Sporathe knowledge network
ES
Universidad Nacional de Colombia

Bogotá - Ingeniería - Maestría en Ingeniería - Ingeniería de Sistemas y Computación · 2020

Sistema para la búsqueda de documentos biomédicos basado en aprendizaje profundo

This dissertation explored the use of strategies based on deep learning to address the information retrieval task in the biomedical domain. We implemented a system composed of three main phases. First, the Recovery Phase, where for a given question we aimed to extract a subset of the most relevant documents out of indexed documents. We obtained the best result using the DFR (Divergence From Randomness) algorithm with the model base I (f) (Inverse term frequency model), the posterior effect L (an information gain model based on Laplace's law of succession) and used H2 normalization (frequency is inversely related to length). The second phase seeked to filter the aforementioned subset of the most relevant documents using deep learning strategies. The use of Word Mover's Distance and Document Centroid was proposed as a baseline and they were compared with two models based on BERT. Here we obtained the best result out of the pre-trained BERT model for Question Answering, which allowed filtering documents that do not contain sufficient evidence to answer the associated question. In the last phase of the system, two document reordering strategies were explored. The first one is based on reordering the documents by passages that combine two types of representation for pairs (of the type question, passage) using textual representation and conceptual representation. This, to take advantage of the use of knowledge generated by experts in ontologies, for example. The second strategy was based on Metric Learning, we used a network of three inputs (question, answer +, answer−) and again took advantage of the use of structured information from biomedical concepts that were extracted from QuickUMLS [Soldaini and Goharian, 2016]. The proposed systems were evaluated using the test sets presented at the BioASQ 2019 competition and were compared against the best solution to a given task in each of the sets. We concluded that the system that obtains the best results for MAP (Mean Average Precision) is the one made up of DFR for initial retrieval, BERT for QA for document filtering and Metric Learning using conceptual information for reordering.

Full text 64 pages with text out of 69

Read the full thesis Record in the repository

Its library's thesaurus