Learning to Rank for Text Summarization: Revisiting the Features and Methods for Turkish Bank Documents

This study revisits features and learning to rank algorithms for domain specific Turkish text summarization. In this study, we elaborate on the effect of sentence-level and word-level features for text summarization using three learning to rank algorithms that belong to different categories. To this aim, we annotated a corpus of Turkish bank transcripts, and designed three experimental settings. In the first setting, we investigate the performance of sentence-level features, in the second word-level features, and in the last the union of both features in Turkish text summarization. The experimental results show that word-level features perform poorer compared to the sentence-level features.