How we chose a machine learning library for working with natural language

In 2019, a new center for work with artificial intelligence (CII) appeared in OTR. Initially, it was created as a kind of experiment in working with new technologies. However, quite soon he received the first combat mission to automate the technical support of users of the GIIS "Electronic Budget". Anatoly Bezryadin, Director of the Directorate for Technologies and Competencies, and CII employees who took part in the ambitious project, spoke about the stages of AI implementation.





Calls to technical support are complex technical requests with many variables - section, subsystem, regulatory documentation, invoice, and so on. In some cases, the proposal text could be up to 20 lines of text. It is sometimes difficult even for a person to understand the context of such an application, and for AI it can even become an insoluble task. 





The list of difficulties was supplemented by tight deadlines for completing the task, as well as restrictions on the software solutions used. Since we work with government customers, we can use either our own developments or open-source solutions. 





Choosing the type of neural networks

ยซ ยป . , . .





. . . , .





. . . , . , . , .





. , . . . , . .





. . , . .





. .





, . . , , :





  • ;





  • ;





  • ;





  • ;





  • GPT;





  • BERT.





, . .





GPT BERT. , . BERT , , Google. .





:





  • open-source-;





  • ;





  • ;





  • ;





  • ;





  • ;





  • .





:





  • MATLAB + Deep Learning Toolbox;





  • Theano:





  • Torch;





  • PyTorch:





  • TensorFlow/Keras.





MATLAB โ€” The MathWorks. . Deep Learning Toolbox. , . 





. , . , . 





Theano โ€” Python. . , .





, Theano . . 





Torch โ€” . Lua. , . 





Theano, Torch . . . 





PyTorch โ€” . , Torch. Python. 





PyTorch , , . . 





TensorFlow โ€” , Google. Keras, . . 





TensorFlow Keras . , , . Google.





Google TensorFlow Keras . โ€”  , 10 . 





75% . . . . 85%.





ยซ ยป. . .





, . , .








All Articles