“PAWLIN Technology” company developed a system for automatic neural classification of textual messages in the field of airport security. The system was designed for bilingual usage: both Russian and English.
Now the system is successfully embedded and being used with the aim of determining the level of danger in the airports, that sufficiently eases the work of highly qualified experts, who classify the news messages manually.
Within system designing a module for generation of text messages vector descriptions was successfully applied, it was based on such advanced technologies as Word2vec and LSTM, which show the best results in the field of NLP nowadays.
The system includes automatic conveyors for learning, additional learning and classification of textual messages, which interact with each other.
The algorithm of system operation includes several stages:
- Creating and learning of a neural network model based on news messages obtained from database;
- Online additional learning of created model;
- Decision: to update created model or to keep its latest version;
- Classification of unmarked news obtained from database in real time.
In classification accuracy check mode research results are displayed in the form of matrix normalized to the number of objects.
Table 1. Example of news classification
|Classified message||The probability of relation to the category of threats to airports, calculated using algorithm, in %||True value|
|Two aircrafts with passengers made an emergency landing at the John Kennedy airport. The reason for the forced landing was an anonymous message about the possible bomb exploding on board at any moment.||98, 3||The message belongs to anonymous messages category and is related to airport security|
|“Russia” Airlines said that the aircraft with the national team of Saudi Arabia landed on two working engines. According to preliminary data, a bird hit the blade but it did not lead to a fire. This was reported to “RBK” by the carrier representative. “Passengers safety was not at risk. The alarm was not announced during the aircraft landing in Rostov-on-Don”.||7, 2||The message is not related to airport security. In spite of presense of such words as ‘risk’, ‘hit’,’safety’,’alarm’ — message was NOT classified as having relation to terrorism|
The advantages of a developed system:
- Fully automated learning process, additional learning process and classification of textual messages;
- It is bilingual;
- High speed performance;
- The ability of use regardless of presence or absence of a GPU on the platform.
Table 2. Accuracy results of experimental studies of a system health module
|Probability of a correct comparison with a similar textual description||not less than 90||97,4|
|Probability of a false comparison with textual descriptions with different meaning||not more than 5||2,6|
Table 3. Speed results of experimental studies of a system health module
|Stage||Number of uploaded news messages||Runtime on a CPU, min|
|Creating and learning||9381||6.27|
The system can be widely used both in aviation and other fields, that are faced with the tasks of event classification and forecasting, such as business analytics, financial operations, enterprise workflow planning, etc.