Klasifikacija bioloških anotacija na velikoj skali koristeći reprezentacije riječi izvedene iz dokumenata biomedicinske znanstvene literature

master's thesis

• Title (english): Classification of Large-Scale Biological Annotations Using Word Embeddings Derived from Corpora of Biomedical Research Literature
• Author: Adriano Baćac
• Mentor: Mile Šikić (mentor)
• Committee member: Mile Šikić (predsjednik povjerenstva)
• Committee member: Marko Subašić (član povjerenstva)
• Committee member: Krešimir Križanović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2017-07-10, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Naučeni su vlastiti Word2vec i GloVe modeli prikaza riječi za znanstvenu literaturu u području biomedicine, kao i tri klasifikacijske metode za diskriminaciju fenotipova, dvije temeljene na agregaciji vektorskog prikaza riječi, i jedna na rekurentnim neuronskim mrežama. Prikazi riječi su trenirani na velikom korpusu znanstvenih članaka i njegovim tematski specifičnim podskupovima. Rezultati klasifikacije su testirani na 6 izvora dokumenata. Pokazano je da Word2vec postiže bolje rezultate kada je treniran na tematski specifičnom podskupu koji je sačinjen od 4.9% od ukupnih članaka nego kada je treniran na cijelom korpusu. Korištenje rekurentnih neuronskih mreža imalo je problema s prenaučenosti, moguće zbog predugačkih dokumenata ili premalog skupa za učenje. Iako predloženi modeli nisu bili bolji od stroja potpornih vektora koristeći prikaz vreće riječi, pokazano je da korištenje agregacijskih metoda uz bazni model povećava količinu ispravne klasifikacije manjinske klase kod nekih fenotipova za oko 10%.

Abstract (english)

Custom Word2vec and GloVe embeddings for scientific literature in the biomedical domain were trained, as well as three classification methods for discriminating phenotype traits, two of which were based on aggregating word embeddings and one on recurrent neural networks. Word embeddings were trained on a large corpus of scientific articles and its more subject-specific subsets. Classification performance was tested on 6 document sources. It was shown that Word2vec achieves better performance when trained on a subject-specific subset corpus comprised of 4.9% articles, than when trained on the entire corpus. Using recurrent neural networks had an overfitting problem, possibly because the documents were too long or the training set too small. Although the proposed models did not outperform support vector machine using bag-of-words, it was shown that using the aggregation methods alongside the baseline model increases the amount of correctly classified minority class in some phenotype traits by around 10%.

• Keywords: vektorski prikaz riječi; Word2vec; GloVe; RNN; LSTM; klasifikacija fenotipova; specifičnost korpusa
• Keywords (english): word embedding; Word2vec; GloVe; RNN; LSTM; phenotype classification; corpus specificity
• Language: english
• URN:NBN: urn:nbn:hr:168:032239
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2019-03-13 18:43:47