Here at ICWSM we are getting great presentations from all the various fields concerned with social media. These include: text mining, artificial intelligence (especially NLP/CompLing), psychology, graph algorithms, social network theory, data visualization/UI design and data mining. One of the major roles and purposes of the conference is to bring these areas together to better model, support and leverage social media. To do this, I think we need to be explicit about where the intersections lie. Of course, drawing a graph of these relationships is likely to produce a fully connected graph, but before we do that we can at least highlight some key relationships.
Psychology and Text Mining
Much of the work on text mining focuses on the document – and yet we are concerned with social media, in other words, media produced by individuals in a social context. Where is the individual or community? Understanding opinion must take into account some model of what opinion is (a belief), the psychological processes of opinion (how do we change opinion?) and so on. Documents do not have opinions, people do!
Text Mining and Social Networks
The most obvious thing I see here is the relationship between appraisal (sentiment/opinion/…) and notions of community, influence and authority. Opinions are formed by people in the context of some environment. Often, that environment will include one or more individuals who are expressing the same (or opposite) opinion on the same topic. Is there pressure (in a public forum) to express agreeable opinions? If so, should we discount them in some way?
Text Mining and AI/NLP
While text mining often eschews deep representations, when we look at reviews and other opinion laden text, we can see clear indications of both expressions of ontological structure (‘the camera has a great lens’) and discourse structure (‘it has a great lens, however…’). Certainly we must avoid problems of AI or NLP completeness, but we should also build more detailed models of these aspects of opinion expression. There are some publications already that weave discourse structure into sentiment mining which is a great start.
Text Mining and Data Visualization/UI Design
Opinion mining is challenging, but we are getting to the point where some practical value can emerge. However, improving the value of the technologies is not something that needs to be limited to improving the basic accuracy of the inference piece. By exploring how results can be amplified and the impact of errors can be diffused we can get more out of what we already have.
I believe that attendees of ICWSM have a great opportunity (and many of us are already benefiting from it) to fully leverage interactions between different research communities. In some ways, this is what defines our meeting.