been used for many applications beyond product/process access, including user or job interview, often involving open-ended questions, such as “what’s the best of you” You are most welcome to chat w/ Juji and ask “What’s up, buddy?” :slight_smile:

Question 8: Could you give us recommendations for CS staring in the Democratzing?

Answer: See answer below - similar question.

Question 9: Could you give us any recommendations for computer science researchers to get started in the field of democratizing AI? Especially in other fields of AI such as Machine Learning or Multi-agent systems applications.

Answer: Check out the recent publications in the following ACM conferences and journals on explainable AI as well as advances in intelligent human-computer interactions:
https://iui.acm.org/2021/
https://dl.acm.org/journal/tiis

Also see my interview with ACM on the last question:
People of ACM - Michelle Zhou"

Question 10: Learning is more than just spoken language, but non-verbal in English may contain more than 60% of information being communicated. AI systems will not, looking next 5 - 10 years, be able to extract meaningful non-verbal information because it is so varied, and ambiguous at times. Comments?

Answer: You are absolutely right: learning about a person or a person in a conversation is more than just about what is being said, it’s also about non-verbal cues. Currently Juji has not considered such cues because language alone already poses a huge challenge. As technology advances, hopefully Juji will be able to use all the cues in the future.

Question 11: Can you share Tutorial for Creating Cognitive Assistant?

Answer: Check out our tutorials on Youtube: https://www.youtube.com/c/hellojuji and our online documentation: Juji Documentation

Question 12: How is useful this cognitive AI assistant?

Answer: You need to ask our clients :slight_smile:

Question 13: Will the speaker talk about the technology behind this chatbot?

Answer: "We tried to give a high-level overview in this talk. If you are interested in knowing more about what’s behind the scene, check out our documentation(Juji Documentation) or publications (https://juji.io/publications/)

Question 14: Hello, very interesting what you show. I would like to ask, does the bot makes its own responses and how?

Answer: A bot responds with what humans put into them. But it needs to automatically discern when and how to respond with such knowledge depending how a conversation unfolds.

Question 15: Even after training Alexa more than 50 times my voice, Alexa still indicates, “your voice is new”. Professional cons are very compelling as they have learned how to use their non-verbal to help sell a story. AI assistant cannot, at this time, learn non-verbal.

Answer: You have to ask Amazon why Alexa always indicates “your voice is new” after training it over 50 times :slight_smile: You’re right, Juji AI assistants currently don’t use non-verbal cues yet, hopefully in the future, they will.

Question 16: It’s look like FSA (Finite State Automata)…is it ?

Answer: Yes, Finite State Automata - also known as finite state machines - we use it to extract certain patterns and recognize certain entities.

Question 17: Is English the only language used for conversations?

Answer: See answer below - similar question

Question 18: Do you support multi language bots, both in topic discovery and ML part? Or English only?

Answer: See answer below - similar question

Question 19: What about different languages, how different is learning or model?

Answer: Actually our AI architecture and models are language independent. Depending on the requirements, in most cases, it should be easy to support additional languages: https://juji.io/ai-chatbot-api/#multilingual-bot

Question 20: I have used Micorsoft cognitive services (LUIS). How is what you are doing different from that and the equivalent service offered by google or amazon? I can show you what we have done with LUIS in a zoom meeting any time, if you like. thanks.

Answer: See answer below - similar question

Question 21: Could you comment on the differences between your tool for creating chatbots and the others on the market right now, such as the Watson Assistant and the Microsoft Bot Framework from Azure?

Answer: The main differences are three: (1) Juji supports a much richer set of conversational capabilities, including supporting of mixed-initiative (two-way) conversations and managing highly complex, context-sensitive conversations; (2) Juji supports deeply personalized conversations during which a Juji AI chatbot can infer a user’s unspoken needs and wants and personalize each engagement; and (3) Juji enables non-IT professionals to create an AI assistant with superior conversation capabilities as described in (1) and (2), with no coding, 20-50x faster than using any of other platforms.

Question 22: Since the training requires a lot of relevant sample data which may not be available or of low quality, is JuJi doing any work in the area of automatically generating training data?

Answer: Yes, Juji leverages trained models as well as auto-generates many training data - that’s why you saw in our demo, a chatbot designer/developer often provided just a few examples.

Question 23: How good is the training when the sample data is faulty?

Answer: Training results depend on the sample data. If sample data is faulty, the training will result in whatever the sample data provides.

Question 24: Are your tools open source?

Answer: No, it is not open-sourced yet.

Question 25: Can a chatbox be embedded into a robot?

Answer: See answer below - similar question.

Question 26: Have you ever used your API to interface with robots, to have an AI physical robot instead of an on-screen chatbot?

Answer: Although we haven’t done it, we have actually thought about using our AI to drive physical robots’ interactions with their users. We don’t have access to a robot, if you have and are interested in collaborating with us, feel free to reach out: hello@juji.io

Question 27: The question disappeared before I finished typing it. I was trying to ask whether the bot can draw on knowledge in an OWL ontology or RDF triple store.

Answer: Where are you trying this? Feel free to reach out to us on forum.juji.io to ask questions and report bugs.

Question 28: What would happen if the user answered in Yen or Pounds?

Answer: Since the topic used in this case expects “dollar” as the currency. If someone answers “Yen” or “Pounds”, the chatbot will inform you what it expects. Similar topics can be used to handle different currencies.

Question 29: The cognitive ai engine is answering the same way for the same set of questions. Do you think it will be easy to reverse engineer the pattern? A human response will almost never be as repetitive.

Answer: Actually a chatbot designer/developer can put alternative responses for the same question to make an AI sound less repetitive (check out AI design tips: Design Tips - Juji Documentation) In the future, such alternative expressions can also be auto-generated to put more “natureless” into a machine – the technology is just not mature enough to be used yet.

Question 30: What other kinds of knowledge models can it work with?

Answer: It should be able to work with any kinds of knowledge models. Our AI architecture is model agnostic.

Question 31: At the outset, I missed your definition of ‘democratizing AI’? Could you repeat, please?

Answer: Making AI widely accessible to people and organizations that don’t have AI expertise or programming skills. Here is an analogy: back in 1970s, computers are only accessible to organizations such as big banks that can afford them, now “democratizing computing” means that almost everyone can access and use a computer (cell phone is a mini computer too).

Question 32: Are we talking about a fully cognitive AI or a chatbot which is context aware? I feel there is a big difference. It will be great to your perspective.

Answer: While I don’t know what the word “fully” fully means in your statement, what Juji is creating are AI assistants in the form of chatbots with cognitive intelligence: (1) language abilities (2) attention + memeory; and (3) the abilities to read people and personalize engagements.

Question 33: How are linked topics handled. For example: If you put in a weekend day in the restaurant bot would it be able to warn you that the coupon is not valid on that day?

Answer: In Juji, a chatbot designer/developer can easily call a function or API to retrieve relevant information. For example, to check if a coupon is still valid, you could feed the coupon code to a function or API call and the call will return the validation of the coupon. Similarly, if a user asks if I could use the coupon today, by feeding the date into the call, the chatbot could inform the user accordingly.
One can call a function or make an API request in a simple GUI (we are updating the documentation, feel free to reach out to use on forum: forum.juji.io).

Question 34: Great presentation! I have a question on conversation topics and context. How is a chat-bot building context data on specific conversation. What technology is used. Can memory of conversation be used to store user preference, etc. ?

Answer: We have used a number of technologies to enable an AI to understand context, e.g., optimization algorithms that tries to maximize the likelihood of conversation coherence. In Juji, a chatbot designer/developer can use custom “attributes” to store various information gathered during a conversation and keep them around to “remember” things. Check out our documentation: Chatbot Monitoring - Juji Documentation

Question 35: Is it easier to develop domain-specific (for example, health and financial domains) AI assistants compared to general, all-domain, AI assistant?

Answer: It really depends on the requirements. If a domain-specific AI assistant requires in-depth knowledge and deep conversations, it is as hard as the task of developing a general purpose AI assistant.

Question 36: Could you share the DOI of your paper, please? :slight_smile:

Answer: Which paper? Here is our publication website that includes links to most of our publications: https://juji.io/publications/

Question 37: If this tech becomes good enough, it can perform psychological analysis of its users. What are your thoughts about the dangers of influencing users to much?

Answer: Yes, we do realize the benefits and risks such technologies may bring. Check out our scientific articles on such discussions: https://dl.acm.org/doi/10.1145/3232077

Question 38: Does the system try to understand the user’s emotions during the dialogue and use it to adapt responses?

Answer: Yes, we created cognitive AI assistants to better understand a user from various aspects, including his/her emotions and sentiment. Juji AI assistants have some built-in responses that try to empathsize with users’ emotions. For example, if you create a Juji chatbot and add a topic that asks a user “how are you feeling today”. Then preview the chatbot, you will see how it handles a user’s emotions (e.g., responding with “I’m doing great” vs. “Feeling terrible today”.)

Question 39: I see potential for bias that will impact an AI’s assistant. How might you evolve an AI assistant to prevent bias behavior?

Answer: You are right, biases could exist for any AI assistants, e.g., the training data used, models trained, or how conversations are designed (just like us humans who hold many different types of biases). Establishing ethical AI practice guidelines and enforcing the following of such guidelines would one way to help eliminate such biases.

Question 40: Is the platform compatible with text-to-speech and speech-to-text?

Answer: Yes, someone actually used Google API STT and TTS to enable voice interaction of a Juji AI assistant: https://juji.io/ai-chatbot-api/#voice-bot

Question 41: How is Juji updating real-time POI information such as hours of operation, changes in menu, time to wait for a table, etc. It is interesting that Juji knows what days of the week a coupon is valid, but how does it do so without knowing which restaurant?

Answer: Business-specific information, such as hours of operations, changes in menu, or time to wait for a table, needs to be fed to a Juji AI assistant. Such information could be manually updated by a human staff (e.g., restaurant manager) or automatically fetched from a database (e.g., e-commerce product inventory database). This article describes how to update a Juji chatbot including live updates without disrupting ongoing chats: https://juji.io/blog/q-a-dashboard/

Question 42: How does Juji handle low confidence in intention prediction?

Answer: Check out this article that details how Juji handles low confidence during its interpretation of user input: https://juji.io/blog/question-recommendation/

Question 43: Can Juji detect and handle unknown values related to different expressions of user requests?

Answer: Just sign on Juji, create a chatbot and then add a topic that asks a user “how tall are you”? Then preview the chatbot and input a value is NOT a height measure and see how Juji handles it.

Question 44: Do you address privacy issues in the development of AI assistants?

Answer: Thanks for asking this question! We take privacy issue very seriously in Juji’s practice. We also advise Juji clients to always explicitly ask for their users’ permission/consent before collecting any private information. Moreover, we have also put “watermarks” in certain conversation topics (e.g., topics that elicit user private/sensitive information such as social security number), so during a chat, a Juji chatbot will automatically warn a user to be cautious about disclosing personal information to the chatbot.

Question 45: Can an expert system emulate Juji?

Answer: I’m not sure what kind of expert system you have in mind… Some of our university collaborators have already thought of developing simulation systems to interact with Juji and better improve Juji.

Question 46: Did you train any model?

Answer: Yes, we train many. many models to improve Juji AI’s capabilities to understand its users for different tasks across domains.

Question 47: How can a library department make use of such chatbot?

Answer: Just like any other academic departments, a library department could use a Cognitive AI assistant to reduce administrate burdens, improve student experience, and elevate staff morale (e.g., they don’t need to answer repetitive student questions over and over again!). We’d love to show you concrete use cases that could help a library department. Feel free to email us hello@juji.io

Question 48: Which programming language would you suggest for creating a chat bot?

Answer: We use a combination of programming languages to create Juji, including Clojure, Java, and Python. On Juji, you don’t need to progam to create a chatbot. If you do want to program, try Clojure.

Question 49: I’ve been married to my beautiful bride for more than 30 years. There are times she does not even understand what I am saying, or trying to say. How can AI be democratized as our perspectives (background and belief system)( are so different and this will influence communications.

Answer: AI is created by humans and follows humans’ commands to understand and not to understand. Your beautiful bride may sometimes choose “not” to understand what you’re saying, it’s not because she does not really understand and it’s just what she chose to do so :slight_smile: Machines/AI however would not have such freedom as we humans have. If we create AI to understand, it will.

Question 50: Is it open source?

Answer: No, it is not open-sourced.

Question 51: Besides developing for the two roles of single (chatbot %2B human), do you anticipate your platform to support group or team interactions (multi-party %2B chatbot, human-party %2B two separate role chatbot, etc.)? Would your current model/approach scale this way or would it require a different architecture/approach?

Answer: Juji’s infrasture is in fact designed to support multi-party conversation from the very beginning asJuji supported Wizard-of-Oz studies (both humans+machines interacted with users) to better understand user interaction patterns and needs. Check out this short article of ours: [PDF] Building Real-World Chatbot Interviewers: Lessons from a Wizard-of-Oz Field Study | Semantic Scholar

If you deploy a Juji chatbot on Facebook (choosing the FB deployment option), you can also easily check out multi-party conversations in a Juji chat as both humans and the chatbot can chat with FB page followers in one chat window.

Question 52: How does Juji maintain the semanticity of the conversation?

Answer: As mentioned in our talk, we have developed mixed-methods (known as neural-symbolic approaches) to interpret the semantics and identify specific concepts expressed by a user during a conversation.

Question 53: As mentioned in our talk, we have developed mixed-methods (known as neural-symbolic approaches) to interpret the semantics and identify specific concepts expressed by a user during a conversation.

Answer: We are in fact collaborating with unviersities to help more people including those with cognitive impairment. Since the work is still at its early stage, we’d be happy to keep you posted on the progress. Just follow us on Linkedin where we post recent advances in cognitive AI: https://www.linkedin.com/company/juji

Question 54: Would it be just as easy to build a chatbot for another language?

Answer: Depending on the requirements, in most cases, yes, it should be easy: https://juji.io/ai-chatbot-api/#multilingual-bot

Question 55: Is there any language beside English?

Answer: While we focus on English, we can certainly add other language support per clients’ request.

Question 56: Just chatted with Juji, “do you understand Japanese?” and unable to answer

Answer: Thanks for letting us know! Juji is far from perfect and does not understand many things and we are working on improving it.

Question 57: Accessibility & disability? - screen reader/cognitive?

Answer: Similar to screen reader that help vision impaired better access information, our work helps non-IT people better access/adopt AI.

Question 58: Is there an ethical framework for what the chatbot platform can or should be used about? For example, it can be used by agents that intentionally attempt to promote fake news etc. Is there some provision about that?

Answer: Great question and thanks for asking! While we don’t have a formal ethical framework per se, we have started working on formulating a set of ethical guidelines. Check out this recent paper in ACM CHI 2021 that lists a comprehensive set of chatbot evaluation criteria, including evaluting ethics: https://dl.acm.org/doi/10.1145/3411764.3445569

Question 59: What is the little D on top left of T?

Answer: The little “D” stands for “Digression” - all the topics with a little D represent all possible conversation digression topics. As mentioned in our talk, there are two types of topics managed by a cognitive AI assistant, the topics in the main flow (called agenda topics) and the rest (digression topics). Any topic other than the topic in focus could be a digression topic. That makes conversational management extremely challenging as the AI must recognize which topic should be switched to given a user input.