Transition-based neural network model for extracting composite objects and their attributes from natural language texts

Abstract: Extracting structured information from user feedback texts is a task of great scientific and commercial value. However, modern methods of extracting information from texts either do not take into account the structural relationships in the extracted knowledge or have low accuracy of their extraction. The paper proposes a transition-based neural network model for extracting composite objects and their attributes. The model is trained in multitask fashion when span extraction and link prediction tasks solved by the same components. The quality of the proposed model was tested on the tasks of processing reviews from Amazon and AliExpress stores and processing user requests from the Google Play store. The experimental results demonstrate that quality of the extraction of links between spans increases by value from 0.07 to 0.172 F1 depending on the task.

Keywords: natural language processing, machine learning, neural networks, sentiment analysis

For citation:
Ehlakov Yu. P., Gribkov-Egor-Igorevich E. I. Transition-based neural network model for extracting composite objects and their attributes from natural language texts. Doklady Tomskogo gosudarstvennogo universiteta sistem upravleniya i radioelektroniki, 2020, vol. 23, no. 1, pp. 47–52. DOI: 10.21293/1818-0442-2020-23-1-47-52