A Framework for Summarizing Stock Market News with Language Models: BORSUM
Güncel SARIMAN Nebi Berke İÇÖZ
AbstractAlthough the increasing data density in recent years positively affects our learning and development speed, it also brings complexity and difficulties in decision-making processes. The time allocated to interpret and understand all the data turns into a process that users do not want. Especially in areas such as the media sector, many news, video and audio content produced cannot be used because of their length. In financial markets, a lot of content is produced on the same subject throughout the day due to foreign policy and politics. Markets are affected by a wide range of news such as corporate earnings reports, mergers and acquisitions, regulatory changes and geopolitical events. It becomes a very difficult process for investors to follow these news. At this point, automatic text and voice summarization stands out as a powerful tool for processing complex information quickly and effectively. This paper deals with a specially created dataset for summarizing Turkish stock market news and the automatic summarization process with language models adapted to this dataset. The Turkish language poses special challenges to natural language processing models due to its unique agglutinative structure and limited data resources. Compared to classical natural language processing approaches, open source language models that have emerged in recent years have achieved significant success in different languages by making inferences from big data. In this context, mt5-base, mbart-50, Gemini and GPT 3.5 were used in our study. As a dataset, 240 news texts were obtained from Dunya.com. As the second dataset, 7000 Turkish economy news texts from the Hugging-Face platform were used. The two datasets were used for fine tuning approach and comparison of the models. The datasets were divided into 70% training, 10% validation and 20% testing. The models were run with a learning rate of 2e-5 and a number of cycles ranging from 20 to 100. Rouge, Bleu and BertScore metrics were used in the evaluation phase. The summarization models in the study were converted into APIs. The models published as open source on the Github platform can be used in various interfaces using API addresses.