REPOSITORY > RESULTS

Doctoral dissertation

Mining text-enriched heterogeneous information networks

Author(s): Miha Grčar (Author), Nada Lavrač (Supervisor)

Thesis defense date: 09.10.2015

Organization: MPŠ - Mednarodna podiplomska šola Jožefa Stefana

PID: 20.500.12556/ReVIS-15068

Views: 6 | Downloads: 0

Abstract

Text mining involves text preprocessing, modeling, knowledge discovery, visualization, and eval-uation techniques to discover, present, and evaluate knowledge from large collections of text documents (text corpora). This thesis addresses the problem of discovering knowledge from large text corpora enriched with relational links between the texts. If different relations are involved, such relational data can be described in the form of a heterogeneous information network, a generalization of the standard information network involving a single relation between the net-work nodes. If viewed from the network analysis perspective, the same problem can be interpreted as the problem of discovering knowledge from heterogeneous information networks enriched with texts. We call such networks text-enriched heterogeneous information networks or TEHINs for short.
The main hypothesis researched in the thesis is that structural/relational data, often available in real-world scenarios, can be exploited to improve the performance of algorithms employed for solving text mining tasks such as text classification and ranking. To support this hypothesis, the developed methodology should be applicable to a wide range of data analysis problems, and to large corpora of text documents accompanied with relatively large heterogeneous information networks. The main motivation for this work is due to the fact that the current general-purpose text mining tools are unable to handle texts and relational information in a common knowledge discovery setting. The goal of this thesis is thus to develop a general-purpose methodology for mining TEHINs in a typical text mining framework.
The main contribution of this thesis is the developed methodology for mining text-enriched heterogeneous information networks, named TEHmINe. It is designed as an easy-to-understand workflow, composed of well-established data and text mining components. The main functionality of the developed workflow is the projection of texts and structures into a common vector space in which knowledge discovery is performed. The methodology can be applied to a wide range of data mining problems that involve heterogeneous networks, texts, or a combination of the two data types. As an example, we show how a set of methodology building blocks can be used for very efficient centroid-based classification of vertices in heterogeneous networks and for drawing relatively large graphs and networks.
We showcase the developed methodology in two real-world use cases. In the video lecture categorization use case, we employ the TEHmINe methodology to mine a TEHIN formed out of textual data and structured information. We show that the TEHIN contains a lot of useful infor-mation and that by employing the methodology, we are able to significantly outperform the standard text mining approach. Furthermore, in the ontology querying use case, the general idea is to rank ontology entities with respect to a search query. To this end, we have adapted the proposed methodology for the task of ontology querying. We refer to the derived approach as the OntoBridge methodology. It is shown that by combining textual data and relational structure, we can significantly improve the performance of the developed ranking system over the baseline achieved with a standard text mining approach.

Attachments

Cite this work