Our alumni

Tom Green

PhD Thesis: Using NLP to Resolve Mismatches Between Jobseekers and Positions in Recruitment

Supervisor: Dr Diana Maynard

Industry partner: Tribepad

Cohort: 1

Thesis examiners: Dr Mark Stevenson (University of Sheffield) and Dr Rudy Arthur (University of Exeter)

Viva date: 29 November 2023

Current employer: Department for Work and Pensions, Government of the UK


Biography

Tom joined the CDT with a background in Psychology and Product Management. Keen to apply NLP techniques to live settings, Tom collaborated with recruitment software provider TribePad to investigate job recommendation and candidate ranking algorithms. He was supervised by Diana Maynard and Chenghua Lin. 


PhD Summary 

Recruiting through online portals has seen a dramatic increase in recent decades and it is challenging for jobseekers to evaluate the overwhelming amount of data to efficiently identify positions that align with their skills and qualifications. This research addresses this issue by investigating automatic approaches that leverage recent developments in Natural Language Processing (NLP) that search, parse, and evaluate the often unstructured data in order to find appropriate matches. 


Impact

This research informs the early stages of the recruitment process when candidates are searching for jobs to submit an application for that align with their particular skills and experience. This research improves upon simplistic methods that are commonly used, by leveraging state-of-the-art NLP techniques that are able to learn from past application outcomes, and considering the semantic content of salient entities in candidate profiles and job descriptions. Jobseekers will benefit from this research in terms of a more appropriate ranking of the available jobs which will reduce the amount of data to review to find appropriate matches, and recruiting agents will benefit from this research in terms of more appropriate ranking of candidates which will reduce the amount of data to review to find appropriate matches.    


Links

Will Ravenscroft

PhD Thesis: Speech Separation in Noisy Reverberant Acoustic Environments

Supervisor: Professor Thomas Hain

Industry partner: 3M Health Information Systems

Cohort: 1

Thesis examiners: Dr Yoshi Gotoh (University of Sheffield) and Prof Patrick Naylor (Imperial College London)

Viva date: 2 May 2024

Current employer: Bose


Biography

I got my start in academia in electronic engineering and mathematics. I also had a keen interest in audio signal processing. I came to Sheffield to dive deeper into the machine learning of audio processing with respect to speech. My thesis was focused on speech enhancement and separation technologies with a view to multi speaker speech recognition.


PhD Summary 

Speech separation has seen significant advancements in recent years due to advanced deep learning techniques. However speech and reverberation still degrade the performance of these models. In this thesis, some inherent assumptions about the design of these models is challenged and a number new techniques for designing and training these models are proposed. These techniques result in improved model generalization, robustness and computational efficiency.


Impact

This thesis demonstrated the massive redundancies in some common approaches to speech separation research. It showed that performance of lightweight models can by improved by more intelligent analysis instead of massively increasing computational requirements. It also showed the same from the opposite end where large models contained significant redundancies because researchers had as yet failed to challenge preconceived notions about other model architectures.


The main benefit here is in the reduced training requirements for these models. Firstly, this is better for the environment which is better for society as a whole. Secondly, it demonstrates to researchers and practitioners they can save time and cost if they design their models more intelligently and take the time to challenge preconceived notions that have dominated this research field for many years now. 


Links

Meg Thomas

PhD Thesis: A multidisciplinary investigation of conversation and disfluencies in cognitive decline

Supervisors: Dr Traci Walker and Professor Heidi Christensen

Industry partner: Apple

Cohort: 1

Thesis Examiners: Dr Stuart Cunningham (University of Sheffield) and Dr Leendert Plug (University of Leeds)

Viva date: 13 June 2024

Current employer: Forensic Voice Centre, York, UK

Cohort 1 Student Representative


Biography

Dr Megan Thomas is a forensic speech and audio analyst at the Forensic Voice Centre. She recently completed her PhD in speech and language technology from the University of Sheffield where she was investigating the viability of using speech disfluencies as an early indicator of cognitive decline. Before this, she completed an MSc in Forensic Speech Science at the University of York, and a BSc in Arabic and Linguistics from the University of Westminster. 


PhD Summary 

This thesis examined cognitive decline, with a focus on neurodegenerative dementias such as Alzheimer’s disease, which often present with language impairments. Speech was explored as a non-invasive biomarker, with disfluencies like pauses offering insights into disease progression. The study investigated the diagnostic utility of disfluency analysis and task-specific speech elicitation. Advances in machine learning have enabled automatic cognitive decline classification systems, but challenges in generalisation and transparency persist. By integrating disfluency features, this research enhances the accuracy of these systems. Additionally, conversational analysis was explored to distinguish dementias from mild cognitive impairment. 


Impact

This PhD allowed me to explore my interests in speech and language technology, and introduced me to the key concepts of machine learning and artificial intelligence. The CDT gave me opportunities to work with industry leaders such as Apple and Zoo Digital, which gave me invaluable insight into work outside of academia. 


Links

Peter Vickers

PhD Thesis: Navigating Multimodal Complexity: Advances in Model Design, Dataset Creation, and Evaluation Techniques

Supervisor: Professor Nikos Aletras and Dr Loïc Barrault

Industry partner: Amazon

Cohort: 1

Thesis Examiners: Dr Mark Stevenson (University of Sheffield) and Prof Yulan He (King’s College London)

Viva date: 7 June 2024

Current employer: AI Solutions Hub, Northeastern University, USA


Biography

Peter Vickers is a Computer Scientist specialising in Natural Language Processing and Multimodal AI. He completed his Ph.D. at The University of Sheffield, focusing on augmenting language models with multimodal information. His research enhances models' capabilities for tasks like Visual Question Answering and Text-to-Image Retrieval. Peter's Ph.D. was dual-funded by the UKRI Centre for Doctoral Training in Speech and Language Technologies and an Amazon Studentship Grant.


Currently, Peter works as a Data Scientist at the AI Solutions Hub, Northeastern University, applying cutting-edge Machine Learning research to business problems. He has participated in JSALT summer workshops and interned as an Applied Scientist at Amazon UK. Peter holds an MSc in Computer Science with Speech and NLP from the University of Sheffield and a BA in English Language and Literature from Magdalen College, Oxford.


Outside of his academic pursuits, Peter is an avid backcountry skier, with experience in Norway and the High Arctic. He lead the 2024 UK Stauning Alps expedition to Eastern Greenland.


PhD Summary 

My thesis explores how artificial intelligence systems can integrate information from diverse data types, ranging from highly structured knowledge graphs to unstructured images and text. We investigate this through three tasks: visual question answering, eye-tracking prediction, and citation recommendation. Our research focuses on developing novel multimodal AI models, creating diverse and diagnostic datasets, and improving evaluation metrics for complex classification tasks. We introduce the concept of "knowledge density" to categorize different data modalities and examine how models perform when combining information across this spectrum. Our work aims to advance multimodal AI systems' ability to reason with heterogeneous data sources for real-world applications.


Impact

The impact of this PhD project could be significant in several ways:


Links

Danae Sánchez Villegas

PhD Thesis: Beyond Words: Analyzing Social Media with Text and Images

Supervisor: Professor Nikos Aletras

Industry supporter: Emotech

Cohort: 1

Thesis Examiners: Dr Carolina Scarton (University of Sheffield) and Prof Andreas Vlachos (University of Cambridge)

Viva date: 23 October 2023

Current employer: University of Copenhagen, Denmark


Links

Sebastian Vincent

PhD Thesis: Context-Based Personalisation in Neural Machine Translation of Dialogue

Supervisor: Dr Carolina Scarton

Industry partner: ZOO Digital

Cohort: 1

Thesis Examiners: Prof Nikos Aletras (University of Sheffield) and Dr Alexandra Birch-Mayne (University of Edinburgh)

Viva date: 28 November 2023

Current employer: ZOO Digital


Biography

I was born in Poland, moved to study in the UK in 2016. I have a BSc degree in Computer Science and AI. My PhD thesis is entitled "Context-Based Personalisation in Neural Machine Translation of Dialogue”. After the PhD I became an AI Research Scientist at ZOO Digital. 


PhD Summary 

My thesis explores personalisation of neural machine translation through the use of extra-textual information such as character and production metadata, in the domain of scripted dialogue. It puts forward a novel framework for working with such information and described an evaluation scheme for capturing how specific the queried translation (human or machine sourced) are to the provided context.


Impact


Links

Guanyu Huang

PhD Thesis: My Fair Robot: Shaping a Mismatched Conversational Partner via Affordance Design

Supervisor: Professor Roger K Moore

Cohort: 2

Thesis Examiners: Dr Chaona Chen (University of Sheffield) and Dr Mary Ellen Foster (University of Glasgow)

Viva date: 19 December 2024


Viva passed with minor corrections


Biography

My journey as a researcher has been guided by a fascination with communication, particularly in the areas of human-robot interaction, sociality and inclusivity. I have been fortunate to work on designing affordances for social robots that promote meaningful, human-centred interactions. My background in applied linguistics, with a focus on sociolinguistics and second language acquisition, has also allowed me to explore how second language teaching (SLT) methodologies can enhance language acquisition research.


Much of my perspective has been shaped by my experiences as a foreign language user and teacher. These experiences taught me how easily interactions between mismatched partners can falter, and inspired me to seek ways to bridge these gaps. Drawing on insights from different disciplines, I aim to contribute to user-centred, value-driven and evidence-based approaches that improve communication and foster better understanding - whether between humans or between humans and robots.


I am particularly passionate about exploring human-centred, technology-driven solutions that address real-world challenges in communication and interaction. By focusing on the needs and values of individuals, I strive to develop technologies that not only bridge gaps, but also promote inclusivity and understanding in meaningful ways.


PhD Summary

My thesis examines how the design of a social robot's appearance, sound and behaviour - its affordances - shapes human interaction. It investigates communication mismatches, anthropomorphic features, and proposes 'honest affordance design' to align robot capabilities with user expectations. Results show that while human likeness increases likeability, it is not the most influential factor; users prefer robots that embrace their non-human identity. Contextual adaptability in affordance design is crucial to meet expectations of warmth and competence in different social roles. This research provides theoretical, experimental and computational insights for transparent and consistent robot design to enhance user experience.


Impact

The results of my PhD research have implications for improving the design and use of social robots in various domains. By addressing communication mismatches and proposing 'honest affordance design', my work contributes to the creation of robots that align their external signals with their internal capabilities. This alignment increases user trust, comfort, and engagement, ensuring that robots are better equipped to meet human expectations in diverse contexts.


My research challenges the overemphasis on anthropomorphism and provides evidence-based guidelines that prioritise transparency and contextually adaptive affordances. These insights are particularly valuable for industries developing robots for healthcare, education, customer service, and other social roles, where robots must perform effectively while fostering positive human interactions. The proposed frameworks, such as the Theory of Lenses and the Interactive Lemon model, provide a conceptual guidance to improve user experience, contributing to the advancement of human-centred robotics. The reach of this research spans multiple sectors and stakeholders, including robot manufacturers, policy makers and the general public. Beyond industry, the public will benefit from a better understanding of how to interact with and use robots in everyday life, fostering acceptance and reducing fear of automation.


Overall, this research has the potential to shape the development of more trustworthy, effective and widely accepted social robots, enriching the lives of individuals and supporting innovation across sectors.


Links

Rhiannon Mogridge

PhD Thesis: An exemplar-informed approach to Speech and Language Tasks

Supervisor: Dr Anton Ragni

Cohort: 2

Thesis Examiners: Professor Roger K Moore (University of Sheffield) and Dr Sam Kirkham (Lancaster University)

Viva date: 7 November 2024

Industry partner: Toshiba


PhD Summary

The field of machine learning has drawn heavily from the fields of psychology and neuroscience, in particular in the development of neural network architectures, which are based on simplified versions of structures in the brain. While effective for many tasks, neural networks do not, in general, incorporate any way of storing specific experiences, instead using training data to parameterise a model, and then discarding the training date prior to inference. We explore an alternative option: a simple, explainable model from the field of human psychology called Minerva 2, which uses previously seen examples to perform classification or regression. By comparing Minerva 2 with neural networks, we demonstrate that Minerva 2 is in fact a neural network itself, with parameters taken directly from the data, rather than being trained by backpropagation. We propose new architectures, which are based on Minerva 2 and incorporate both a memory of previous examples and parameterisation that allows model flexibility. We show that feature representation is crucial for this type of model, which might explain the lack of representation of this type of model in the literature. Speech and text representations have improved rapidly in recent years, however, and if this trend continues, simple, interpretable models such those proposed here will become more competitive. As evidence of this, we use high quality speech representations in conjunction with a Minerva-based model to demonstrate state-of-the-art performance on a speech intelligibility task.


Links

George Close

PhD thesis: Perceptually Motivated Speech Enhancement

Supervisor: Dr Stefan Goetze

Cohort: 2

Thesis Examiners: Professor Roger K Moore (University of Sheffield) and Prof. Dr.-Ing. Timo Gerkmann (University of Hamburg)

Viva date: tbc

Industry partner: Toshiba


PhD Summary

Speech Enhancement (SE) is a vital technology for online human communication. Applications of Deep Neural Network (DNN) technologies in concert with traditional signal processing approaches to the task have revolutionised both the research and implementation of SE in recent years. However, the training objective of these Neural Network Speech Enhancement (NNSE) systems generally do not consider the psychoacoustic processing which occurs in the human auditory system. As a result, enhanced audio can often contain auditory artefacts which degrade the perceptual quality or intelligibility of the speech. To overcome this, systems which directly incorporate psychoacoustically motivated measures into the training objectives of NNSE systems have been proposed.

A key development in speech audio processing in recent years is the emergence of Self Supervised Speech Representation (SSSR) models. These are powerful foundational DNN models which can be utilised for a number of more specific speech processing tasks, such as speech recognition, emotion detection as well as SE. Finally, the methods of evaluation of SE systems have been revolutionised by DNN technology, that is to say the creation of systems which are able to directly predict Mean Option Score (MOS) ratings of Speech Quality (SQ) or Speech Intelligibility (SI) derived from human listening tests.

This thesis aims to investigate these three areas; psychoacoustic training objectives of NNSE, the incorporation of SSSR features and the prediction of human derived labels of speech directly from audio signals. Further, the intersection of these areas and combined use of techniques from these areas will be investigated.

A widely adopted approach for psychoacoustically motivated NNSE training is the MetricGAN framework. Here, a NNSE network is trained as generator adversarially (pitted against in competition) with a metric prediction discriminator. The discriminator is tasked with predicting the score assigned to the input audio by a (typically non-differentiable and thus unable to be used as a loss function directly) metric function, while the generator uses inference of the discriminator to obtain a loss value for its outputs. While MetricGAN has proved effective and is becoming a widely adopted technique, there is scope to improve it in several areas. Several of the contributions of this thesis are related to these improvements including the introduction of an additional DNN tasked with improving the range of inputs to the metric prediction Discriminator, changes to the Neural Network (NN) structure of both components and the prediction of non-intrusive measures among others. A key finding of this work is that perceptually motivated NNSE systems tend to overfit towards the target perceptual metric, resulting in degraded ”real world” enhancement performance. The concept of the metric prediction is further developed into systems proposed for the related task of DNN based human MOS prediction. This can be done intrusively meaning that the system has access to a non-distorted version of the signal under test as a reference or non-intrusively meaning that only the signal under test is available. Here, human labels of SQ or SI are directly predicted from the audio signal stimulus. SI prediction is mainly investigated in the domain of hearing aid SE system evaluation in this work. State of the art performance is achieved by SQ prediction systems developed and presented in this work.

Two novel applications of SSSR are presented. Firstly, as feature space representations in the loss function of NNSE systems. In particular, it is found that using earlier intermediate DNN layer outputs in this application is particularly effective, and a strong correlation between the SSSR distance measure and psychoacoustic metrics and MOS labels is shown. Secondly, SSSR representations are proposed for use as feature extractors for the discriminator DNN components of the MetricGAN framework, as well as for MOS estimators.


Links