PhD Defense Mr. L. Abzianidze, MSc
Title: A Natural Proof System for Natural Language
Supervisor: Prof. J.M. Sprenger
Co-supervisor: Dr. R.A. Muskens
Natural language represents a device for communication where
its phrases carry meanings. The same meaning can be expressed with several
natural language phrases. Capturing this variation of natural language is a key
for making machines understand natural language.
To account for this problem,
the thesis proposes a novel automatized method of modeling semantic relations
between natural language phrases. Given two phrases, a premise and a
conclusion, e.g., (1) "this dissertation is boring" and (2) "not
all PhD theses are interesting" respectively, the method can prove that
(2) is entailed from (1). It can also recognize phrases with semantically
contradicting or neutral meanings.
Instead of translating the semantics of natural language phrases in some formal language, which in itself is a complex task, the method employs a proof technique used for formal languages and adapts it to natural language. We automatized the method as a reasoning system and showed that on standard test datasets it achieves high scores comparable to the state-of-the-art results. In contrast to most existing reasoning systems, our system is able to process multiple premises. Advances in natural language understanding directly contribute to automated question answering and semantic search (opposed to a primitive word matching search).
Location: Cobbenhagen building, Auditorium (access via Koopmans building)