Extending the recurrent neural network model for improved compositional modelling of text sequences

doctoral thesis

• Title (croatian): Proširenje modela povratnih neuronskih mreža za poboljšano kompozicijsko modeliranje tekstnih sljedova
• Author: Martin Tutek
• Mentor: Jan Šnajder (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Committee member: Goran Glavaš (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; (Department of Electronics, Microelectronics, Computer and Intelligent Systems); Zagreb
• Defense date and country: 2022-07-22, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Universal decimal classification (UDC): 004 - Computer science and technology. Computing. Data processing

Abstract

The thesis focuses on exploring extensions to the recurrent neural network (RNN) algorithm for natural language processing (NLP) in terms of improving its capabilities of semantic composition, investigating the possible benefits of leveraging multi-prototype word representations and improving its overall interpretability. While RNNs have received a strong competitor in form of the Transformer model, both approaches to processing natural language sequences possess their own set of issues. This thesis investigates methods of inducing sparsity in neural networks in order to learn shared sense representations and also tackles the problem of semantic composition in recurrent networks. The thesis introduces a novel approach for building recursive representations of language which is better suited to the hierarchical phrasal structure of language. The original scientific contributions of the thesis are: 1. Empirical analysis of the convergence of recurrent neural network algorithms on text sequence modeling tasks with respect to different input word representations; 2. An algorithm for learning word representations serving as input to text processing models based on contextualized word representations; 3. An extension of the recurrent neural network model for processing text sequences with mechanisms for processing linguistic phenomena such as polysemy, semantic composition, and coreference.

Abstract (croatian)

Središnje pitanje ovog doktorskog rada su proširenja algoritma povratnih neuronskih mreža (RNN) za obradu prirodnog jezika (NLP) u vidu poboljšanja semantičke kompozicije, nadogradnje kroz korištenje višeprototipnih reprezentacija riječi i sveukupno poboljšanje interpretabilnosti algoritma. Usprkos modelu Transformer koji predstavlja snažnu konkurenciju povratnim neuronskim mrežama, oba pristupa analizi teksta posjeduju svoje nedostatke. U okviru doktorskog rada istraženi su pristupi ugrađivanja rijetkosti u neuronske mreže s ciljem učenja dijeljenih reprezentacija značenja te problem semantičke kompozicije u povratnim mrežama. Doktorski rad uvodi novi, rekurzivni pristup građenju reprezentacija teksta prikladniji za specifičnosti jezika. Znanstveni doprinosi ovog doktorskog rada su: 1. Empirijska analiza konvergencije povratnih neuronskih mreža na zadacima analize teksta u ovisnosti o različitim ulaznim reprezentacijama riječi; 2. Algoritam učenja reprezentacija riječi baziran na kontekstualiziranim reprezentacijama riječi, s ciljem njihovog korištenja kao ulaza modelima obrade teksta; 3. Proširenje algoritma povratnih neuronskih mreža za obradu prirodnog jezika s naglaskom na obradu jezičnih fenomena poput polisemije, semantičke kompozicije i koreferenciranja.

• Keywords: word representations; multiprototype representations; semantic composition; interpretability; recurrent neural networks; natural language processing
• Keywords (croatian): značenje riječi; povratne neuronske mreže; algoritam pozornosti; transparentnost dubokih neuronskih mreža; obrada prirodnog jezika
• Language: english
• URN:NBN: urn:nbn:hr:168:818357
• Promotion: 2022
• Study programme: Title: Computer Science; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti (Doktor znanosti)
• Type of resource: Text
• Extent: ix, 93 str. : graf. prikazi ; 30 cm. + CD-ROM.
• File origin: Born digital
• Access conditions: Open access
• Created on: 2022-11-24 09:36:13