Institute for Web Science and Technologies · Universität Koblenz - Landau
Institute WeST

Semantically Enhanced and Minimally Supervised Models For: Ontology Enrichment, Text Classification and Document Recommendation

[zur Übersicht]
Wael Alkhatib

This dissertation proposes new approaches, empowered by the combination of NLP , lexical-semantic resources and deep learning structures, for addressing the challenges in three major tasks concerning data management on the web. The tasks in hand are, namely ontology construction and enrichment, multi-label text classification and recommendation systems. Methods proposed for tackling these tasks are challenged by the data variety spanning from diverse data sources, contextual information around data, to structures and formats. Also the high dimensionality of data and feature space impose new challenges on designing and training the models tackling the different NLP tasks. We will address the main challenges related to these three main tasks, review the state-of-the-art techniques in order to highlight the current research gaps and propose new models to overcome the aligned challenges.

Research goals and contributions:

  • Designing minimally supervised methods for ontology construction and enrichment (addressing: The intensive reliance on complicated feature engineering and linguistic analysis for ontology construction)
    • DeepOnto.KOM: a minimally supervised ontology learning system using convolutional neural networks and knowledge-bases
  • Designing semantically enhanced models to improve the state-of-the-art in text classification (addressing: The high dimensionality of feature space, label imbalance and training overhead in multi-label text classification)
    • Designing semantic-based feature selection methods for text classification
    • Analyzing the feasibility of using deep learning for multi-label text classification ( small dataset of long documents)
    • Designing an ontology-based training-less classifier
  • Personalized citation recommendation systems (addressing: Adaptivity and personalization in recommendation system of documents, we focus on citation recommendation)
    • Designing a Joint Model of Deep Semantic Similarity Model (DSSM) and Bibliographic Information for Personalized Citation Recommendation
    • Using Adversarially Regularized Graph Autoencoder for Personalized Citation Recommendation

21.11.19 - 10:15
B 016