Modelling Convergent Social Tagging Processes

Next up at Web Science 2016 is Paul Seitlinger, whose interest is in social tagging practices. These provide a valuable insight into human cognition and offer an opportunity to validate lab-based models 'in the wild'. One key question in this is how semantic stabilisation, or consensual use of tags, comes into being. This is influenced by the interplay of both human and non-human actors.

As multiple users tag the same piece of content, unique new tags may be introduced – but over time, the likelihood of new tags being introduced declines, and there is a gradual settling on a shared set of most prominent tags. This may be influenced by both a personal vocabulary of common background knowledge, and by processes of imitation, especially if the tagging system promotes such imitation by highlighting the tags used by previous taggers.

The process for tag stabilisation can thus be modelled using relatively simple rules; the assumption is that there are two stochastically independent processes at play here. The outcome validity of such models, as compared to processes observable in real life, is high; however, we also know that users are reflective actors and that the underlying assumptions of the model are very simplistic – there is low process validity in this model, therefore. So, how can we mode reflection?

What is required here is a computational model of human memory search: this search of the memory is initiated by engagement with the item to be tagged, and searches for relevant context available within the user's memory. For a given article, for instance, such context could be information about the author, the topics covered in the article, or other relevant information – and this contextualisation might activate several subsequent memory search processes that generate further contextual information.

Each iteration would then yield a new tag to be used in the tagging process, for instance; the longer this goes on, the focus of the tags moves away from the immediate, commonly used tags (the spotlight drifts). A test of this model through a multi-agent simulation, based on Wikipedia data, generates a good match with real-life observations, confirming the 'drifting spotlight' hypothesis.