Dr Martha Lewis
BA (Hons), MSc, MRes, PhD
Current positions
Contact
Press and media
Many of our academics speak to the media as experts in their field of research. If you are a journalist, please contact the University’s Media and PR Team:
Research interests
Quick links: https://marthaflinderslewis.github.io/; Google Scholar; DBLP
How are humans able to combine concepts to form new ones? Why do unexpected or 'emergent' attributes arise in these combinations? I use computational linguistics, conceptual spaces theory, quantum theory, and category theory to try and answer questions like these.
Before Bristol, I held a Veni fellowship at the ILLC at the University of Amsterdam. I was a postdoc in the Quantum Group in the Department of Computer Science, University of Oxford, under the project ‘Algorithmic and Logical Aspects of Meaning’ funded by AFOSR. I did my PhD at the University of Bristol, in the Bristol Centre for Complexity Sciences, and before that the Evolutionary and Adaptive Systems (EASy) MSc at the University of Sussex.
Research InterestsI use interdisciplinary approaches to modelling language and concepts. My research splits into three broad areas - compositional approaches to language and meaning, applications of quantum theory in modelling language, and evolutionary approaches to language.
Compositional approaches
Humans are able to generate and to understand completely novel combinations of concepts. How can we do this? I look at ways to integrate symbolic composition with statistical or fuzzy representations of concepts. Examples are: integrating grammar with conceptual spaces, a hierarchical approach to concept composition, and logical structure in vector spaces
I have also recently started working on metaphor and my student Xiaoyu Tong developed a comprehensive metaphor paraphrase dataset which we are currently extending.
Applications of quantum theory
The compositional framework I work in has its roots in quantum theory, and there is a wide range of work that examines quantum approaches in cognitive science. I provide a thorough review of this research. Density matrices are a notion taken from quantum theory that I use to model hyponymy and entailment within a compositional vector-based model of meaning. The same framework is used to model lexical ambiguity where we show how to learn density matrix representations of words from a corpus and how these representations disambiguate in the process of composition.
Evolutionary approaches
A key aspect of human concept use is that concepts and words evolve over time. I have examined how shared concepts can emerge in a community of artificial agents, and the impact of linguistic hedges and concept conjunction on the resulting concepts.
Publications
Selected publications
11/09/2020Towards logical negation for compositional distributional semantics
IfCoLoG Journal of Logics and their Applications
Compositional hyponymy with positive operators
International Conference on Recent Advances in Natural Language Processing in a Deep Learning World, RANLP 2019 - Proceedings
Interacting Conceptual Spaces I
Conceptual Spaces: Elaborations and Applications
Recent publications
01/01/2023EXTRACT: Explainable Transparent Control of Bias in Embeddings
Compositional Fusion of Signals in Data Embedding
Compositional Fusion of Signals in Data Embedding
Quantum Computing and Cognitive Simulation
Quantum Computing in the Arts and Humanities: An Introduction to Core Concepts, Theory and Applications
Modelling Lexical Ambiguity with Density Matrices
Proceedings of the 24th Conference on Computational Natural Language Learning
Modelling the interplay of metaphor and emotion through multitask learning
EMNLP-IJCNLP 2019 - 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint Conference on Natural Language Processing, Proceedings of the Conference