A Guide to Training the Mind Expression Engine

“I live in terror of not being misunderstood”
-Oscar Wilde

Artificial Dumbness

Try to have a simple conversation with your Alexa speaker or a customer service chatbot and you’ll quickly realize that despite recent advances, the ‘intelligence’ part of artificial intelligence is still more aspirational than real. Tell Alexa “Call me an ambulance” and it might conclude that your name is ‘an ambulance’ and update your contact information — while you’re sprawled on the floor having a heart attack!

It’s not really the chatbot’s fault. Human language is complex and nuanced. Words can have different or even ambiguous meanings, depending on the context. And intention is often conveyed through body language or intonation, which even humans sometimes miss. Then of course there’s sarcasm, double-entendres, subtle humor, and innuendo. We know from our own lived experience that we don’t always get one another. Can you imagine then how challenging it must be for a computer to understand us, especially when you throw in misspellings, emojis, and slang?

Building a Better Bot

Big Tech’s approach to this problem is based on the premise that the more processing power and the larger the data set, the more proficient and ‘intelligent’ our AI systems will become. Though skilled at detecting patterns in human language, algorithms based on a machine learning approach fall short in natural language understanding (the ability to determine intent). That’s because they ignore the logic embedded in human language and rely instead on statistical computations to generate predictive approximations of human intention. In contrast, the Mind AI reasoning engine learns through contextualization and abstract reasoning, just like humans do. This ‘Canonical’ approach gives the Mind AI engine an edge in working out human intention, which is why Mind AI decided to focus on chatbots for its first commercial application.

The Mind AI reasoning engine can reason and draw logical conclusions based on inputs. For example, if I teach the AI that “my phone battery is dead” is a cause for “my phone won’t turn on”, then the AI will understand that a phone needs power from its battery to function. It will know this when it encounters the same issue in the future. To understand human requests across multiple domains, the Mind AI reasoning engine needs a robust database built on natural language inputs. To build this database, Mind AI teamed up with Synesis One to crowdsource the knowledge (through our train2earn App) the AI needs to create a ‘mental map’ of the world. This includes both domain specific knowledge and more general knowledge.

Here’s how it works. Our architects create campaigns (based on client needs) to crowdsource the domain specific knowledge needed to answer customer queries. Builders choose which campaigns to work on and use their creativity to come up with ‘utterances’ (from linguistics, the smallest unit of speech) to express every possible way that someone might make a particular query in a given domain. The Validators then reviews and either rejects or approves each submission. Builders whose utterances are approved are paid in SNS tokens. In this way, over time, the Mind AI reasoning engine expands its ‘mental map’ of the world, which will help it contextualize human queries and make logical inferences about human intentions.

The Guide

The Mind AI engine classifies utterances into three types: specific, general, and entailment. These categories are rooted in linguistic theory but also map on to Mind AI’s unique three-node data structure (called canonicals) that the engine uses to establish logical relations between ontologies (see technical white paper). Mind AI’s linguists have determined that we need approximately 300 validated utterances with pattern diversity per topic for the AI to grasp the various ways humans might express a given query. Crafting utterances that pass validation and improve the AI can be challenging, especially when you’re just getting started. To help you increase your success rate in train2earn, we’ve put together the following guidelines.

Specific Utterances
A specific utterance, as the name suggests, is one that is related to the topic of the campaign, but with more detail — such as details of an action, a situation, or a specific intent. Typically, this means using more words, which makes the utterance longer. You can do this by using more detailed words than the original topic or changing the sentence structure (syntax) to express the same idea. Consider the following example in the domain of electronics:

In the above utterances, the underlined word(s) provide greater specificity to the topic installation services availability. These utterances teach the AI different ways that someone might inquire about installation services, which will enable the AI to understand customer inquiries on this topic in the future. Note that the validators will reject utterances with the same sentence structure but different words for the same thing. So, for example, ‘Do you offer installation services for television / stereos / satellite dishes / home theater / etc. will be rejected, as it doesn’t teach the AI anything new.

General Utterances
A general utterance, as you might guess, refers to the theme described in the topic subject example, but with fewer details. General utterances will include less details of action, situation, and specific intent, thus making them shorter.

If the utterances are too broad and fall outside the scope of a given domain (in this case, telecommunications), then they will be rejected by the validators. Consider the following:

✘ Can I buy a package plan?

✘ I am interested in buying a package plan

These utterances do not describe “mobile package plan” in a more general sense, but rather invite different interpretations in different domains (such as travel package plan).

Entailment
Entailment describes a relationship between two sentences such that if the first one is true, the second must also be true. Entailment allows the AI to understand and reason logically based on natural language inputs, thereby improving the AI’s ability to make its own hypotheses about a given topic. Consider the following example from the Food & Beverage domain:

(1) Topic Subject: May I have some water?

(2) Entailment Utterance 1: I am thirsty.

(3) Entailment Utterance 2: I need water.

In this case, (1) entails the meaning of both (2) and (3). We can say that the Topic Subject entails the meaning of Entailment Utterances. Builders should create utterances that can replace the Topic Subject and keep the original intent intact.

The following utterance does not fulfill the topic subject’s intent though it is entailed by the topic subject.

✘ People can misunderstand how to use a laptop

To be validated, utterances should fulfill the intent of the Topic Subject, meaning they can replace the Topic Subject and still convey the same meaning. Given that the Topic Subject is true, utterances must also be true. Finally, entailment utterances should capture a part of information from the Topic Subject.

Conclusion

Natural language works for humans because we have shared experiences. From childhood, we learn about the world, the kind of things in it, and how they’re related. To use an earlier example, we understand what it means to be thirsty because we’ve all experienced it. Computers, on the other hand, don’t have human experiences. They are limited to their training data sets. What Mind AI is attempting to do is to provide their AI with a ‘mental map’ of the world, which will help the AI intuit human meaning and intentions.

Thanks to Synesis One’s train2earn platform, the AI can crowdsource the data it needs to make sense out of human queries. It won’t share our experiences, but it can perhaps begin to understand them well enough to make hypothesis about our intentions. The Mind AI and Synesis One teams believe this approach will narrow the gap between today’s limited machine learning and tomorrow’s artificial general intelligence.