boy holding an enourmous pencil

The first stage in any project is to train the system to answer your particular question. This may sound like science fiction but it really is what we do. Our software gives computers the ability to recognise certain features of web pages. Our (human) team identifies examples of relevant comments that i-sieve computers then analyse, literally learning how to distinguish relevant from irrelevant, positive from negative, interesting from unimportant.

The team usually assign relevant comments into one or more clusters, that is, comments that follow a particular theme. For example there may be a cluster of comments about the music used on an ad, a cluster of comments about the effectiveness of the product, and so on.

This process typically takes between 2 days and a week. Once the computers have learned what they need to know, they can take over. We typically look at around half a million online resources, selecting those that address the question you asked, and classify them accordingly (routinely achieving more than 90% accuracy).

Sometimes the system will find resources that are clearly relevant but that don't fit into the pre-defined clusters of results. In this situation, the team are alerted and are able to define a new cluster, even if it is not what was originally asked. It was in this way, for example, that an analysis of sentiment concerning a drug sold as a back pain relief was discovered to be talked about among users as particularly good at relieving menstrual pain.