Modeli dubokog učenja za ekstraktivno sažimanje dokumentata

undergraduate thesis

• Title (english): Deep Learning Models for Extractive Document Summarization
• Author: Juraj Vladika
• Mentor: Jan Šnajder (mentor)
• Committee member: Jan Šnajder (predsjednik povjerenstva)
• Committee member: Domagoj Alagić (član povjerenstva)
• Committee member: Martin Tutek (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2019-07-12, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Područje obrade prirodnog jezika nastalo je kao spoj lingvistike, računarske znanosti i umjetne inteligencije, a bavi se analizom i sintezom ljudskog jezika u svim oblicima. U zadnje vrijeme najčešće se problemi unutar područja rješavaju modelima strojnog i dubokog učenja. Ekstraktivno sažimanje dokumenata jedan je od problema u području, a cilj mu je izraditi kratak prikaz sadržaja teksta dokumenta birajući pridom najreprezentativnije rečenice. U ovom je radu predloženo nekoliko matematičkih modela za rješavanje problema. Glavni od njih temeljio se na povratnoj neuronskoj mreži, točnije njezinim arhitekturama LSTM i GRU. Dan je pregled teorijskih osnova potrebnih za razumijevanje rada umjetnih neuronskih mreža te prikaza teksta u računalu s pomoću vektorske reprezentacije riječi. Testirana je važnost značajki koje su određivale prikladnost uključivanja rečenica dokumenta u sažetak kao što su sadržajnost, isturenost, novost te relativni i apsolutni položaj. Dana je analiza i pregled rezultata nekoliko varijanti modela te usporedba s referentnim (osnovnim) modelom. Primijećena je značajna prednost modela dubokog učenja nad tradicionalnim modelima.

Abstract (english)

Natural language processing is a field that emerged as a mixture of linguistics, computer science and artificial intelligence, and it deals with analysis and synthesis of human language in all of its forms. Lately the most common method of solving problems in the field was using models of machine learning and deep learning. Extractive document summarization is one of the problems in the field, and its goal is to make a short overview of document's content by chosing the most representative sentences. In this thesis several mathematical models for solving this problem were suggested. The main one was based on a recurrent neural network, more precisely its LSTM and GRU architecture. An overview of theoretical foundation needed for understanding the work of artificial neural networks and representation of words using vectors was given. The importance of features that determined the likelihood of including a hiven sentence in the summary was tested, such as content, novelty, salience, and relative and absolute position. A significant imporvement of models based on deep learning in comparison to traditional models was noticed.

• Keywords: strojno učenje; obrada prirodnog jezika; neuronske mreže; LSTM; TFIDF; GRU; povratne neuronske mreže; klasifikacija rečenica
• Keywords (english): machine learning; natural language processing; neural networks; LSTM; TFIDF; GRU; recurrent neural networks; sentence classification
• Language: croatian
• URN:NBN: urn:nbn:hr:168:232052
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2020-01-07 22:53:20