Debut of AnswerGrid

Fri, 19 Dec 2008 8:15:14 -0700

Welcome to the debut of yet another project developed in the Ziemantics Lab, called AnswerGrid, or AG in short.

As its name suggests, AnswerGrid is a question & answering system. It is designed to answer your questions asked in natural language, but it is not like a search engine that presents the list of webpages related to your query. Instead, it gives exact answer in natural language.

AnswerGrid, an artificial intelligence?

AG is not the SkyNet from the movie "The Terminator", but in order to build AG, we are using the techniques and methodologies from the artificial intelligence field.

AI that we know from Hollywood movies can get into a real conversation and--of course--action. In real life, however, there is no such system, but bots. Bots, like A.L.I.C.E., can create meaningful conversation to some extend but their capacity is only enough to keep conversation going. They shape the conversation based on pre-definite templates. Their primary purpose is to make a conversation rather than helping you achieve a goal.

With AnswerGrid, we will try to break that. Our goal is to develop a system that you can communicate with it to learn more about specific topics and achieve a goal. These systems are also called expert systems.

As you may notice, homepage of AG looks like a homepage of any search engine with an input field at the center of the page to enter your query. As you enter your questions and get the responses, you will realize that it is more like a chat panel. So, it looks like you are having a conversation with AG.

AnswerGrid, a proving ground

At its current stage, it is far from being finished, because it has very limited knowledge of the world and it cannot even keep conversation going and remember what you say.

Hence, currently, it is only a proving ground to show that our preliminary design works. In fact, the system behind the AG will be used in number of projects in the future and AG on its own will be a debut area for our technology being developed in Ziemantics, LLC Lab.

AnswerGrid, what it can do

Our goal is to train our system with fact-based information on as many specific domains as possible and to be able to answer questions asked about those trained domains.

To begin with, we chosed "geography" domain, because data in this domain is fact-based and there is a public database that we can work on, which is called "The World FactBook" published by CIA.

The World FactBook is a public database where fact-based information about all countries are presented in specific format that we can easily convert it to our internal conceptual representation. For example, number of people living in one country, its national currency, its capital and many other types of information can be found in this database.

To achieve this, we determined the types of questions that we can answer via this database and came up with over 250 questions. Some of them are as easy as "how many people live in China" and other can be as difficult as "which countries produce more electricity than Australia does." We will talk more about question types and their difficulty level on another post. Following the determination of these questions, we created a model of this domain and introduced concepts that system can use to understand the question, locate its answer in the knowledge-base and lexicalize found answer for output.

At this point, it can process up to 200 questions and find the answer of about 140 questions out of those 200. But it can only generate answer in natural language for about 90 questions of those 140 answerable ones. Confused? Dont be. You can find those questions and more [ here ].

What is yet to come

Development of AG is far from being finished and this release is actually an early birth. Since we designed and developed this technology from scratch, without making use of another application, there are many iterations in this development and each iteration takes time while enhancing the representation, reasoning and conversation capabilities of the system.

At this point, we passed couple of milestones, such as this one, the debut of the system. The next milestone is to incorporate logic and numeric reasoning. So far, we only implemented fact reasoner and when we implement those other two reasoners, the system will be complete and should handle any logical and numerical relation.

To sum up,

Even though AG is online and it is not good enough to satisfy the demand of information-hunger users, what we are doing here today is to show how far we came so far and what we are up to. We hope you appreciate this little progress and make use of it despite its limited capacity.

Later on, we will get into the system and describe its parts in short. Till then, if you have any question about our project or like to share your thoughts with us, please email us or send a feedback from our website.