Natural Language Generation System for Stock News Generation

The subject of this research is related to the field of natural language generation (NLG). NLG systems are usually built to make automatic the process of creating simple texts such as weather report or financial news, etc. The goal of this thesis was to build an NLG system that would generate stock news based on the data on MOEX and RTSI indexes changes and to describe the process of its realization. In this regard we had some problems to solve, namely, to build the architecture of the system, to choose methods for index changes analysis, to examine the features of financial news texts, to choose an NLG approach, to write a program in Python and to evaluate the results. In the thesis we describe the system implementation process and discuss the decisions we used to make to achieve the result. First of all we made some research on index changes and index behavior in general, defined the types of stock news which the program should generate and created the layouts for the texts. Our program has two modules: the analysis module that takes stock data as input and determines the index changes using the special algorithm and the NLG module that generates news on basis of the information it gets from the analysis module. The NLG module uses rules and templates to create the news texts. It also uses financial lexicon that was collected for this purpose. While working on the NLG module we examined how the means of coherence function in the automatic text generation and which rules can manage this process. The evaluation has shown that our program is adequate both in index analysis and NLG.