Extractive Text Summarization Using Deep Learning

U ovom radu dan je pregled nad područjem obrade prirodnog jezika kroz opis odabranih modela dubokog učenja korištenih kroz povijest. Opisani su postojeći pristupi za semantičko sažimanja teksta, a naglasak je stavljen na ekstraktivno sažimanje. Kao praktični ishod rada, ostvaren je web sustav za ekstraktivno sažimanje teksta. Dan je opis arhitekture i programskog ostvarenja sustava. Sustav se zasniva na modelu BERT i algoritmu grupiranja k-sredina. Ekstraktivni sažetak se stvara slijedom operacija. Ulazni tekst se prvo dijeli na rečenice. Pomoću modela BERT se računaju vektorski prikazi rečenica. Vektorski prikazi rečenica se grupiraju algoritmom k-sredina, a zatim se odabiru rečenice koje ulaze u sažetak. Ostvareni sustav vrednovan je mjerama ROUGE na skupovima podataka CNN/Daily Mail i Newsroom.This paper provides an overview of the natural language processing field through a description of selected deep learning models used throughout history. Existing approaches to semantic text summarization are described, with an emphasis on extractive text summarization. As a practical outcome of the paper, a web system for extractive text summarization is realized. A description of the architecture and software implementation of the system is given. The system is based on the BERT model and the k-means clustering algorithm. An extractive summary is created using a sequence of operations. The input text is first divided into sentences. Using the BERT model, vector representations of sentences are calculated. Vector representations of sentences are clustered using the k-means algorithm, followed by the selection of the sentences that enter the summary. The realized system is evaluated using the ROUGE metric on CNN/Daily Mail and Newsroom datasets