Deep learning for noise-tolerant RDFS reasoning

Abstract

This thesis aims to provide a stepping stone towards bridging the Neural-Symbolic gap, specifically targeting the Semantic Web field and RDFS reasoning in particular. This is accomplished through layering Resource Description Framework (RDF) graphs and encoding them in the form of 3D adjacency matrices. Each layer layout in the 3D adjacency matrices forms what we termed as graph word. Every input graph and its corresponding inference are then represented as sequences of graph words. The RDFS inference becomes equivalent to the translation of graph words that is achieved through neural network translation.

Recent research work on semantic reasoning with noise-tolerance focuses on type inference and does not aim for full RDF Schema (RDFS) reasoning. This thesis documents a novel approach that takes previous research efforts in noise-tolerance in the Semantic Web to the next level of full RDFS reasoning by utilizing advances in deep learning research. Deep learning techniques—even though robust to noise and very effective in generalizing across a number of fields including machine vision, natural language understanding, speech recognition etc.–require large amounts of data of low-level representation rather than “symbolic representations used in knowledge representation” ([3]).

In 2010, Hitzler and van Harmelen called for questioning the model-theoretic semantic reasoning and investigation of machine learning (ML) for semantic reasoning [1] because ML techniques are more robust to noisy data. Four years later, a position paper about machine learning on linked data [2] considered research efforts to couple both fields to be still “disappointing”.

Since the introduction of the Semantic Web vision in 2001 as an extension to the Web, where machines can reason about the Web content, the main research focus in semantic reasoning was on the soundness and completeness of the reasoners. While these reasoners assume the veracity of the input data, the reality is that the Web of data is inherently noisy.

The evaluation confirms that deep learning can in fact be used to learn RDFS rules from both synthetic as well as real-world Semantic Web data while showing noise-tolerance capabilities as opposed to rule-based reasoners.

This challenge constitutes what is referred to as the Neural-Symbolic gap.