- process the meta-data network, automate common patterns
- gather characterstics of the dataset -> will point towards types of methods, might reveal problems that need to be be addressed
- do some basic experiments such as term tracking, dependency parsing, etc -> will restrict and/or guide the choice of algorithms
- automate term tracking -> probably useful in any case
- gather tasks from basic experiments in suite of tools (& set up properly)
- pre-process the linguistic data, the objects' descriptions, including vectorisation and (basic) feature extraction
-
map experts' perspectives on social bias (domain experts are philosophers, social scientists, activists, really anyone but me):
- get a workable definition which can be translated into linguistic/statistical properties
- define boundaries of bias in opposition to other concepts
-
manifesto: carefully explicate our own goals and requirements to (1) be judged against and (2) make process transparent and reproducible
-
review previous work:
- already identified domains, problems and formalisations
- inspiration from methods already in use
- off-the-shelf algorithms and libraries
- limitations
-
formalise the world of biases (types, aspects, levels) to be detected in the first iteration
=> goal for February: be ready to implement a battery of detection algorithms for experimentation and feedback from the stakeholders
TBD