• Ali Al Miman

Student thesis: Phd


This thesis aims to develop a Natural Language Inference (NLI) engine that is more robust and accurate than can be obtained through the current standard approaches. There are currently two main approaches to NLI: shallow and deep. Shallow approaches are based on lexical overlap, pattern matching, and distributional similarity [Giménez and Márquez, 2007] while deep approaches employ semantic analysis, lexical and world knowledge, and logical inference [Blackburn and Bos, 2003]. Both of these approaches have advantages and disadvantages. Shallow approaches rate, as their name suggests, superficial. They cannot make use of background knowledge to link a query to a body of knowledge because there is no way of chaining through a series of rules. Deep approaches are fragile, since they can only be employed with texts that can be accurately parsed, and there are no existing parsers that can reliably analyse the structure of arbitrary input texts. The goal here is to create an in-between approach that takes advantage of the most useful points of each of the existing approaches. The way to achieve this solution is by taking the pre-processing stage from shallow approach (dependency trees) and the inference stage from the deep approach, where an inference engine (theorem prover) has been created in a different standard. This Inference engine will obtain the required information from natural language directly, without translating inputs into any logical formula. Therefore, the goal is to apply robust logic to natural language. To achieve this goal, a theorem prover must be designed so that it can accept NL snippets. In particular, we replace the standard unification algorithm used in first-order theorem prover by an approximate algorithm for matching parse trees. We have tested this approach to NLI using rules extracted from several online dictionaries and with syllogistic patterns extracted from the FraCaS test set.
Date of Award31 Dec 2017
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorAllan Ramsay (Supervisor)


  • Inference
  • Natural language
  • understanding
  • Dependency trees
  • Automated reasoning
  • Theorem prover

Cite this