Deep machine learning for syntactic annotation projection

master's thesis

• Author: Tina Marić
• Mentor: Goran Igaly (mentor)
• Committee member: Goran Igaly (predsjednik povjerenstva)
• Committee member: Mladen Vuković (član povjerenstva)
• Committee member: Marko Erceg (član povjerenstva)
• Committee member: Goranka Nogo (član povjerenstva)
• Granter: University of Zagreb; Faculty of Science; (Department of Mathematics); Zagreb
• Defense date and country: 2018-03-01, Croatia
• Scientific / art field, discipline and subdiscipline: NATURAL SCIENCES; Mathematics

Abstract

U ovom radu istražuje se prijenosno učenje kroz više jezika s ciljem omogućavanja sintaktičke analize jezika koji nemaju dovoljno označenih podataka za učenje. Najbolji pristupi rješavanju problema uključuju projekciju oznaka sintaktičkih ovisnosti preko paralelnih tekstova, iz jezika koji imaju mnogo označenih podataka za učenje u jezike koji imaju nedovoljno. U prvom poglavlju opisuju se osnovni pojmovi morfološkog označivanja rečenica i parsanja njihovih ovisnosnih stabala kao i način označivanja sintaktičkih ovisnosti. Prvi pristup rješavanja problema projekcije oznaka sintaktičkih ovisnosti je opisan u drugom poglavlju. Zasnovan je na algoritmu predstavljenom u znanstvenom radu Multilingual Projection for Parsing Truly Low-Resource Languages [10]. Predložene su prilagodbe algoritma koje vode poboljšanju rezultata. U trećem poglavlju predstavljena je ideja o upotrebi neuronskih mreža za projekcije oznaka sintaktičkih ovisnosti te nekoliko ideja kojima rad u budućnosti može biti unaprijeđen.

Abstract (english)

The purpose of this thesis was to explore cross-lingual transfer learning to dependency parsing, with a goal of enabling syntactic analysis for low-resource languages. The best approaches involve annotation projection: the transfer of dependency structures via parallel texts, from resource-rich to low-resource languages. In the first chapter, basic concepts of part of speech tagging and dependency parsing are described as well as the way of annotating texts. The first approach to solving an annotation projection problem is described in the second chapter. It is based on the algorithm presented in the paper Multilingual Projection for Parsing Truly Low-Resource Languages [10]. We propose the way of adjusting the existing algorithm which leads to the improvement of results. In the third chapter, the idea how to use neural networks for annotation projection is presented, and also some of the ideas how the work done in this thesis can be extended in the future.

• Keywords: prijenosno učenje kroz više jezika; sintaktička analiza jezika; sintaktička ovisnost; morfološko označivanje rečenica; parsanje ovisnosnih stabala
• Keywords (english): cross-lingual transfer learning; syntactic analysis; low-resource languages; annotation projection; speech tagging; dependency parsing
• Language: croatian
• URN:NBN: urn:nbn:hr:217:780941
• Study programme: Title: Computer Science and Mathematics; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra računarstva i matematike (magistar/magistra računarstva i matematike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2018-09-06 10:29:13