ISSN 2587-814X (print),
ISSN 2587-8158 (online)

Russian version: ISSN 1998-0663 (print),
ISSN 2587-8166 (online)

Egor Gribkov 1,2, Yuri Yekhlakov  2
  • 1 TomskSoft LLC, 8, Nahimova Street, Tomsk 634034
  • 2 Tomsk State University of Control Systems and Radioelectronics , 40, Prospect Lenina, Tomsk 634050

Neural network model for user request analysis during software operations and maintenance phase

2020. No. 1 Vol.14. P. 7–18 [issue contents]

      This article offers a transition-based neural network model for extracting informative expressions from user request texts. The configuration and transition system that turns the process of informative expression extraction into the execution of a sequence of transitions is described. Prediction of transition sequence is done using a neural network that uses features derived from the configuration. To train and evaluate a proposed model, a corpus of annotated Android mobile application reviews from the Google Play store was created. The training procedure of the model for informative expressions extraction and selected model’s hyperparameters are described. An experiment was conducted comparing the proposed model and an alternative model based on a hybrid of convolutional and recurrent neural networks. To compare quality of these two models, the F1 score that aggregates recall and precision of extracted informative expressions was used. The experiment shows that the proposed model extracts expressions of interest better than the alternative: the F1 score for spans extraction increased by 2.9% and the F1 for link extraction increased by 36.2%. A qualitive analysis of extracted expressions indicates that the proposed model is applicable for the task of user request analysis during operation and the maintenance phase of software products.

Citation: Gribkov E.I., Yekhlakov Yu.P. (2020) Neural network model for user request analysis during software operations and maintenance phase. Business Informatics , vol. 14, no 1, pp. 7–18. DOI: 10.17323/2587-814X.2020.1.7.18
Rambler's Top100 rss