Testing Talla Against Our Own A.I. Bullshit Detector

Posted by Rob May on Jan 12, 2018 6:35:05 PM


We recently launched an ebook called "Bullshit, Hype, and a Little Bit of Magic: How To Make Sense of It All When Buying A.I. Products." Today I want to take a quick tour of Talla to understand where we stack up against our own analysis.

Download: Bullshit, Hype, and a Little Magic

The purpose of the ebook is to sift through the hype and the reality to find those little pieces of magic in A.I. products. Talla's core product was created to help bring a little A.I. magic to your knowledge management and company info. Now let's see how the Talla product answers the questions in the ebook.

The first question you should ask when evaluating an A.I. product is about where the training data comes from. For the most part, because we are dealing with company data, we don't use many pre-trained models. We rely on data that comes in from your company, as you enter it, and we train custom per-company models because business vernacular varies quite a bit company to company and industry to industry. Training on something like Wikipedia, which is what most NLP models do, isn't as effective for this.

Because we rely on your own data, we also make it really easy for you to annotate it. The same way you add a little bold or italics formatting to make certain things stand out so your content is easier for other humans to read, we make it easy for you to add some simple tags (in a similar style to how you add bold and italics) to make your content easier for machines to read. This makes it fast and efficient to train machine learning models custom to your content and company information.

Talla learns as it goes. The more data you input, the more Talla knows. We can deal with an initial data dump, but, models are constantly retrained as you create new data so they stay updated, and we can start with a small amount of data if that is your preferred method. In fact, what we often see is people use Talla's knowledge base functionality for one particular type of company data, until they realize how powerful it can be and decide to look at moving everything over to Talla.

The second question we encourage you to ask in the ebook is about ramp up time. Once an A.I. system has data, how long before it is useful and "smart?" Talla is useful the very first day you start. The ramp up time is minutes.

Because we provide a web interface and also chat interfaces in Slack and Microsoft Teams, Talla's knowledge base is useful even if you don't use the artificial intelligence, because you can use it like any other knowledge base. But if you choose to do some simple annotation, Talla quickly learns things like a company glossary, information ownership, when things go out of date, and how concepts tie together into a knowledge graph.

The third major question is about error correction. How do inaccuracies get resolved? Machine learning models are probabilistic, like humans, and aren't always 100% correct. When they make mistakes, what do we do? This is where Talla really excels. The Talla system is full of workflows that constantly check and cross check information, and provide simple ways for humans to "teach" Talla when Talla is wrong about something. These workflows are difficult to describe effectively, so if you want to see them, just contact sales@talla.com and ask for a demo.

What we have built is a tool that makes A.I. applied to your company information as easy as possible. Our system is trained with your data, quickly, and is easy to continue to train if it makes errors or needs to be retrained. We believe that artificial intelligence can change the way that you work and have massive impacts on your productivity and effectiveness, and Talla in particular should be a key component of your A.I. strategy. Please email us if you want to learn more, or download the ebook if you want to learn the questions to ask any A.I. vendor.

Contact Sales

Topics: Artificial Intelligence, knowledge management