Athens University Economics and Business (AUEB)
Large scale hierarchical text classification
Abstract
dc:descriptionHierarchies are becoming increasingly popular for the organization of documents, particularly on the Web. Web directories, such as the Υahoo! Directory and the Dmoz Directory, are typical examples. Along with their widespread use, comes the need for automated classification of new documents to the classes of the hierarchy. In this thesis, we call this problem Large Scale Hierarchical Text Classification. It is a large scale classification problem, since the classes are thousands and the documents can be hundreds of thousands or even millions.It is also hierarchical, since the classes are connected by parent-child relations.An important issue in hierarchical classification is the evaluation of different classification algorithms, an issue which is complicated by the hierarchical relations among the classes.Several evaluation measures have been proposed for hierarchical classification using the hierarchy in different ways without however providing a unified view of the problem. In this thesis, we study the problem of evaluation in hierarchical classification by analysing and abstracting the key components of the existing performance measures. We also propose two alternative generic views of hierarchical evaluation and introduce two corresponding novel measures. The proposed measures, along with the state-of-the-art ones, are empirically tested on three large datasets from the domain of text classification. The empirical results illustrate the limitations of existing approaches and how the proposed methods overcome most of them across a range of cases.We then focus on the simplest case of large scale hierarchical text classification, where the hierarchy is a tree and each document belongs in a single leaf class of the hierarchy.A popular method of hierarchical classification is cascade classification, which greedily traverses the hierarchy from the root to the predicted leaf. In order to perform cascade classification, a classifier must be trained for each node of the hierarchy, but in the upper levels the number of features can be prohibitively large. It is therefore desirable to reduce the dimensionality of the feature space at these levels. We examine the computational feasibility of the most common dimensionality reduction method (Principal Component Analysis) for this problem, as well as the computational benefits that it provides for cascade classification and its effect on classification accuracy. Furthermore, we propose a probabilistic cascading approach, which outperforms the traditional greedy cascade, by making better use of the probabilities estimated by the classifiers.Finally, we consider a more complex domain, known as biomedical semantic indexing, where biomedical documents have to be classified to the classes of a large biomedical taxonomy. This domain is more complex in that the taxonomy is a directed acyclic graph, rather than simply a tree, the same document may belong in several classes, and the correct classes are not necessarily leaves of the taxonomy. We examine the use of dense word vectors, also known as word embeddings, as a method of dimensionality reduction. We consider several efficient approaches for the transition from dense word vectors to vectors that represent entire texts, proposing a simple weighted centroid approach that is suitable for this domain. We show that by adopting this approach, hierarchical text classification algorithms become sufficiently scalable for large scale semantic indexing, without being less effective than then usual bag of words representation. We experiment with flat and hierarchically expanded k-nearest neighbor classifiers that employ our centroid representations of article abstracts, examining the effect of various parameters. We also present a high precision system that can be combined with the widely used Medical Text Indexer (MTI) system of the National Library of Medicine to improve its performance.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2015
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kosmopoulos, Aris
- Κοσμόπουλος, Άρης
Subjects
dc:subject × 8Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/36242