Reasoning with Natural Language

Course goals

1. Understand how formal logic can be used to express the meaning of natural language sentences;
2. Understand and get hands-on experience with reasoning over the meaning of natural language sentences using automated theorem proving;
3. Learn about the pros and cons of logic-based and deep learning-based models of reasoning;
4. Know the pros and cons of evaluation methods in Natural Language Reasoning;
5. Get acquainted with recent research on Natural Language Inference and Reasoning;

Content

Natural language reasoning is a complex task in the field of Natural Language Processing (NLP) that requires understanding the meaning of natural language expressions and recognizing semantic relations between them. The course is organized around this interdisciplinary problem and exploits methods from formal logic and deep learning.

The course discusses empirical methods of creating reasoning datasets and developing deep learning models for evaluating them on these datasets. The focus is on the Natural Language Inference (NLI) task, a classification of semantic relations between sentences. The course covers the creation (either manually, automatically, or semi-automatically) of NLI datasets and the semantic challenges they contain. Such benchmark datasets contain a mixture of logic-style deductive and common-sense inferences. They are an integral part of the system evaluation on Natural Language Understanding.

Part of the course is dedicated to logic-based approaches to natural language reasoning, in particular, how one can automatically detect that a premise sentence semantically entails or contradicts a hypothesis sentence. This includes systematic translation of the natural language meaning in some formal logic, i.e., a formal language with formal semantics, and the use of theorem-proving techniques for automated reasoning.

The first part of the course covers the introduction to the NLI task and diverse methods of solving it, with relevant group assignments, to provide students with background knowledge that will be sufficient to carry out a final group project and write a short paper. The projects most likely will be software projects where a group of students is expected to answer a research question in the domain of natural language reasoning.

The part of the evaluation also involves presenting research papers on natural language reasoning and asking questions to your classmates during their Q&A session of the presentation. Sometimes there will be interactive polls during the lectures to assess students' knowledge of already covered materials.

You might find the course interesting if:
(a) You are interested in NLP and would like to learn more about symbolic approaches (and combine them with deep learning) to model meaning in language;
(b) You like logic and linguistic semantics and would like to use them on empirical data, beyond paper & pencil exercises;
(c) The current description already sounds interesting to you.

N.B. AI students who wants to register should send an email to the course coordinator before the late registration period. This is because the course belongs to the Humanities Faculty, which has an earlier registration period for block 2 courses. Students can also register for the course in Osiris during the late registration period, typically in the 2nd half of October, but there is no guarantee for free places.

Additional information

This is the former [TLMV13020] Logic and Language course rebranded with a new, more descriptive name.

AI students who wants to register should send an email to the course coordinator before the late registration period.