Aim is to utilise machine learning (ML) to extract and generate metadata from the existing files by creating, training and improving a model (Nerual Network) to process currently existing files and the filestructures to extract labled metadata such as project date-time, dimensionality, projetc name and/or other metadata.

Desired attributes are:

A folder exists for each subproject, further README files exist in them for more information

• Filepath metadata extraction: extract different metadata from the file path of each file. (Current)
• Document metadata extraction: extract different metadata from coduments such as pdf reports.
• Hierarchical analysis and metadata sharing: copy metadata to related files based on hierachicl structure.

Tensorflow (TF) provides all the machine learning (ML) functionality and is a widely used scalable ML tool kit. Thus TF will be used for this project.

The ML process will generate and improve in an itterative process a model that can be used to predict metadata from new file path inputs. Note that the once a prediction has been made, it can be confirmed or corrected by a human and this response can be fed back into the ML process to improve the models accuracy.

The RNN will be able to process input of varying lengths, thus rather than feeding the entire path at once its fed character by character. The RNN will utilise Long Short Term Memory (LSTM) to process the entire path character by character and continously output information discovered at that stage.

Unlike regualr Neural Neworks (NN) which process each input seperately, RNN feeds data from the previous steps (previous characters processed) into the current computation.

Likly This Network will become a Deep Neural Newtork (DNN) which means it has more than one hidden layer. Note that the NN can be both recurrent (RNN) and deep (DNN).

DNN are known for better processing of more complex data and thus is likly to increase model accuracy. Similarily to RNN DNN are more complex than regular NN.

In conclusion we will initially attempt to implement and test a Deep Rrecurrent Neural Network (DRNN) and possibly compare it to RNN and DNN implementation depending on performance results.

Geoscience Information Team, Geological Survey of Queensland, Department of Resources, Brisbane, QLD, Australia, [email protected]