≡ Menu

First Look – Cogito


Neil and I caught up with Expert System, a Semantic Intelligence company, last week to discuss their Cogito product. Expert System is based in Italy, has 145 employees and originally worked on spell checkers for Microsoft. While they have a lot of business in Europe they are now growing in the US. Their core pitch is, frankly, the classic attack on keyword search – that a lack of understanding of the language makes certain keyword searches useless and undermines the quality of results in many others. These problems include different meanings for the same word, different words for the same meaning and different words that have closely related meanings. Their Cogito product aims to deliver meaning based processing to address these and other issues. It is a horizontal technology designed to take a wide range of unstructured information (documents, email, web or mobile content, information services) and conduct linguistic analysis to generate a conceptual map that can be used to power categorization, discovery, semantic search etc.

Cogito is based on four approaches used in human comprehension – understanding word forms, the parts of speech, how words relate to other words and the context of key words (morphological, grammatical, logical and semantic analysis). At the heart of Cogito is a semantic network (CAT is a MAMMAL, MAMMAL is an ANIMAL etc). Cogito has these networks for English, Arabic, Italian, German and French and the English one, for instance, has 350,000 words with 2.8m relationships. In addition some 3,500 rules are defined for these relationships. The development studio allows the development and management of custom rules in addition so that users can create their own. They claim to have developed the product with a real focus on performance – a small memory footprint, good scalability, high throughput etc.

They had some interesting case studies like automatic categorization of incoming news feeds, knowledge management, customer sentiment analysis and customer care email processing. We got a short demo and it did show some nice disambiguation from the context such as telling apart gas for gasoline and gas for accelerator pedal. The ability to add your own rules and include the product through APIs both struck me as particularly interesting.