One of the most difficult tasks in the field of data mining is the development of universal tools for the analysis of texts written in the literary and business styles. A popular path in the development of algorithms for processing text document corpus is the use of machine learning methods that allow one to solve NLP (natural language processing) tasks. The basis for research in the field of natural language processing is to be found in the following factors: the specificity of the structure of literary and business style texts (all of which requires the formation of separate datasets and, in the case of machine learning methods, the additional feature selection) and the lack of complete systems of mass processing of text documents for the Russian language (in relation to the scientific community-in the commercial environment, there are some systems of smaller scale, which are solving highly specialized tasks, for example, the definition of the tonality of the text). The aim of the current study is to design and further develop the structure of a text document corpus processing system. The design took into account the requirements for large-scale systems: modularity, the ability to scale components, the conditional independence of components. The system we designed is a set of components, each of which is formed and used in the form of Docker-containers. The levels of the system are: the data processing level, the data storage level, the visualization and management of the results of data processing (visualization and management level). At the data processing level, the text documents (for example, news events) are collected (scrapped) and further processed using an ensemble of machine learning methods, each of which is implemented in the system as a separate Airflow-task. The results are placed for storage in a relational database; ElasticSearch is used to increase the speed of data search (more than 1 million units). The visualization of statistics which is obtained as a result of the algorithms is carried out using the Plotly plugin. The administration and the viewing of processed texts are available through a web-interface using the Django framework. The general scheme of the interaction of components is organized on the principle of ETL (extract, transform, load). Currently the system is used to analyze the corpus of news texts in order to identify information of a destructive nature. In the future, we expect to improve the system and to publish the components in the open repository GitHub for access by the scientific community.
Предметные области OECD FOS+WOS
- 5.02.DI БИЗНЕС