Publications
Selected publications across my research. See the profiles below for the full record.
Uses a custom tool to elicit intermediate reasoning from frontier models, comparing performance with native reasoning and analysing differences in token efficiency and reasoning structure.
Develops a generalised extraction framework for diffusion language models, analysing how masking patterns and sampling resolution affect memorisation and training-data leakage.
Shows how colluding semi-honest nodes can recover model updates from secure aggregates in sparse decentralised federated-learning topologies, enabling reconstruction of private training data.
Introduces a behaviour-driven language-similarity measure based on perturbation-induced cross-lingual transitions, showing stable patterns across five multilingual models and two corpora covering 44 languages.
Compares six vision-language models with human speakers in a Hungarian information-structure task, finding that models track discourse context but over-regularise interacting linguistic pressures into narrow response templates.
Introduces CRAFT, a few-shot black-box attack that recovers captions and sensitive attributes from image embeddings across five open-source and proprietary embedders, while tested defences fail to close the leak without reducing downstream utility.
Introduces Cue-Resistant Memorization to separate genuine memorisation from reconstruction enabled by lexical cues, showing across 32 languages that much apparent PII leakage is cue-driven.
Introduces a multilingual, multi-hop benchmark grounded in knowledge-graph paths for evaluating and mitigating LLM hallucinations.
Introduces an explicit language-sampling framework and metrics, demonstrating that language selection materially affects the generalisability of multilingual evaluation claims.
Audits non-English Wikipedia with quality filters usually applied to noisy web data, showing that multilingual NLP must treat data quality—not only coverage—as a first-class concern.
Grounds reasoning traces in knowledge-graph paths and improves factuality across six open-domain question-answering benchmarks using 3.9K training examples.
Introduces benchmarks for financial information extraction from earnings calls, documenting domain-shift failures and evaluating open-ended LLM extraction with human experts.
Introduces a systematic framework for evaluating LLM-driven threat-modelling tools, assessing output quality, LLM integration, and governance risks.
Introduces a financial information-extraction dataset with hierarchical labels, supporting evaluation of KPI classification, extraction, and structured extraction.
Uses teacher–student LLM simulations to generate educational feedback data, improving smaller models’ feedback quality in expert and model-based evaluations.
Studies memorisation across 95 languages, model families, scales, and definitions, showing that language similarity reveals cross-lingual patterns beyond training-data volume.
Develops a taxonomy of how knowledge graphs can support the detection, evaluation, and mitigation of LLM hallucinations, and identifies open problems at the knowledge-graph–LLM interface.
Connects NLP-security research with established cybersecurity ethics, develops the idea of “white-hat NLP,” and offers guidance for harm minimisation and responsible disclosure.
Shows that partially successful text-embedding inversion is possible from a single example by aligning victim and attack embedding spaces, including transfer across domains and languages.
Studies multilingual embedding inversion across 20 languages, eight language families, and 12 scripts, identifying script- and family-specific vulnerabilities alongside predictable language confusion.
Introduces Language Confusion Entropy to quantify unintended output-language switches and links these failures to linguistic typology and multilingual embedding inversion.
Uses Shapley Head Values to analyse how BERT and RoBERTa attention heads organise around related morphosyntactic phenomena.
Defines multilingual and cross-lingual embedding-inversion attacks, showing that risk varies by language and English-centred defences do not automatically transfer.
Shows that NLP papers use “typological diversity” inconsistently, and provides measures for more defensible language sampling and multilingual evaluation.
Provides benchmarks for eight NLP tasks across as many as 28 Creole languages, including newly developed datasets.
Critically examines whether typological feature prediction serves its claimed purposes in NLP and linguistic documentation, using a survey of typologists to offer recommendations for better interdisciplinary alignment.
Learns language representations from cross-linguistic colexification graphs, providing a semantic-typology route to modelling language similarity.
Tests methods for removing language-identity signals from multilingual representations across 19 languages, clarifying when representation alignment improves cross-lingual transfer and when intuitive normalisation methods fail.
Uses meta-learning to select source-language training instances for zero-shot transfer across 15 languages, with a typological analysis of when learned cross-lingual sharing helps.
Introduces a question-answering dataset with subjectivity annotations for questions and answer spans across six review domains, making subjectivity effects measurable in extractive QA.
Disentangles structural, genetic, and geographical signals in language representations learned from translated text, finding structural similarity strongest while genetic relatedness can act as a confound.
Models languages and typological features jointly through exponential-family matrix factorisation, connecting feature covariance with generalisation to unseen languages.
Shows that multilingual language embeddings encode typological structure across phonology, morphology, and syntax, supporting their use for linguistic analysis and cross-lingual transfer.
Introduces a multilingual corpus whose translations share compositional meaning representations based on Discourse Representation Theory.
Introduces semantic tagging as a fine-grained task for multilingual semantic parsing and presents a deep residual architecture for the task.
Combines formal meaning representations and logical inference with statistical classification for textual entailment and semantic similarity, placing among the top systems in the SemEval 2014 entailment task.