Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.798465
Title: Text classification
Author: Singh, Gaurav
ISNI:       0000 0004 8507 5393
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
There is an abundance of text data in this world but most of it is raw. We need to extract information from this data to make use of it. One way to extract this information from raw text is to apply informative labels drawn from a pre-defined fixed set i.e. Text Classification. In this thesis, we focus on the general problem of text classification, and work towards solving challenges associated to binary/multi-class/multi-label classification. More specifically, we deal with the problem of (i) Zero-shot labels during testing; (ii) Active learning for text screening; (iii) Multi-label classification under low supervision; (iv) Structured label space; (v) Classifying pairs of words in raw text i.e. Relation Extraction. For (i), we use a zero-shot classification model that utilizes independently learned semantic embeddings. Regarding (ii), we propose a novel active learning algorithm that reduces problem of bias in naive active learning algorithms. For (iii), we propose neural candidate-selector architecture that starts from a set of high-recall candidate labels to obtain high-precision predictions. In the case of (iv), we proposed an attention based neural tree decoder that recursively decodes an abstract into the ontology tree. For (v), we propose using second-order relations that are derived by explicitly connecting pairs of words via context token(s) for improved relation extraction. We use a wide variety of both traditional and deep machine learning tools. More specifically, we used traditional machine learning models like multi-valued linear regression and logistic regression for (i, ii), deep convolutional neural networks for (iii), recurrent neural networks for (iv) and transformer networks for (v).
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.798465  DOI: Not available
Share: