In order to teach an AI model to answer questions as correctly as possible, it is necessary to combine data selection and digitization efforts and involve experts in their fields in the project. Pyotr Ermakov, Brand Director of Machine Learning at Yandex, talks about this erythrocyte.
Image source: Google DeepMind / unsplash.com
“We specifically prepare and select the data used to train the model – this is the first stage. Equally important, not all data is available. There is a lot of non-digital media. We work with universities and libraries to digitize the material.”– said Mr. Yermakov.
Next comes the second phase of training, with trainers including: teachers, associate professors and doctors of science. An example of this type of work is the creation of a legal-focused AI model – in order for its answers to be correct, it requires the involvement of professional lawyers. “We hire experienced attorneys to ask the right questions and correct the model’s answers so that the model then learns the most relevant answers possible.”“, the expert explained.
When testing legal AI models, a special set of data and questions are generated that the system must answer while complying with the law. The effectiveness of the control depends on the size of the model and the “throughput”, i.e. the people responsible for checking the model’s response.
If you find an error, select it with your mouse and press CTRL+ENTER.
