M1 Thesis
A Hybrid DeepMet and Rule-Based NLP Framework for Identifying Personification Metaphors in Twitter Discourse on Artificial Intelligence
Summary: This article examines how personification metaphors shape public perceptions of generative AI through a computational analysis of Tweets posted during the emergence of ChatGPT (2022-2023). Using a hybrid methodology combining rule-based NLP and the DeepMet neural network model, we identify AI personification patterns and discuss their ideological implications. Results reveal that most metaphors employ Subject-Verb-Object constructions framing AI as an active agent. The major source domains are COGNITION (e.g. think), ACTION/CHANGE (e.g. replace), COMMUNICATION (e.g. listen), and EMOTION (e.g. love, fear). Sentiment analysis shows predominantly positive attitudes, particularly when framing AI as a collaborative assistant, while negative sentiment is strongly associated with job displacement concerns. These patterns suggest that users assign agency and autonomy to AI systems with critical implications: personification underpins an implicit social contract where augmentation is embraced and replacement resisted, yet it simultaneously obscures accountability by positioning AI as an independent actor rather than a product of corporate design and deployment. Moreover, as LLM outputs become more sophisticated, the figurative distance between computational systems and human behaviour risks collapsing entirely. Our findings show that metaphorical language dynamically constructs social realities about generative AI, raising stakes for public trust, responsibility attribution, and technology governance.
Summary: This article examines how personification metaphors shape public perceptions of generative AI through a computational analysis of Tweets posted during the emergence of ChatGPT (2022-2023). Using a hybrid methodology combining rule-based NLP and the DeepMet neural network model, we identify AI personification patterns and discuss their ideological implications. Results reveal that most metaphors employ Subject-Verb-Object constructions framing AI as an active agent. The major source domains are COGNITION (e.g. think), ACTION/CHANGE (e.g. replace), COMMUNICATION (e.g. listen), and EMOTION (e.g. love, fear). Sentiment analysis shows predominantly positive attitudes, particularly when framing AI as a collaborative assistant, while negative sentiment is strongly associated with job displacement concerns. These patterns suggest that users assign agency and autonomy to AI systems with critical implications: personification underpins an implicit social contract where augmentation is embraced and replacement resisted, yet it simultaneously obscures accountability by positioning AI as an independent actor rather than a product of corporate design and deployment. Moreover, as LLM outputs become more sophisticated, the figurative distance between computational systems and human behaviour risks collapsing entirely. Our findings show that metaphorical language dynamically constructs social realities about generative AI, raising stakes for public trust, responsibility attribution, and technology governance.
M2 Thesis
Do LLMs Use Metaphors of Mental Illness as Well as Humans? A Linguistic Comparison in Mental Health Counseling Contexts.
This thesis is funded by the Empirical Foundations of Linguistics (EFL) IdEx project, a six-year Université Paris Cité initiative (2025–2030) bringing together 11 research teams across 4 partner universities.
This thesis examines whether large language models use metaphors of mental illness in ways comparable to human counsellors in mental health counseling dialogues. It compares human-human counseling data with LLM-generated counseling dialogues in order to assess both metaphor use and its interactional effects. More specifically, the project analyzes whether LLM counsellors produce metaphors for mental health as human therapists do, and whether these metaphors lead to meaningful effects in dialogue rather than merely imitating surface form.
The methodology combines a hybrid NLP pipeline with interactional analysis. Metaphors are identified through DeepMet, a RoBERTa-based token-level metaphor detector, followed by top-down lexical filtering based on recurring source domains such as darkness, burden, entrapment, downward motion, fragility, and journey. A manually evaluated subset is then used to assess reliability. The project further investigates the effect of metaphor in interaction through measures such as engagement, elaboration, self-disclosure, and sentiment shift across counseling turns.
The study contributes to research on counseling linguistics, computational metaphor analysis, and AI-mediated mental health communication, while also addressing broader questions about how safely and effectively LLMs can approximate human therapeutic discourse.
This thesis is funded by the Empirical Foundations of Linguistics (EFL) IdEx project, a six-year Université Paris Cité initiative (2025–2030) bringing together 11 research teams across 4 partner universities.
This thesis examines whether large language models use metaphors of mental illness in ways comparable to human counsellors in mental health counseling dialogues. It compares human-human counseling data with LLM-generated counseling dialogues in order to assess both metaphor use and its interactional effects. More specifically, the project analyzes whether LLM counsellors produce metaphors for mental health as human therapists do, and whether these metaphors lead to meaningful effects in dialogue rather than merely imitating surface form.
The methodology combines a hybrid NLP pipeline with interactional analysis. Metaphors are identified through DeepMet, a RoBERTa-based token-level metaphor detector, followed by top-down lexical filtering based on recurring source domains such as darkness, burden, entrapment, downward motion, fragility, and journey. A manually evaluated subset is then used to assess reliability. The project further investigates the effect of metaphor in interaction through measures such as engagement, elaboration, self-disclosure, and sentiment shift across counseling turns.
The study contributes to research on counseling linguistics, computational metaphor analysis, and AI-mediated mental health communication, while also addressing broader questions about how safely and effectively LLMs can approximate human therapeutic discourse.
Course Projects
Phonetics and Phonology
- Analyzing F1 and F2 Formant Frequencies of English Monophthongs Produced by Nigerian Speakers
- Prosodic Boundaries in Human and LLM Speech: A Comparative Acoustic Analysis
- Tools: Praat, R
Cross-Linguistic Comparison
- A Corpus Analysis of the Uses of “Literally” in American English Movie Dialogue
- Tools: Sketch Engine
Historical Linguistics and Reconstruction
- The Use of the Singular Pronouns “Tu” and “Você” in Portuguese: Diachronic Perspectives and a Synchronic Case Study of Santa Catarina
- Tools: Corpus do Português: Genre/Histórico, Sketch Engine