Posted by Libby Miller on
The Data team have been doing a hackweek about social networking, using the week as an opportunity to try novel (to the team) techniques for comment classification, sentiment analysis and network analysis.
One sub-team found several clusters of tweets related to the BBC, each with their own distinct vocabulary, and they attempted classification using rules-based and machine-learning techniques (machine learning came out on top but suffered from a lack of training data).
A second team used parts of speech (‘aspect based sentiment analysis’) to identify the entity that a comment refers to, allowing comments directed at the BBC as an organisation to be distinguished from comments about particular programmes or people.
The final team researched some of the groups of people tweeting about the BBC and its programmes to understand if and how they are related.
Internet and Society have been pursuing their projects around making machine learning understandable and controllable, with Ben testing a bird identification model on a Raspberry Pi, and Jess researching XAI (‘Explainable AI’), which is a relatively new research field. They are working with David and Alicia to make an application which explains how the model comes to its conclusions about what sort of bird is in a photo, using various techniques including Misa’s brilliant FlashTorch project, a visualisation tool for CNNs.
A group called Ciné-Real (a 16mm film club and associated podcast) watched King Kong using BBC Together as part of the experience, which was a nice test for us (preplanned events have different requirements to spontaneous watch parties).
Meanwhile Anansi are preparing to publish a series of blog posts about how they made the synthetic voices project (the survey is still open), and have also been planning their next steps as a team. Henry’s been writing a paper on voice, and also has a blog post about Looking for Nigel coming up.
This post is part of the Internet Research and Future Services section