Recently there’s been a lot of hype around Artificial Intelligence (AI). Google, Microsoft and Facebook are all investing heavily in these technologies, there seems to be a new startup every day and the press is full of dire predictions about automation and the future of work. Google’s Go program beat another world master, translation programs continue to improve and Gartner predicts that by 2018, over 3 million workers around the world will be supervised by a “roboboss”.
There is significant debate about how realistic these predictions are. After all, say the cynics, the promise of intelligent agents has not been realized, despite years of anticipation. The quality of translations, while measurably improved, leaves much to be desired even though Google, Facebook and other industry giants are spending vast sums on machine learning.
The problem is a simple one to articulate and a difficult one to solve. True “artificial intelligence” can only be achieved when computers understand and can respond appropriately to language, in all its richness and variance. Machine learning algorithms, however promising, have not yet been able to overcome this barrier. They are effective, but only within well-defined and circumscribed problem spaces.
Current linguistic theory describes the process of language acquisition as a categorization exercise. Children are exposed to spoken words and associate them with physical “things”. They mentally organize the concepts into categories and intuit the relationships between the categories. In this way, they learn how to use language to communicate. The mechanism by which this categorization and relationship building takes place is not well understood and as a result, it cannot be replicated in software - this is the chasm that artificial intelligence cannot yet cross.
The Semaphore semantic model is a collection of concepts and relationships. From the model, we generate rules that classify information assets (by determining what they are about and attaching precise, complete and consistent metadata) or extract critical facts, entities and relationships. Natural language processing, statistical analysis and other techniques are used to surface the evidence within the information assets that causes the rules to fire. Semaphore bridges the AI chasm today and provides capabilities that solve real-world business problems in a way that true AI cannot.
At Smartlogic, we believe there are currently some aspects of AI that can be applied to support human expertise. To address our clients need to solve business problems today, we use machine learning and AI to assist humans and simplify the process of capturing domain knowledge and putting it work. To this end, we leverage machine learning algorithms like Support Vector Machines, to mine large corpuses of public information and return ordered collections of concepts that might be relevant to the model. Subject matter
experts can then determine where and if they fit and quickly and drag and drop them into the model.
Another area where machine learning can assist human experts is in better refinement of the text analytics. Hidden Markov Models (HMM) are machine learning algorithms that use Bayesian statistics to identify parts of speech –a necessary component of analyzing text to determine its meaning. However, the subtleties of language across organizations provide additional challenges as HMM typically come with static statistical models.
One of our current research focus areas is in how we can use machine learning to subtly modify the Hidden Markov Model over time as we detect different language usage patterns within a specific corpus or problem domain. This will make the evidence surfaced to the rules more exact and improve the precision of classification.
Until the mechanism for language acquisition is understood we believe true AI is not achievable. Today, Smartlogic AI – Assisted Intelligence – is assisting business subject matter experts and information scientists rather than replacing them.
